---
kind: concept
slug: convolutional-neural-network
title: Convolutional neural network
summary: >-
  A network that slides learned filters over local neighbourhoods, stacking convolution, nonlinearity and down-sampling. Variants differ in dimension, depth, skips, recurrence, branching, sparsity, regularisation and hybrid transformer blocks. The 71 papers span segmentation, registration, saliency, text, pruning, transfer, video, point clouds, grasping, mineral mapping and robot force signals.
tags: [computer-vision, deep-learning, architecture]
papers: ["1505.00387", "1505.03046", "1506.00327", "1601.06615", "1603.01564", "1603.03827", "1610.02055", "1610.08815", "1612.05360", "1704.02422", "1704.06841", "1705.02544", "1705.08922", "1706.01040", "1706.06480", "1706.09911", "1708.02709", "1708.09427", "1711.10684", "1801.05746", "1803.01164", "1803.07835", "1804.02386", "1804.07265", "1806.00468", "1806.05258", "1807.03361", "1807.11809", "1811.04241", "1811.05588", "1812.01718", "1812.11446", "1901.07031", "1909.06581", "1910.10536", "2004.08919", "2004.11540", "2006.03541", "2007.04417", "2007.13184", "2012.10071", "2101.06742", "2101.07511", "2101.10861", "2103.01415", "2103.07138", "2103.07678", "2103.12731", "2103.15808", "2104.11892", "2105.03889", "2106.07229", "2106.11342", "2107.10356", "2108.09091", "2110.00791", "2112.15538", "2201.03299", "2201.03545", "2202.06709", "2211.05778", "2303.11851", "2303.14126", "2304.03198", "2307.09283", "2402.10220", "2410.17725", "2605.11989", "2605.30099", "2606.23286", "2609.40212"]
related: [semantic-segmentation, continuous-convolution, long-short-term-memory, teleoperation, optical-flow, local-global-attention, depthwise-separable-convolution, convolutional-recurrent-network, vision-transformer, self-attention, data-augmentation, dropout, transfer-learning, federated-learning, image-registration, image-manipulation-detection, time-series-forecasting, attention-mechanism, object-detection, instance-segmentation, multi-scale-feature-fusion, deformable-convolution, deep-unfolding, data-consistency-layer, multi-label-classification, noisy-labels, dataset-curation, text-classification, recurrent-neural-network, multi-task-learning, linear-probing, metric-learning, contrastive-learning, graph-neural-network, efficient-vision-backbone, knowledge-distillation, domain-adaptation, salient-object-detection, local-feature-matching, test-time-augmentation, sentiment-analysis, text-embedding, multi-layer-perceptron, image-thresholding, multimodal-fusion, interpretable-machine-learning, image-super-resolution, generative-adversarial-network, algorithmic-fairness, positional-encoding, channel-attention, time-series-classification, time-series-imputation, residual-learning, robotic-grasp-generation, post-training-quantization, network-pruning, image-enhancement, bibliometric-analysis]
status: draft
updated_at: 2026-11-01
---

## Overview

A convolutional layer applies the same small set of learned filters at every
position of its input, so each output value depends only on a local
neighbourhood. Stacking such layers with nonlinearities and down-sampling gives
features that cover a larger area as depth grows. The cited papers use this one
recipe at very different scales: a two-layer classifier on 32 x 32 images
[arxiv:2303.14126 fig. 5], a one-dimensional network over 200-point GPS
segments [arxiv:1804.02386 §3.2], a three-dimensional network that predicts
a deformation field between two volumes [arxiv:1807.03361 §2.3], and deep
encoder-decoder networks for per-pixel labels [arxiv:1612.05360 §3.1].

The authors of one cited analysis read convolutions as high-pass filters that
keep fine detail, in contrast to self-attention, which they read as a
low-pass smoothing step [arxiv:2202.06709 §3]. That reading is the authors'
interpretation, not a measured property of every CNN. A 2016 survey gives the
same building blocks and the AlexNet recipe as a starting point
[arxiv:1601.06615 §2]. A 2018 survey gives the convolution and pooling formulas
and the output-size rules, then walks through LeNet to DenseNet
[arxiv:1803.01164 §III]. Most cited papers are applications with their own
measurements. A 2021 survey of detectors lists the classifier backbones that
detectors reuse, such as ResNet and EfficientNet, and copies their ImageNet
top-1 from the source papers [arxiv:2104.11892 §4] [arxiv:2104.11892 tab. 2].
A free textbook covers the same building blocks with runnable code, and
its convolutional chapter starts on page 239 [arxiv:2106.11342 p. 239]. One of
the cited papers is a position paper on brain-inspired AI, with no equations
or experiments and no convolutional design detail [arxiv:1706.01040 §3], so it
adds no evidence for the choices below.

A 2021 systematic review of machine learning for disease diagnosis names CNNs
as the most common method. That is a count of the reviewed papers, not a
controlled comparison [arxiv:2112.15538 §7].

Two 2021 reviews of application areas show the same pattern. Of the 232
articles and proceedings that one review coded for UAV imagery, 91.2% use a
CNN [arxiv:2101.10861 §3]. A weed-detection survey notes that pre-trained CNNs
are common, while the accuracies it lists come from different crops, datasets
and metrics and are not comparable [arxiv:2103.01415 §10.1.1] [arxiv:2103.01415 §11].
Both are counts of the reviewed work, not controlled comparisons.

CNNs can also learn attributes that clinicians cannot see. Standard CNNs trained on medical scans predicted self-reported race with one-vs-all AUC of about 0.96 to 0.99 on chest X-rays [arxiv:2107.10356 tab. 3], and the authors could not isolate the image features responsible [arxiv:2107.10356 p. 21].

## Variants

### Encoder-decoder with residual blocks and summation skips

Four down-sampling levels with 64 to 512 feature maps, a 1024-map bridge at
40 x 40, and a decoder that deconvolves and adds the matching encoder map
instead of concatenating it [arxiv:1612.05360 §3.1] [arxiv:1612.05360 tab. 1].
Each level uses residual blocks. Reported on one electron-microscopy volume
per dataset, with V_rand 0.97804 against 0.97276 for U-Net on Drosophila
[arxiv:1612.05360 tab. 2] and Dice 0.9461 against 0.9085 on zebrafish
[arxiv:1612.05360 tab. 3]. Card: [FusionNet](/papers/1612.05360).

### Fully convolutional segmentation with object-level voting

A VGG16-based fully convolutional network, fine-tuned from ImageNet weights,
labels every pixel of 1000 x 1000 px patches. The patches are stitched back
into the full image, and each object takes the majority label of its pixels
[arxiv:1706.06480 §4.2] [arxiv:1706.06480 fig. 3]. The authors use this to
avoid resizing objects, since resizing destroys the texture they need
[arxiv:1706.06480 §4.1]. Training is staged: the network grows from FCN-32s
to FCN-16s and FCN-8s, with a learning rate that drops at each stage
[arxiv:1706.06480 §4.3.3]. On 21 steel image pairs, with an 11/10 image split,
the authors report 93.94% object accuracy against 66.50% for an object-based
VGG16 [arxiv:1706.06480 tab. 1] [arxiv:1706.06480 §5.1]. Class-balanced
patches and rotation augmentation account for the rise from 90.97% to 93.94%
[arxiv:1706.06480 tab. 2]. Card: [Advanced Steel Microstructural Classification](/papers/1706.06480).

### Inception-style recurrent residual blocks

Each block runs parallel 1 x 1 and 3 x 3 branches whose convolutions are
unrolled over time steps, adds a pooled branch, and adds the block input back
as a shortcut [arxiv:1811.04241 §3] [arxiv:1811.04241 eq. 3] [arxiv:1811.04241 eq. 4].
The experiments use two time steps of recurrence after one feed-forward
convolution [arxiv:1811.04241 §4]. On BreakHis at 40x, the authors report 97.95%
image-level binary accuracy with augmentation, against 97.16% without it
[arxiv:1811.04241 tab. 4]. Card: [IRRCNN](/papers/1811.04241).

### Small two-layer image classifier

Two convolution blocks with 32 filters and max pooling, a 64-unit dense layer,
and one sigmoid output [arxiv:2303.14126 §3.3] [arxiv:2303.14126 fig. 5]. It
reaches 93.55% accuracy and F1 0.936 on the authors' 20,000-image test split
[arxiv:2303.14126 tab. 7] [arxiv:2303.14126 tab. 11]. The paper's headline
92.98% comes from a different topology with 128 filters [arxiv:2303.14126 tab. 2].
No external baseline is reported [arxiv:2303.14126 §4.2]. Card: [CIFAKE](/papers/2303.14126).

### Two-convolution gesture classifier converted with TensorFlow Lite

Two blocks, each a 3 x 3 convolution, ReLU, 2 x 2 max pooling and dropout, with
32 and then 64 filters, followed by a 128-unit dense layer and a five-way
softmax [arxiv:2110.00791 §2.2] [arxiv:2110.00791 tab. 1]. The same design is
trained at four square input sizes, from 64 to 256 pixels, and the authors
choose 96 x 96 [arxiv:2110.00791 §2]. Training uses Adam with batch size 32, at
most 25 epochs, and a stop once validation loss has risen for three epochs
[arxiv:2110.00791 §2.2]. The trained model is then converted with TensorFlow
Lite, which the authors say lowers the precision of the parameters. They also
describe structured pruning on a representative dataset, but give no ratio
[arxiv:2110.00791 §3]. At 96 x 96 the authors report 98.11% accuracy and
47,864,936 bytes before conversion, against 96.9% and 15,942,712 bytes after
[arxiv:2110.00791 tab. 2]. The two accuracies come from different sets, 100
validation images against 32 test images, and the paper does not show that
they are comparable [arxiv:2110.00791 tab. 2]. The paper also does not say
what "model size" measures [arxiv:2110.00791 §3]. Card:
[Optimizing Neural Network for Computer Vision task in Edge Device](/papers/2110.00791).

### One-dimensional CNN over sensor sequences

Each 200-point segment is a 1 x 200 x 4 tensor of speed, acceleration, jerk and
bearing rate [arxiv:1804.02386 §3.2.1]. The network uses 1 x 3 convolutions with
stride 1, 1 x 2 max pooling, ReLU and a five-way softmax [arxiv:1804.02386 §3.2].
A single model reaches 79.8% test accuracy, and a seven-member bootstrap
ensemble reaches 84.8% [arxiv:1804.02386 tab. 4]. Both numbers use a test-set
epoch choice, so they are optimistic [arxiv:1804.02386 §5.1]. Card: [Transportation modes from GPS](/papers/1804.02386).

### Convolutional features feeding an LSTM for robot force signals

Each sample is 1000 frames by 24 channels, the joint and force readings of a
teleoperated arm sampled at 100 Hz [arxiv:2402.10220 §III-D] [arxiv:2402.10220 §IV-A].
Each channel is z-scored [arxiv:2402.10220 eq. 6]. A stack of four convolution and
pooling blocks, then flatten, dropout and dense layers, gives a six-way softmax.
The text and the figures disagree on the convolution dimension and pooling type
[arxiv:2402.10220 fig. 7a] [arxiv:2402.10220 §IV-A]. For forecasting, the same convolutional encoder feeds an
LSTM with four stacked layers of 100, 100, 50 and 50 units and a 200-unit dense
layer, which predicts the next 500 frames [arxiv:2402.10220 fig. 8d] [arxiv:2402.10220 §IV-E].
The paper reports a single-stream classifier at 86% accuracy, a dual-stream at 84%
and a quadruplet-stream at 89% on its second dataset [arxiv:2402.10220 §IV-C]. The
forecaster's RMSE in the table is 1.91e-05 for the quadruplet hybrid, against 0.065
for an LSTM fed the spatial mean of the channels [arxiv:2402.10220 tab. III]. The
authors do not say whether the convolutional encoder is frozen or fine-tuned during
forecasting [arxiv:2402.10220 §IV-E.2]. Card: [Towards Semi-Autonomous Robotic Arm Manipulation Operator Intention Detection from Force Data](/papers/2402.10220).

### Three-dimensional CNN for weakly supervised registration

A 3D encoder-decoder with four down-sampling and four up-sampling blocks. Each
block has two convolution, batch-norm and ReLU layers, and the first
convolution uses a 7 x 7 x 7 kernel, while the rest use 3 x 3 x 3
[arxiv:1807.03361 §2.3]. The output is a dense displacement field that warps
a moving MR volume onto a fixed ultrasound volume. Training uses only sparse
anatomical label pairs, scored with a multiscale soft Dice, plus a
bending-energy penalty [arxiv:1807.03361 §2.1] [arxiv:1807.03361 §2.2]. Under
12-fold patient-level cross-validation on 108 prostate pairs, the authors report
a median landmark error of 3.6 mm and a median gland Dice of 0.87
[arxiv:1807.03361 tab. 1]. Training ran for 48 hours on one 12 GB GPU
[arxiv:1807.03361 §3.2]. Card: [Weakly-Supervised Convolutional Neural Networks for Multimodal Image Registration](/papers/1807.03361).

### One-dimensional CNN encoders for molecules and proteins

A library exposes a 1D CNN over SMILES characters and another over protein
residues, each with global max pooling. The two embeddings are concatenated and
passed to an MLP [arxiv:2004.08919 §2.1] [arxiv:2004.08919 app. A]. On the DAVIS
benchmark the authors report MSE 0.254 and concordance index 0.879 for the
CNN+CNN pair, against 0.262 and 0.870 for DeepDTA [arxiv:2004.08919 fig. 1]. These
values are read from a low-resolution figure, and the paper does not state the
training hyperparameters used for them [arxiv:2004.08919 §3]. Card: [DeepPurpose](/papers/2004.08919).

### Small U-Net with depthwise-separable convolutions and attention

A U-Net whose double convolutions are depthwise-separable, with a CBAM
attention block on each skip path, predicts rain maps 30 minutes ahead from 12
five-minute radar frames [arxiv:2007.04417 §3.1] [arxiv:2007.04417 §4.1]. The
model has 4.1 million parameters against 17.3 million for U-Net [arxiv:2007.04417 tab. 1].
On the authors' NL-50 test set, both models reach MSE 0.0122, and the model
has slightly lower NMSE, 416.10 against 416.38 [arxiv:2007.04417 tab. 4]. The
training loss in the released code is a per-image sum, not the mean of the
paper's equation [arxiv:2007.04417 eq. 1]. Card: [SmaAt-UNet](/papers/2007.04417).

### Deep pretrained-style backbone inside federated learning

A VGG16 backbone with one extra convolution and three dense layers before the
output, trained on grayscale 256 x 256 images with Adam at a learning rate of
0.0001 and focal loss [arxiv:2101.07511 §3.2.2]. Clients are grouped by
imaging modality, and the server averages weights within and across groups
[arxiv:2101.07511 §3.1] [arxiv:2101.07511 alg. 1]. Card: [Collaborative federated learning](/papers/2101.07511).

### CNN backbone with local self-attention at stage ends

The AlterNet design is a pre-activation ResNet-50 in which some convolution
blocks at the end of stages are replaced by local multi-head attention
[arxiv:2202.06709 app. E]. The paper states a 4 x 4 window for CIFAR; the
released code defaults to 7 [arxiv:2202.06709 app. E] [arxiv:2202.06709 §4.1].
The authors report that it beats CNN baselines on CIFAR-100 under corruption
[arxiv:2202.06709 §4.2], with the values given only in figures, and is on par
with Swin in ImageNet accuracy, a comparison that uses different training
pipelines [arxiv:2202.06709 app. E]. Card: [How Do Vision Transformers Work?](/papers/2202.06709).

### Generic scene features from a scene-centric dataset

AlexNet, GoogLeNet and VGG-16 are trained on a dataset of about 2.4 million
images in 205 scene categories, and on a 365-category subset of about 1.8
million images [arxiv:1610.02055 §2.3.1] [arxiv:1610.02055 §3]. The
authors' VGG model scores 58.90% top-1 on Places205 test and 55.19% top-1 on
Places365 test, with 10-crop averaging [arxiv:1610.02055 tab. 1] [arxiv:1610.02055 tab. 2].
For reuse as a fixed feature extractor, they take fc7 activations (4096-d) or
the pooled layer of GoogLeNet (1024-d) and train a linear SVM with C = 1
[arxiv:1610.02055 §3.3]. The paper does not state the learning rate, batch
size or iteration count for these CNNs [arxiv:1610.02055 §3]. Card: [Places](/papers/1610.02055).

### Seven-layer detector for computers without a GPU

Built from Tiny-YOLOv2 by trial and error: the authors vary layer count,
filter count and input size, keep the best variant, and retrain it
[arxiv:1811.05588 §3.1]. The chosen network has seven convolutions with 16 to
256 filters, 2 x 2 max-pooling after the first five, and a linear 1 x 1
convolution that outputs a grid of five boxes per cell [arxiv:1811.05588 §4.1].
The authors remove batch normalisation, on the argument that it costs time in
a small network [arxiv:1811.05588 §3.5]. Trained on PASCAL VOC, it reaches
33.57% mAP at 21 FPS on one laptop without a GPU, against 40.48% and 2.4 FPS
for Tiny-YOLOv2 [arxiv:1811.05588 tab. III]. The paper gives 12.26% mAP on
COCO at the same speed [arxiv:1811.05588 tab. V]. Card: [YOLO-LITE](/papers/1811.05588).

### Structured sparsity: pruning convolution weights at four grain sizes

Magnitude pruning removes convolution weights at one of four grain sizes:
single weights, sub-kernel vectors, whole kernels or whole filters. Each grain
is scored by the sum of its absolute weights, and the lowest-scoring grains go
first [arxiv:1705.08922 §3.3] [arxiv:1705.08922 eq. 1]. The four grains are
compared at equal density, with the same method and training schedule, on
ImageNet [arxiv:1705.08922 §4]. At matched density, fine-grained pruning gives
the highest top-5 accuracy for all five models, and vector pruning lands
between it and kernel pruning [arxiv:1705.08922 tab. 1]. Coarser grains need
fewer index bits. At the baseline accuracy, vector pruning gives the lowest
storage ratio for VGG-16, GoogLeNet and ResNet-50, while fine-grained pruning
does so for AlexNet and DenseNet-121 [arxiv:1705.08922 tab. 2]. In a simulated
SCNN-style dataflow on VGG-16 convolutions, vector pruning needs 30 to 35%
fewer output memory references than fine-grained pruning [arxiv:1705.08922 tab. 3].
The results are author-reported. Card: [Exploring the Regularity of Sparse Structure in Convolutional Neural Networks](/papers/1705.08922).

### Pretrained CNN as a starting point for fine-tuning

Eleven ImageNet-pretrained PyTorch CNNs, from AlexNet to MnasNet, are adapted to
five small image sets. Each is tuned in one of two ways: only the classifier is
retrained, or every layer is [arxiv:2605.11989 §4]. The score is accuracy divided
by parameter count, which the authors call accuracy density [arxiv:2605.11989 eq. 1].
Training uses Adam, batch size 10 and learning rates from {1e-2, 1e-3, 1e-4},
with early stopping on validation accuracy [arxiv:2605.11989 §4]. The authors
report SqueezeNet as the densest model under full tuning and ResNet-18 under
classifier-only tuning, with one pass over the data [arxiv:2605.11989 §5.1.1]
[arxiv:2605.11989 §5.1.2]. With ten passes, ResNet-18 is again the most efficient
under classifier-only tuning [arxiv:2605.11989 §5.2.2]. The text gives no
numeric accuracy, density or time, and the values appear only in plotted figures
[arxiv:2605.11989 fig. 3] [arxiv:2605.11989 fig. 4]. Card: [A Transfer Learning Evaluation of Deep Neural Networks for Image Classification](/papers/2605.11989).

### Linear convolutional network (theory of implicit bias)

No nonlinearities: the network is a product of full-width circular
convolutions, and its last layer is a dense vector [arxiv:1806.00468 §2].
With exponential loss on separable data, gradient descent on the weights
converges in direction to a stationary point of a penalty on the Fourier
coefficients, the l_{2/L} quasi-norm, so the depth L changes which separator is
chosen [arxiv:1806.00468 §3]. A fully connected linear network of any depth
converges to the hard-margin SVM direction instead [arxiv:1806.00468 §3]. The
paper has no experiments [arxiv:1806.00468 §6]. Card: [Implicit Bias of Gradient Descent on Linear Convolutional Networks](/papers/1806.00468).

### Classic AlexNet-style stack as the reference design

Five convolutional layers and three fully connected layers, with 96 first-layer
filters of size 11 x 11 x 3 [arxiv:1601.06615 §2.1] [arxiv:1601.06615 fig. 1].
Max pooling follows layers 1, 2 and 5, dropout is applied to the last two fully
connected layers, and the survey names mini-batch SGD with batches of about 100
as the usual optimiser [arxiv:1601.06615 §2.5] [arxiv:1601.06615 §2.3.1]. The
learning rate, momentum, dropout rate and weight decay were hand-tuned and are
not given in the survey [arxiv:1601.06615 §2.5]. The survey runs no experiments
of its own. Its figures of about 3% from augmentation and 1 to 2% from ensembles
are summaries of cited studies [arxiv:1601.06615 §2.4]. Card: [A Taxonomy of Deep Convolutional Neural Nets for Computer Vision](/papers/1601.06615).

### Text CNN over word vectors for sentence labels

Each sentence becomes a 50 x 100 matrix of word2vec vectors, padded or cut to
50 tokens [arxiv:1704.06841 §3]. Two groups of two convolution layers follow,
each with 256 filters of width 5 and ReLU, and each group ends in max pooling.
A dropout of 0.5, a 128-unit dense layer, a second dropout and a 26-way softmax
complete the network [arxiv:1704.06841 §3] [arxiv:1704.06841 fig. 1]. The
authors report 0.68 accuracy, against 0.28 to 0.51 for four baselines
[arxiv:1704.06841 fig. 2]. The sentences come from encyclopedia-style articles
with one label per article, the dataset is not released, and there is no
separate test split [arxiv:1704.06841 §3] [arxiv:1704.06841 §4]. Card: [Medical Text Classification using Convolutional Neural Networks](/papers/1704.06841).

### Text CNN with FastText embeddings for user-level labels

Follows Kim (2014) with filter size 3 and pooling size 4, trained for 100
epochs with one model per condition. A user's posts are concatenated, cut to
15,000 tokens and embedded with FastText vectors [arxiv:1806.05258 §5]. Across
nine conditions it is best only on eating, with binary F1 of 44.53, while
FastText is best on 7 of 9 binary tasks [arxiv:1806.05258 tab. 4]. The authors
attribute FastText's lead to reading all of a user's posts, where the CNN
samples a portion of them [arxiv:1806.05258 §5]. Card: [SMHD](/papers/1806.05258).

### Greedy layer-wise CNN with auxiliary classifiers

Each block is a 3 x 3 convolution with bias and ReLU. Downsampling is an
invertible 2 x 2 space-to-depth step at some depths, not pooling
[arxiv:1812.11446 §3.1]. Each block is trained with a small auxiliary classifier
on the frozen output of the blocks before it, so no end-to-end backpropagation
is needed [arxiv:1812.11446 §3.2] [arxiv:1812.11446 alg. 1]. On ImageNet, the
single-crop top-5 accuracy is 79.7% for one-hidden-layer blocks, against 79.1%
for end-to-end AlexNet, and 88.7% for three-hidden-layer blocks, against 88.0%
for end-to-end VGG-11 [arxiv:1812.11446 tab. 3]. On CIFAR-10 the authors report
88.3% and 91.7% for one- and three-hidden-layer blocks, against 89% for
end-to-end AlexNet and 92.5% for end-to-end VGG [arxiv:1812.11446 tab. 2]. The
memory saving is an author claim from CIFAR-10, where end-to-end training uses
about five times more GPU memory; no ImageNet memory figures are given
[arxiv:1812.11446 §4.1]. Card: [Greedy Layerwise Learning Can Scale to ImageNet](/papers/1812.11446).

### Grid CNN baseline for urban traffic

Grid models stack the input time steps into channels, or apply convolutions to
each cell of a city grid [arxiv:2108.09091 §4]. The plain CNN is the grid
baseline. On single-step TaxiBJ prediction (six input steps, one output step),
it reaches RMSE 23.550, against 18.702 for ST-ResNet, which uses residual CNN
blocks over closeness, period and trend inputs, and 17.826 for STDN, which
combines local CNNs with LSTM and attention [arxiv:2108.09091 tab. 4]
[arxiv:2108.09091 §4.1] [arxiv:2108.09091 §5.1]. The authors say no grid model
dominates, with STDN generally strongest [arxiv:2108.09091 §5.2]. Card: [DL-Traff](/papers/2108.09091).

### ConvNeXt-B shared encoder for matching images across views

One ConvNeXt-B backbone, about 88 million parameters, embeds both views, and
features are mean-pooled with no attention pooling [arxiv:2303.11851 §4.2]
[arxiv:2303.11851 §3.2]. On the VIGOR same-area split, the ConvNeXt-B row
reaches R@1 75.81, against 69.89 for the authors' ViT variant of similar size
[arxiv:2303.11851 tab. 4]. The ConvNeXt choice rests on this ablation
[arxiv:2303.11851 §4.3.1]. Card: [Sample4Geo](/papers/2303.11851).

### Highway layers: gated skip connections

A highway layer outputs a sigmoid-gated mix of a transform H(x) and the
untouched input x [arxiv:1505.00387 §2, eq. 3]. The gate bias starts negative,
so each layer begins by carrying its input [arxiv:1505.00387 §2.2]. The
convolutional version, a 19-layer network trained directly, reaches 92.24% on
CIFAR-10 test [arxiv:1505.00387 tab. 1] [arxiv:1505.00387 §3.2]. The authors
report that 100-layer fully connected highway nets train with plain SGD where
plain nets of that depth degrade [arxiv:1505.00387 §3.1]. The paper reports no
variance across seeds [arxiv:1505.00387 §3.1]. Card: [Highway Networks](/papers/1505.00387).

### Patch-trained CNN converted to a whole-image classifier

A patch classifier is trained on lesion-centred patches. Its convolutional
layers are then run on a whole image, followed by two convolutional top blocks,
global average pooling and a softmax output, with no heatmap in between
[arxiv:1708.09427 §2.2] [arxiv:1708.09427 fig. 1]. Training is two stages on
image labels: first the new top layers, then all layers [arxiv:1708.09427 §3.1.2].
On the DDSM test split, a four-model ensemble with flip averaging reaches
per-image AUC 0.91 [arxiv:1708.09427 §3.1.4] [arxiv:1708.09427 fig. 3]. The best
single model reaches 0.88 with augmentation [arxiv:1708.09427 tab. 2]. Models
transferred to INbreast reach 0.95 each and 0.98 as an ensemble on a 107-image
validation split [arxiv:1708.09427 §3.2.2] [arxiv:1708.09427 tab. 4]. The
authors' code is non-commercial, so this variant is described from the paper
alone. Card: [Deep Learning to Improve Breast Cancer Early Detection on Screening Mammography](/papers/1708.09427).

### Residual U-Net without cropping

The encoder-decoder keeps U-Net's shape, but each level is a residual unit in
full pre-activation order, downsampling uses a stride-2 convolution, and the
decoder needs no cropping because input and output have the same size
[arxiv:1711.10684 §2.1.3] [arxiv:1711.10684 tab. I]. It has 15 convolution
layers against 23 in U-Net [arxiv:1711.10684 §2.1.3]. It is trained with mean
squared error on 30,000 random 224 x 224 patches, without augmentation
[arxiv:1711.10684 §3.2]. On the Massachusetts roads test set the relaxed
break-even point is 0.9187, against 0.9053 for U-Net [arxiv:1711.10684 tab. II].
The text puts it at about 7.8M parameters against 30.6M for U-Net
[arxiv:1711.10684 §3.4]. The paper does not say how the shortcut is reshaped
where stride or filter count changes [arxiv:1711.10684 §2.1.2]. Card: [Road Extraction by Deep Residual U-Net](/papers/1711.10684).

### Unrolled cascade with data-consistency layers

A CNN block is followed by a data-consistency layer that puts back the measured
k-space samples. The pair is repeated and trained end to end [arxiv:1704.02422 §3]
[arxiv:1704.02422 §4]. Each block uses 3 x 3 convolutions with 64 filters, ReLU
and a residual connection [arxiv:1704.02422 §6]. For dynamic sequences, a
data-sharing layer adds k-space averages from neighbouring frames as extra input
channels [arxiv:1704.02422 §5] [arxiv:1704.02422 §6]. At 3-fold undersampling on
2D frames, the MSE is 0.89 x 10^-3, against 2.12 x 10^-3 for DLMRI
[arxiv:1704.02422 tab. 1]. The authors time one frame at 23 ms on a GTX 1080
[arxiv:1704.02422 §7.2.3] and one 3D sequence at 8.21 s [arxiv:1704.02422 §7.3.2].
These timings are the authors' own. Card: [Cascaded CNN for dynamic cardiac MRI](/papers/1704.02422).

### Mixup-trained residual network for character images

Three character-image benchmarks compare a 4-nearest-neighbour classifier, a
small two-layer CNN and PreActResNet-18, the last with and without mixup
[arxiv:1812.01718 tab. 1] [arxiv:1812.01718 §3.1]. On Kuzushiji-MNIST the
mean per-class accuracy is 91.56% for the nearest-neighbour baseline, 95.12%
for the small CNN, 97.82% for PreActResNet-18, 98.41% with input mixup and
98.83% with manifold mixup [arxiv:1812.01718 tab. 1]. The paper gives no
training details for these baselines [arxiv:1812.01718 §3.1]. Some classes in
the two larger datasets have few samples [arxiv:1812.01718 §2]. Card: [Deep Learning for Classical Japanese Literature](/papers/1812.01718).

### DenseNet121 on chest radiographs with uncertain-label handling

The model takes 320 x 320 radiographs and outputs one probability per finding
[arxiv:1901.07031 § Training Procedure]. Four ways of handling uncertain labels
are compared: ignore them, map them to 0 or to 1, self-train on the model's own
soft labels, or use a three-way softmax [arxiv:1901.07031 § Uncertainty Approaches].
No single method wins everywhere. On validation AUROC, U-Ones is best for
Atelectasis (0.858) and U-MultiClass for Cardiomegaly (0.854), and the paper
reports significance only for those two findings [arxiv:1901.07031 tab. 3]
[arxiv:1901.07031 § Results]. Ignoring uncertain labels is a poor fit when
uncertainty is common: for Consolidation, 12.78% of studies are uncertain against
6.78% positive [arxiv:1901.07031 tab. 1]. On a 500-study test set, the model
lies above all three radiologists on Cardiomegaly, Edema and Pleural Effusion,
but not above their majority vote [arxiv:1901.07031 § Test Set] [arxiv:1901.07031 fig. 3].
Card: [CheXpert](/papers/1901.07031). The paper's CC BY licence does not cover
the dataset, so check the access terms before use.

### Deformable 3 x 3 operator with shared projection and softmax weights

DCNv3 replaces fixed 3 x 3 weights with learned sampling offsets and modulation
weights [arxiv:2211.05778 §3.1]. Three changes make it trainable from scratch at
scale: one projection shared by all sampling points, channel groups that each
have their own offsets, and a softmax over the points in place of DCNv2's
sigmoid [arxiv:2211.05778 §3.1]. A sigmoid in place of the softmax drops the
small model's top-1 from 83.5 to 65.7 [arxiv:2211.05778 tab. 6]. Larger 5 x 5
or 7 x 7 kernels do not improve top-1 [arxiv:2211.05778 app. B]. The 30M-parameter
variant reaches 83.5 top-1, against 82.1 for ConvNeXt-T [arxiv:2211.05778 tab. 2].
The largest variant reaches 89.6 at 640 px, but its pre-training uses a joint
427M-image set that the paper does not list [arxiv:2211.05778 tab. 2] [arxiv:2211.05778 §4.1].
The authors report that InternImage-B is slower than ConvNeXt-B despite higher
accuracy [arxiv:2211.05778 app. D]. Card: [InternImage](/papers/2211.05778).

### Real-time detector with C3k2 and C2PSA blocks

YOLOv11 is described as a review, with no experiments of its own. C3k2 blocks
replace C2f in the backbone, neck and head, and a C2PSA attention block follows
SPPF [arxiv:2410.17725 §4.1.1] [arxiv:2410.17725 §4.1.2]. The paper gives no
channel counts, strides or kernel sizes per stage [arxiv:2410.17725 §4]. Its
performance figures are approximate values read from a third-party chart
[arxiv:2410.17725 §6] [arxiv:2410.17725 fig. 2]. Card: [YOLOv11: An Overview of the Key Architectural Enhancements](/papers/2410.17725).

### Fine-tuned AlexNet for small photo sets

An ImageNet-pretrained AlexNet has its last three layers replaced for 12
phone classes and is trained on 650 back-side photos [arxiv:2606.23286 §4].
Each training photo is copied nine times with flips, translation of up to 50
pixels, scale between 0.9 and 1.1, and shear [arxiv:2606.23286 §5.1]. At
learning rate 3e-4, SGD with momentum reaches 98.35 ± 2.02% over seven trials,
against 95.26 ± 3.04% for Adam and 93.20 ± 3.86% for RMSProp
[arxiv:2606.23286 tab. 1]. These are validation accuracies, and no separate
test set is described [arxiv:2606.23286 §4]. The paper does not say whether
augmentation is applied before or after the split, so copies of one photo may
land on both sides [arxiv:2606.23286 §4]. Card: [Transfer learning-based method for automated ewaste recycling in smart cities](/papers/2606.23286).

### Reported CNN use in agricultural studies

This is a survey of 40 agricultural deep-learning studies, not a method
[arxiv:1807.11809 §2]. Convolutional networks are the most common architecture,
in 17 of the 40 studies [arxiv:1807.11809 §4.6]. Of the 24 studies reporting
classification accuracy, 19 are above 90%, but only 8 of the 40 tested on data
other than their training data [arxiv:1807.11809 §4.9] [arxiv:1807.11809 §4.10].
The survey warns that the results are not directly comparable
[arxiv:1807.11809 §4.9]. Card: [Deep learning in agriculture: A survey](/papers/1807.11809).

### U-Net with an ImageNet-pretrained VGG11 encoder

The encoder is the convolutional part of VGG11, initialised from ImageNet
weights, and the decoder mirrors it with transposed convolutions
[arxiv:1801.05746 §2]. Training uses binary cross-entropy minus the log of a
soft Jaccard term [arxiv:1801.05746 eq. 4]. On 30 validation images of the
Inria aerial dataset, validation IoU after 100 epochs is 0.593 with random
initialisation and 0.686 with the pretrained encoder [arxiv:1801.05746 §3].
A network pretrained on Carvana reaches 0.687 [arxiv:1801.05746 §3]. Both
pretrained settings converge faster in the authors' curves [arxiv:1801.05746
fig. 3]. Each setting is one run with no seeds or variance reported. Card:
[TernausNet](/papers/1801.05746).

### Parallel CNN and transformer branches with feature exchange

A convolutional branch and a transformer branch share one stem and run side
by side. A Feature Coupling Unit in each block sends convolutional features
into the transformer as pooled patch tokens, and sends transformer tokens back
into the convolutional branch after reshaping and upsampling
[arxiv:2105.03889 §3.2]. The paper leaves out where the upsampled features
enter the CNN block [arxiv:2105.03889 §3.2]. On ImageNet-1k at 224 px,
Conformer-S reaches 83.4% top-1 against 81.8% for DeiT-B [arxiv:2105.03889
tab. 2]. A serial stack of ResNet-26d and DeiT-S reaches 80.2% with 3.7
GMACs, against 81.9% for Conformer-S/32 with 7.0 GMACs [arxiv:2105.03889
tab. 5]. On COCO, Conformer-S/32 in Faster R-CNN reaches 43.1 box AP against
39.4 for ResNet-101 FPN [arxiv:2105.03889 tab. 3]. Its detection throughput
is lower: 13.5 FPS against 15.9 for ResNet-101 on a V100 [arxiv:2105.03889
app. C]. Card: [Conformer](/papers/2105.03889).

### Pure CNN with transformer-style block layout for phones

The block separates spatial mixing, a depthwise 3 x 3 convolution that is
folded into one kernel at inference, from channel mixing, a pointwise
expansion to twice the width [arxiv:2307.09283 §3.2]. The stem, the
downsampling blocks and the squeeze-and-excitation placement follow the
same ViT-inspired choices [arxiv:2307.09283 §3.3] [arxiv:2307.09283 §3.4].
The authors state that operation counts predict on-device latency poorly, so
they measure latency on an iPhone 12 with Core ML [arxiv:2307.09283 §3.1].
With distillation, RepViT-M1.0 reaches 80.0% top-1 at 1.0 ms, against 79.0%
at 1.1 ms for EfficientFormerV2-S1 [arxiv:2307.09283 tab. 1]. Without
distillation the same model reaches 78.6% [arxiv:2307.09283 tab. 2]. The
step-by-step ablation starts from 71.5% for MobileNetV3-L with the aligned
recipe and ends at 77.4% for the full block [arxiv:2307.09283 §3.1]
[arxiv:2307.09283 §3.4]. Card: [RepViT](/papers/2307.09283).

### Random-view 2.5D classifier as a second-stage filter

Each candidate from an existing detector is cropped at several physical scales
and under random shifts and rotations, giving about 100 views per candidate
[arxiv:1505.03046 §III-A]. A small network reads three orthogonal central slices
as one 2.5D patch, and the candidate score is the mean of the per-view
probabilities [arxiv:1505.03046 §II-B] [arxiv:1505.03046 eq. 1]. The network has
two convolution layers with max pooling, two locally connected layers and a
fully connected layer with DropConnect [arxiv:1505.03046 §II-A]
[arxiv:1505.03046 fig. 1]. At 3 false positives per volume on lymph nodes,
sensitivity is 70% in the mediastinum and 83% in the abdomen, against prior
results of 55% and 30% [arxiv:1505.03046 §III-E]. Card: [Improving Computer-aided Detection using Convolutional Neural Networks and Random View Aggregation](/papers/1505.03046).

### Encoder-decoder with side-output supervision for saliency

A VGG16 encoder with its last pooling layer removed feeds three decoders that
start from conv3-3, conv4-3 and conv5-3 and upsample by deconvolution back to
the input size [arxiv:1705.02544 §3.3.1] [arxiv:1705.02544 §3.3.2]. Each side
output has its own pixel-wise cross-entropy loss, and the fused map has one too
[arxiv:1705.02544 eq. 7] [arxiv:1705.02544 eq. 8]. A learned fusion layer
combines the three side maps [arxiv:1705.02544 §3.2]. Trained on 10,000 SALICON
images, it is reported as best on most metrics for four of five eye-tracking
benchmarks, but not on MIT300 [arxiv:1705.02544 §4.3] [arxiv:1705.02544 tab. 4].
On TORONTO, dropping the side-output losses lowers s-AUC from 0.76 to 0.71
[arxiv:1705.02544 tab. 9]. Card: [Deep Visual Attention Prediction](/papers/1705.02544).

### Joint distribution adaptation for a CNN on unlabelled target data

A CNN is trained on labelled source data, then adapted on unlabelled target
data. The penalty has a marginal term and one term per class, and the class
terms use target pseudo-labels from the network [arxiv:1804.07265 §III-A]
[arxiv:1804.07265 eq. 10]. The authors report 97.5% average accuracy over ten
fault-diagnosis transfer tasks, against 87.4% with marginal-only adaptation and
80.0% for a plain CNN [arxiv:1804.07265 tab. III]. The base architecture is not
stated, and the penalty weight was chosen by a grid search whose validation data
is not named [arxiv:1804.07265 §III-B] [arxiv:1804.07265 §IV-A]. Card: [Deep Transfer Network with Joint Distribution Adaptation](/papers/1804.07265).

### Deep linear generator for per-image blur-kernel estimation

The generator has no activation functions: five hidden 64-channel convolutions
with kernel sizes 7, 5, 3, 1 and 1, then a 1 × 1 output with stride 2
[arxiv:1909.06581 §4.1] [arxiv:1909.06581 fig. 4]. Because the layers are
linear, their filters multiply into one kernel, which is the blur estimate
[arxiv:1909.06581 §4.2]. It is trained for 3,000 iterations against a patch
discriminator, on the one low-resolution image alone [arxiv:1909.06581 §2]. With
ZSSR as the super-resolution step, the authors report 30.363 dB at ×2 on
DIV2KRK, which uses synthetic blur, against 29.368 dB for the earlier estimator
[arxiv:1909.06581 tab. 1] [arxiv:1909.06581 §5.2]. Each image needs a full
training run of 61 s on a V100 [arxiv:1909.06581 §5.3]. The official code is
for academic use only [code: LICENSE.txt#L7], so a commercial version has to be
written from the paper. Card: [Blind Super-Resolution Kernel Estimation using an Internal-GAN](/papers/1909.06581).

### Six-dimensional convolution over putative correspondences

Each putative match between two point clouds is a point in six dimensions: the
coordinates of its two ends [arxiv:2004.11540 §3.2]. A 6-D convolutional network
with residual blocks gives each match an inlier probability [arxiv:2004.11540
fig. 2]. A weighted Procrustes solve then gives a closed-form pose from those
weights [arxiv:2004.11540 §3.3]. The per-point features come from a separate
descriptor network [arxiv:2004.11540 §3.1]. On 3DMatch, the authors report 91.3%
recall against 74.9% for RANSAC with eight million iterations, at 1.21 s per pair
including feature extraction [arxiv:2004.11540 tab. 1]. The paper leaves the
weight clip threshold and the refinement optimiser unstated [arxiv:2004.11540 §4.2].
Card: [Deep Global Registration](/papers/2004.11540).

### Text CNN with global average pooling, compared with an LSTM

Each sentence is a sequence of 300-dimensional word embeddings over a 15,000-word
vocabulary. Four one-dimensional convolutions, one pooling step, a global average
pool and a sigmoid output follow [arxiv:2006.03541 §4.4] [arxiv:2006.03541 fig. 9].
The authors say average pooling works better than max pooling, without giving
numbers [arxiv:2006.03541 §4.4]. With word embeddings, the CNN reaches accuracy
0.8006 on Sentiment140, against 0.8282 for the LSTM [arxiv:2006.03541 tab. 2].
Processing time on Sentiment140 with embeddings is 7 min 3 s for the CNN and
1 h 4 min for the LSTM [arxiv:2006.03541 tab. 7] [arxiv:2006.03541 tab. 8].
Card: [Sentiment Analysis Based on Deep Learning: A Comparative Study](/papers/2006.03541).

### Three-block CNN on face and speech features with a sarcasm check

The face branch uses three blocks, each a convolution, batch normalisation, ReLU,
2 × 2 max pooling and dropout [arxiv:2605.30099 §III.D]. A separate CNN reads
audio features such as MFCCs, chroma and a mel spectrogram [arxiv:2605.30099
§III.E]. For each 5 to 8 s sub-video, the top three emotions from each branch are
compared, and the clip is flagged as sarcasm when no emotion is shared
[arxiv:2605.30099 §III.E] [arxiv:2605.30099 alg. 1]. The authors report 95.7%
accuracy for the CNN with this check, against 70.14% to 95.6% for the compared
models [arxiv:2605.30099 tab. 3]. The evidence is thin: the sarcasm claim rests
on one public video [arxiv:2605.30099 §III.F], and the per-class accuracies
differ between the paper's Table 1 and Table 2 [arxiv:2605.30099 tab. 1]
[arxiv:2605.30099 tab. 2]. Card: [Evaluation of Conversational Agents: Understanding Culture, Context and Environment in Emotion Detection](/papers/2605.30099).

### Small CNN pixel classifier on RGB patches

Each pixel is classified from an RGB patch of 5 × 5 to 9 × 9 pixels centred on
it. The CNN has one convolution layer with five 3 × 3 filters and ReLU, followed
by max pooling [arxiv:2609.40212 §2.2.2]. Networks were selected on a validation
split of the training site and tested on a second site [arxiv:2609.40212 §2.2.2].
On the training site the CNN was best, but on the second site the linear and MLP
networks were slightly better [arxiv:2609.40212 §3.2]. The best CNN reaches MCC
0.79 on the second site, against 0.82 for a linear network on 9 × 9 patches and
0.61 for the Excess Blue index threshold [arxiv:2609.40212 tab. 3]
[arxiv:2609.40212 tab. 2]. Card: [Recognition of Urbanized Areas in UAV-Derived Very-High-Resolution Visible-Light Imagery](/papers/2609.40212).

### Window-based sentence CNN with max pooling over positions

A filter of width h slides over every run of h consecutive word embeddings.
Max pooling over positions leaves one value per filter, so the output length
does not depend on the sentence [arxiv:1708.02709 §3.1.1]. The survey notes
that typical models use hundreds of filters of several widths [§3.1.1]. For
word-level labels, the same window is centred on the word being labelled, and a
conditional random field can score the label sequence [§3.1.2]. The dynamic CNN
replaces fixed pooling with dynamic k-max pooling, which keeps the k strongest
features in their original order, so narrow filters can relate distant words
[§3.2]. The survey lists two weaknesses: CNNs need a lot of training data, and
they model long-range order poorly [§3.2]. Its benchmark tables copy results
from the cited papers, and a multichannel CNN is listed at 88.1% on SST-2
against 87.2% for a bidirectional LSTM [arxiv:1708.02709 tab. 6]. A
convolutional sequence-to-sequence model is listed at 26.36 BLEU on WMT14
English-German and 41.29 on English-French [arxiv:1708.02709 tab. 7]. The survey
runs no experiments and does not normalise these rows. Card: [Recent Trends in Deep Learning Based Natural Language Processing](/papers/1708.02709).

### Convolutional token embedding and depthwise projections in a Vision Transformer

A three-stage Vision Transformer. Each stage starts with an overlapping
convolution that turns the image, or the previous stage's token map, into
tokens and shrinks the grid [arxiv:2103.15808 §3.1]. Inside each block, queries,
keys and values come from depthwise 3 x 3 convolutions over the token grid, and
keys and values are subsampled with stride 2 to cut attention cost
[arxiv:2103.15808 §3.2] [arxiv:2103.15808 §3.2.2]. The model drops the position
embedding, because the convolutions supply local order [arxiv:2103.15808 §3.3].
CvT-13, with 20M parameters, reaches 81.6% ImageNet-1k top-1 at 224 px, against
81.8% for DeiT-B with 86M [arxiv:2103.15808 tab. 3]. Removing the convolutional
embedding costs 0.9 points in the authors' ablation, from 81.6 to 80.7
[arxiv:2103.15808 tab. 6]. Stride 1 on keys and values gives 82.3 at 6.55 GFLOPs,
against 81.6 at 4.53 GFLOPs for stride 2 [arxiv:2103.15808 tab. 7]. The
position-embedding ablation is ImageNet-only, and it shows no loss from
removing the embedding [arxiv:2103.15808 tab. 5] [arxiv:2103.15808 §4.4]. The
fine-tuning learning rate is not stated [arxiv:2103.15808 §4.1]. Card: [CvT](/papers/2103.15808).

### Per-window attention for a k x k convolution

RFAConv replaces a k x k convolution. It expands each window into k² slots,
learns a softmax over the slots of each channel at each position, multiplies
that by the expanded features, rearranges each window into a k x k block, and
finishes with a stride-k convolution [arxiv:2304.03198 §3.3] [arxiv:2304.03198 eq. 3].
The attention is computed per window, so the weights are not shared across
overlapping windows the way a spatial attention map is [arxiv:2304.03198 §3.1]
[arxiv:2304.03198 §3.2]. On ImageNet-1k, ResNet18 top-1 rises from 69.59% to
71.23% for 0.16M extra parameters [arxiv:2304.03198 tab. 3]. The authors'
RFCBAMConv variant reaches 72.15% [arxiv:2304.03198 tab. 4]. On COCO2017 with
YOLOv5n, AP rises from 27.5 to 29.0 for RFAConv and to 29.5 for RFCAConv
[arxiv:2304.03198 tab. 7]. The authors state the memory cost: the expanded
features are k² times the input size, and some devices may not afford it
[arxiv:2304.03198 §4.6]. They also say that replacing many convolutions slows
inference, so the count should be limited when speed matters [arxiv:2304.03198 §4.3].
Card: [RFAConv](/papers/2304.03198).

### Time series re-encoded as images for a tiled CNN

A one-dimensional series is turned into three square images: a Gramian angular
summation field, a Gramian angular difference field and a Markov transition
field. The three are stacked as channels [arxiv:1506.00327 §2] [arxiv:1506.00327 §3.3].
A two-layer tiled CNN, pretrained with topographic ICA, feeds a linear SVM
[arxiv:1506.00327 §3.2] [arxiv:1506.00327 fig. 3]. Across 20 UCR datasets the
authors report the lowest error on 9 of them [arxiv:1506.00327 tab. 1]. The
pretraining uses all series, test inputs included, without labels, so the
reported test errors are transductive [arxiv:1506.00327 §3.2]. The same angle
image can be inverted on its diagonal, and a denoising autoencoder on it lowers
interpolation error at the missing points, by 12.18% to 48.02% in the authors'
range, against the raw series [arxiv:1506.00327 §4.2] [arxiv:1506.00327 tab. 2].
Card: [Imaging Time-Series to Improve Classification and Imputation](/papers/1506.00327).

### Small CNN scoring projected grasp images

A two-finger grasp candidate is encoded as a 15-channel 60 x 60 image. Three
orthogonal projections each give a mean normal, an occupied-depth map and an
occluded-depth map [arxiv:1706.09911 §5.1]. A LeNet-style network with two
convolution and pooling stages, a 500-unit layer and a two-class softmax scores
it [arxiv:1706.09911 §5]. The training set is about 1.5 million labelled
candidates from 55 BigBird objects, scanned from 20 views each
[arxiv:1706.09911 §6]. The robot trials use the three-channel form, and the
occluded channels add about 2 accuracy points at roughly double the encoding
time [arxiv:1706.09911 §7.2] [arxiv:1706.09911 §8.6]. The paper's dense-clutter
result is 93% (268 of 288), and 53% without the classifier [arxiv:1706.09911 tab. 3].
Card: [Grasp Pose Detection in Point Clouds](/papers/1706.09911).

### Grasp classifier with occlusion channels and CAD pretraining

The conference version of the same grasp pipeline. Each candidate is voxelised
into a 60 x 60 x 60 grid and projected into 15 channels [arxiv:1603.01564 §3.3].
Adding the occlusion channels raises recall at 99% precision from 82% to 89% on
the 55 BigBird objects [arxiv:1603.01564 §4.2] [arxiv:1603.01564 fig. 5].
Pretraining on rendered CAD depth images reaches at 4000 iterations what a
random start reaches at 20000 [arxiv:1603.01564 §4.3]. The authors report 93%
grasp success from 288 active-sensing attempts, against 73% in their prior work
[arxiv:1603.01564 §5.4]. The labels assume frictionless antipodal grasps
[arxiv:1603.01564 §3.4], which the released code relaxes, so check the card's
Paper vs code section before reimplementing. Card: [High precision grasp pose
detection in dense clutter](/papers/1603.01564).

### Residual encoder-decoder regressing a position map

### Residual encoder-decoder regressing a position map

One RGB face crop goes through one convolution and ten residual blocks to an
8 x 8 x 512 bottleneck, then 17 transposed convolutions produce a 256 x 256
map [arxiv:1803.07835 §3.2]. Each map cell stores the x, y and z position of a
fixed point on the face, occluded points included, so one output gives landmarks,
dense correspondence and shape [arxiv:1803.07835 §3.1]. The loss weights the
cells 16 : 4 : 3 : 0 for landmarks, eyes, nose and mouth, other face, and neck
[arxiv:1803.07835 §3.2] [arxiv:1803.07835 tab. 3]. Training uses Adam at 0.0001,
batch size 16, and 300W-LP with 3DMM-based targets [arxiv:1803.07835 §3.3]. On
AFLW2000-3D the authors report mean NME 3.62, against 4.49 for 3DSTN and 5.42
for 3DDFA [arxiv:1803.07835 tab. 1]. The network runs in 9.8 ms per image on a
GTX 1080 [arxiv:1803.07835 §4.4]. Card: [Joint 3D Face Reconstruction and Dense Alignment with Position Map Regression Network](/papers/1803.07835).

### Modernised pure ConvNet with large depthwise kernels

The ConvNeXt block is a 7 x 7 depthwise convolution, a LayerNorm, a 1 x 1
expansion by four, one GELU, a 1 x 1 projection, a learned per-channel scale
and the residual add [arxiv:2201.03545 §2.4] [arxiv:2201.03545 §2.5]
[arxiv:2201.03545 §2.6] [arxiv:2201.03545 fig. 4]. It is reached from a
ResNet-50 by a sequence of recipe and design changes, each kept only if it
improved ImageNet-1K accuracy [arxiv:2201.03545 §2] [arxiv:2201.03545 tab. 10].
Downsampling uses separate 2 x 2 stride-2 convolutions, each with a LayerNorm,
and the stem is a 4 x 4 stride-4 patchify layer [arxiv:2201.03545 §2.2]
[arxiv:2201.03545 §2.6]. ConvNeXt-T reaches 82.1% top-1 at 224 px, against
81.3% for Swin-T [arxiv:2201.03545 tab. 1]. Card: [A ConvNet for the 2020s](/papers/2201.03545).

### Sentence CNN as the encoder of a dialog-act classifier with history

Each short text becomes one vector from a convolutional encoder. Filters of
height 3 slide over its word vectors, followed by ReLU and max pooling over
positions [arxiv:1603.03827 §2.1.2] [arxiv:1603.03827 tab. 2]. A second stage
reads the vector of the current text together with the vectors of earlier
texts, so the label can depend on the preceding turns [arxiv:1603.03827 §2.2].
The headline CNN accuracies are 65.5% on DSTC 4, 84.6% on MRDA and 73.1% on
SwDA, against 66.2%, 84.3% and 69.6% for the LSTM variant
[arxiv:1603.03827 tab. 4]. The authors say history at the vector level helps
more than history at the label level [arxiv:1603.03827 §4]. For DSTC 4 and
SwDA the paper gives only train and test splits, but early stopping needs a
validation set [arxiv:1603.03827 §3.1] [arxiv:1603.03827 §3.2]. Card:
[Sequential Short-Text Classification with Recurrent and Convolutional Neural Networks](/papers/1603.03827).

### Two-dimensional CNN with temporal difference modules for video

A 2D ResNet runs on a few frames sampled from each video, and two small modules
add motion to it. The short-term module adds a feature computed from RGB
differences between neighbouring frames to the early stages, and the long-term
module compares features of neighbouring sampled segments and uses the result as
a gate in later stages [arxiv:2012.10071 §3.2] [arxiv:2012.10071 §3.3]. Adding
both modules to a plain 2D baseline changes the top-1 on Something-Something V1
from 46.6% to 52.3% in the authors' ablation [arxiv:2012.10071 tab. 1]. With
8 frames on ResNet-50 the authors report 52.3% top-1 on that dataset at 36
GFLOPs, and 76.6% top-1 on Kinetics-400 with 10 clips and 3 crops
[arxiv:2012.10071 tab. 2] [arxiv:2012.10071 tab. 3]. The released code differs
from the paper's equations in several places, so the equations alone will not
reproduce the numbers; see Contradictions. Card:
[TDN: Temporal Difference Networks for Efficient Action Recognition](/papers/2012.10071).

### Continuous convolution with an MLP kernel over point clouds

The kernel is a small MLP of the relative 3D offset between two points, so the
layer works on unevenly spaced points and can produce output at new positions
[arxiv:2101.06742 §3.1] [arxiv:2101.06742 §3.2]. Factorising the weight tensor
keeps the stored kernel values to one per point pair and output channel, not
one per input channel as well [arxiv:2101.06742 §3.3]. On indoor scans the authors
report 58.27 mIOU, against 48.92 for SEGCloud [arxiv:2101.06742 tab. 1]. On
driving data the model on its own scores below the 3D-FCN baseline, at 46.35
mIOU against 49.28, and the best driving result comes from combining it with a
3D-FCN [arxiv:2101.06742 tab. 2]. The driving and flow data are in-house, and
no official code was found [arxiv:2101.06742 §4.2]. Card:
[Deep Parametric Continuous Convolutional Neural Networks](/papers/2101.06742).

### Local two-dimensional self-attention in a ResNet-like backbone

Local attention takes the place of the spatial convolution in each bottleneck.
The image is cut into blocks, and every block attends to one shared window that
extends a few pixels beyond it, with no mask [arxiv:2103.12731 §3.2]. The
windows are not centred on each pixel, so the layer is equivariant only to
shifts that are multiples of the block size [arxiv:2103.12731 §3.2].
Downsampling is a strided attention layer, which replaces attention followed by
pooling [arxiv:2103.12731 §3.2]. Under the same
recipe, HaloNet-50 reaches 80.5% top-1 against 79.5% for ResNet-50
[arxiv:2103.12731 tab. 3]. The largest model, H7, reaches 84.9% top-1 when
trained from scratch [arxiv:2103.12731 §4.1], and pure attention models train
more slowly than the matched EfficientNets on the authors' hardware
[arxiv:2103.12731 §5]. Keeping convolutions in the first two stages cut the
normalised training time from 1.9 to 1.0, at 84.7% against 84.9%
[arxiv:2103.12731 tab. 4]. Card:
[Scaling Local Self-Attention for Parameter Efficient Visual Backbones](/papers/2103.12731).

### Dual colour-space enhancer with an attention blend

Three blocks for underwater photos. An RGB block of eight 3 x 3 convolutions
removes the colour cast. Its output is converted to HSV, where a second block
predicts the knots of four piecewise-linear curves for brightness and
saturation [arxiv:2103.07138 §III-A] [arxiv:2103.07138 eq. 1]. An attention
block sees the raw image and both outputs and predicts per-pixel weights that
blend the two results [arxiv:2103.07138 §III-A]. Training uses L1, SSIM, an
HSV loss and a VGG-19 perceptual loss [arxiv:2103.07138 §III-B]
[arxiv:2103.07138 eq. 2]. Author-reported PSNR on the real-world set is 24.57 dB,
against 21.30 dB for UGAN, the next highest PSNR in the table
[arxiv:2103.07138 tab. II]. The baselines were run with the authors' pretrained
weights, not retrained [arxiv:2103.07138 §IV-A]. The released code differs
from the paper in several layers and losses, so read the card's Paper vs code
section before implementing. Card:
[UIEC^2-Net](/papers/2103.07138).

### Pre-trained CNN features from sentiment, emotion and personality, merged for sarcasm

Three CNNs trained on other labelled data (sentiment, six-way emotion and five
binary personality traits) produce fully connected features. These are
concatenated with the features of a baseline sarcasm CNN, giving a 1,100-value
vector for an SVM. The paper calls this CNN-SVM [arxiv:1610.08815 §4.1]
[arxiv:1610.08815 §5.3]. Whether the pre-trained extractors are frozen or
fine-tuned is not stated [arxiv:1610.08815 §4.2]. With all feature groups the
author-reported F1 is 97.71%, 94.80% and 93.30% on three Twitter sarcasm sets,
against 97.60%, 92.32% and 92.20% for the baseline CNN-SVM [arxiv:1610.08815 tab. 2].
Training on one set and testing on another can fall sharply, to 33.05% in one
direction [arxiv:1610.08815 §5.6]. Card: [A Deeper Look into Sarcastic Tweets](/papers/1610.08815).

### One-dimensional CNNs as baselines for raw satellite series

Two 1-D convolutional networks, MS-ResNet and TempCNN, classify crop types from
13-band Sentinel-2 series of 70 dates. A Transformer and an LSTM are trained on
the same parcels [arxiv:1910.10536 §4.1] [arxiv:1910.10536 §4.2]. In a gradient
analysis of one example per class, the Transformer and LSTM show zero gradient
on cloudy dates, while the convolutional models show non-zero gradient on some
of them [arxiv:1910.10536 §8.3].
The author-reported kappa on raw data is 0.67 ± 0.02 for TempCNN and 0.69 ± 0.03
for MS-ResNet, against 0.71 ± 0.03 and 0.71 ± 0.01 for the Transformer and LSTM
on 23 classes [arxiv:1910.10536 tab. 1]. On the 12-class set the gap is wider:
0.71 ± 0.07 for TempCNN and 0.78 ± 0.02 for MS-ResNet, against 0.82 and 0.83
[arxiv:1910.10536 tab. 1]. Preprocessing raised every model's scores
[arxiv:1910.10536 §8.1]. Card: [Self-attention for raw optical Satellite Time Series Classification](/papers/1910.10536).

### CNN over the last four layers of a language-specific BERT

Each language gets its own BERT-base encoder. The outputs of its last four
layers are stacked as four input channels for a CNN. The CNN has 160 filters,
32 for each height from one to five tokens, followed by ReLU, global max-pooling
and a sigmoid output [arxiv:2007.13184 §4.4] [arxiv:2007.13184 fig. 1]. The
paper gives learning rate 2e-5 and 10 epochs. Its released code runs 6 epochs
and uses binary F1 for checkpoint selection, and several training settings in
the code are not in the paper [arxiv:2007.13184 §4.4]; see the card's Paper vs
code section. Card: [KUISAIL at SemEval-2020 Task 12](/papers/2007.13184).

### Convolutional network under homomorphic encryption

A standard ResNet-20 keeps its trained weights, and the operators are changed so
that the whole network runs on encrypted input. ReLU is a composition of
minimax polynomials, batch normalisation is a scalar multiply and add, a stride-2
convolution is a stride-1 convolution times a 0/1 window, and the softmax is
computed as exp(x/4) normalised [arxiv:2106.07229 §3.3] [arxiv:2106.07229 §3.4]
[arxiv:2106.07229 §3.7]. The network uses 1,149 bootstraps [arxiv:2106.07229 §3.5].
Author-reported agreement with the plaintext network is 98.67% on 75 images,
where the plaintext network reaches 89.33% and the encrypted one 90.67%
[arxiv:2106.07229 tab. 4]. One image takes about 14,694 seconds on a 112-core
CPU [arxiv:2106.07229 tab. 5]. The text does not make clear whether the encrypted
runs are full runs or simulated [arxiv:2106.07229 §4.1] [arxiv:2106.07229 §4.2].
Card: [Privacy-Preserving Machine Learning with Fully Homomorphic Encryption](/papers/2106.07229).

### Regularised CNN: input, internal and label methods

A survey groups recent CNN regularisers by the place where they act. Input
methods change the training images, among them Cutout, Mixup, CutMix and
AutoAugment [arxiv:2201.03299 §3]. Internal methods act inside the network,
among them Dropout variants, DropBlock, Shake-Shake and ShakeDrop [arxiv:2201.03299
§4]. Label methods soften the target, among them label smoothing, TSLA, SLS and
JoCoR [arxiv:2201.03299 §5]. The survey runs no experiments. Its compiled table
gives 10.94% error for PyramidNet with ShakeDrop and PBA on CIFAR-100, against a
17.01% baseline [arxiv:2201.03299 tab. 2], but the rows differ in architecture
and schedule, and the card flags printing errors in the table [arxiv:2201.03299
§7.2]. Card: [Avoiding Overfitting: A Survey on Regularization Methods for
Convolutional Neural Networks](/papers/2201.03299).

### Reported CNN use in mineral-exploration remote sensing

A review of machine learning for mapping rock units, alteration, structures and
minerals from satellite, airborne and ground-based data. It places convolutional
networks in its deep-learning group and names lithology and mineral mapping as
their uses, citing UAV, Landsat and AVIRIS studies [arxiv:2103.07678 §5.5]
[arxiv:2103.07678 tab. 2]. The authors say few studies apply deep learning to
mineral exploration [arxiv:2103.07678 §5.5]. The review runs no experiments, so
its few numbers come from cited studies [arxiv:2103.07678 §6]. Card: [A review
of machine learning in processing remote sensing data for mineral
exploration](/papers/2103.07678).

## When to use which

| Situation | Use | Why |
|---|---|---|
| Dense per-pixel labels, few labelled volumes, need object outlines | Encoder-decoder with residual blocks and summation skips | Beat U-Net on two microscopy datasets [arxiv:1612.05360 tab. 2] [arxiv:1612.05360 tab. 3]. Results rest on single test volumes [arxiv:1612.05360 §4.2], and memory and runtime are not reported [arxiv:1612.05360 §3.3] |
| Large images with small objects and few labelled images, objects must not be resized | Fully convolutional segmentation with object-level voting | 93.94% object accuracy on 21 images [arxiv:1706.06480 tab. 1]. One train/test split, one run, and the headline denominator does not reconcile with the printed counts [arxiv:1706.06480 §5.1] [arxiv:1706.06480 tab. 4] |
| Image classification with limited labels | CNN trained with augmentation | On eight-class BreakHis at 40x, augmentation raised image-level accuracy from 95.69% to 97.09% [arxiv:1811.04241 tab. 3]. The card lists trial and epoch conflicts [arxiv:1811.04241 §4] |
| Two imaging modalities, sparse anatomical labels only, real-time registration wanted | 3-D CNN predicting a displacement field, trained with multiscale Dice | Median landmark error 3.6 mm and Dice 0.87 under cross-validation [arxiv:1807.03361 tab. 1]. Pairwise intensity-based registration on the same data was worse [arxiv:1807.03361 tab. 2]. No separate held-out test set was available [arxiv:1807.03361 §5] |
| Fixed-length sensor sequences, few features to engineer | 1-D CNN with an ensemble | The ensemble beat a single model by 5 points [arxiv:1804.02386 tab. 4]. The simplest CNN in the authors' table (75.6%) scored below the random forest (78.1%) [arxiv:1804.02386 tab. 6] |
| Sequence or string inputs for a pairwise prediction task, with a library at hand | 1-D CNN encoders with a shared MLP head | The authors' CNN+CNN pair is among the top scores on DAVIS [arxiv:2004.08919 fig. 1]. Training settings are not stated in the paper [arxiv:2004.08919 §3] |
| Short-horizon forecasting on a small compute budget | Depthwise-separable U-Net with attention on skips | About a quarter of U-Net's parameters, with MSE equal to U-Net's on NL-50 [arxiv:2007.04417 tab. 4]. The dataset is available only on request [arxiv:2007.04417 §4.1] |
| Quick binary classifier on small images, no baseline required | Two-layer CNN with 32 filters | Reaches about 93.5% on its own dataset [arxiv:2303.14126 tab. 7]. Training settings such as optimiser and batch size are not stated [arxiv:2303.14126 §3.3] |
| Small image classifier on a single-board computer, where memory is the binding limit | Two-convolution CNN at 96 x 96, converted with TensorFlow Lite | Converted file is 15,942,712 bytes against 47,864,936 before conversion, with 96.9% against 98.1% on different test sets [arxiv:2110.00791 tab. 2]. No latency or memory is measured on the board, and the conversion precision is not stated [arxiv:2110.00791 §3] |
| Several imaging sites that differ by modality | Clustered federated learning on a VGG-style backbone | Overall F1 gains over conventional federated learning on both modalities [arxiv:2101.07511 tab. 3]. Experiments are simulated, with single runs [arxiv:2101.07511 §4] |
| Large data, global context wanted | Local attention at stage ends of a CNN | Attention in the last stage improved the authors' CIFAR-100 ablation [arxiv:2202.06709 §4.1]. The authors report that local-attention models beat CNNs on ImageNet [arxiv:2202.06709 app. C]. Set the window size explicitly, since the paper's 4 x 4 and the code's default of 7 disagree [arxiv:2202.06709 app. E] |
| Pretrained scene features for a downstream task with a small labelled set | Places-trained CNN as a fixed extractor with a linear SVM | Places-trained features average 69.92% over eight benchmarks [arxiv:1610.02055 tab. 3]. On Places205 and SUN205, the paper's ImageNet-AlexNet feature with an SVM scores below its Places-trained models [arxiv:1610.02055 tab. 1] |
| Real-time detection on a laptop or phone CPU, where accuracy can be modest | Seven-layer detector from Tiny-YOLOv2 | 21 FPS on one non-GPU laptop, against 2.4 FPS for Tiny-YOLOv2 [arxiv:1811.05588 tab. III]. Speed comes from one device, and the batch-norm ablation gap for this trial is 1.0 point from one run each [arxiv:1811.05588 tab. III] |
| Multi-channel robot force or joint signals, where the next few seconds of task context must be predicted | CNN features feeding an LSTM forecaster, with a classifier on the forecast | The authors' six-task classifier reaches 89% accuracy on their second dataset, against 86% for the single stream [arxiv:2402.10220 §IV-C]. There is no external baseline, the sets are small, and the splits are random [arxiv:2402.10220 §IV-B] |
| Compressing a CNN for a custom accelerator where memory traffic dominates, and under 0.5 points of top-5 accuracy can be spared | Vector (1-D) pruning of convolution weights | Under 0.5 points below fine-grained top-5 at matched density on all five models [arxiv:1705.08922 tab. 1]. About 30 to 35% fewer output memory references in a simulation [arxiv:1705.08922 tab. 3]. Simulation only, with no measured hardware speed-up |
| Highest accuracy at a given sparsity, with no hardware constraint on the sparse pattern | Fine-grained (scalar) pruning | Highest matched-density top-5 for all five models [arxiv:1705.08922 tab. 1]. The paper's storage ratio at baseline accuracy is lowest for AlexNet and DenseNet-121 [arxiv:1705.08922 tab. 2] |
| Small labelled image set, fixed backbone, little compute for training | ResNet-18 with classifier-only tuning | Highest accuracy density under classifier-only tuning [arxiv:2605.11989 §5.1.2] [arxiv:2605.11989 §5.2.2]. No numeric values in the text, and validation and test splits are not defined [arxiv:2605.11989 §3.3] |
| Small labelled image set, all layers may be retrained | SqueezeNet with full tuning | Highest average accuracy density under full tuning [arxiv:2605.11989 §5.1.1]. The paper's ranking is read from a figure without value labels [arxiv:2605.11989 fig. 3] |
| Predicting which solution gradient descent picks in a linear model of depth L | Read the depth-L result, not a design choice | Stated for full-width circular convolutions with exponential loss on separable data [arxiv:1806.00468 §3]. For L > 2 the result is a stationary point, not a global optimum, and finite filters, pooling and nonlinearities are not covered [arxiv:1806.00468 §3] [arxiv:1806.00468 §6] |
| Text where long-range word order matters, or little labelled text | A plain sentence CNN is a weak fit | The survey lists both weaknesses [arxiv:1708.02709 §3.2]. Its tables copy results from papers with different setups and do not normalise them [arxiv:1708.02709 tab. 6] |
| Underwater photos with colour cast, low contrast or dim brightness, where the output is an enhanced image | Dual colour-space enhancer with an attention blend | Author-reported real-world PSNR of 24.57 dB, against 21.30 dB for UGAN [arxiv:2103.07138 tab. II]. Full-reference scores on one 90-image real set, with baselines run as pretrained [arxiv:2103.07138 §IV-A]. The released code differs from the paper in several places, so check the card's Paper vs code section first |
| Image backbone that needs convolutional locality inside a Transformer-style stack | Convolutional token embedding and depthwise projections (CvT) | 81.6% ImageNet-1k top-1 at 20M parameters, against 81.8% for DeiT-B at 86M [arxiv:2103.15808 tab. 3]. Ablations are on ImageNet only [arxiv:2103.15808 §4.4], and hardware is not stated [arxiv:2103.15808 §4.1] |
| Replacing the k x k convolutions of a CNN, with memory headroom for k² times the input | RFAConv or RFCAConv | ResNet18 top-1 69.59% to 71.23% for 0.16M extra parameters [arxiv:2304.03198 tab. 3]. Memory and inference cost rise [arxiv:2304.03198 §4.6] [arxiv:2304.03198 §4.3] |
| Sentence-level labels from word vectors, with no pretrained encoder available | Text CNN with two 256-filter convolution pairs | 0.68 accuracy against 0.28 to 0.51 for the authors' doc2vec, mean-embedding and bag-of-words baselines [arxiv:1704.06841 fig. 2]. The baselines are doc2vec, mean embeddings and bag-of-words [arxiv:1704.06841 §4], and the dataset is not released [arxiv:1704.06841 §3] |
| User-level text with many posts per user and binary labels | FastText or a linear model before a CNN | FastText is best on 7 of 9 binary conditions, and the CNN is best only on eating [arxiv:1806.05258 tab. 4]. The CNN sees a sample of each user's posts [arxiv:1806.05258 §5] |
| ImageNet-scale training where end-to-end backpropagation is too costly in memory | Greedy layer-wise CNN with auxiliary classifiers | Three-hidden-layer blocks reach 88.7% top-5, against 88.0% for end-to-end VGG-11 [arxiv:1812.11446 tab. 3]. End-to-end ResNet-152 reaches 94.1% top-5 in the same table, and the memory gain is shown only on CIFAR-10 [arxiv:1812.11446 §4.1] |
| City-scale grid forecasting, with a plain CNN as the first baseline | Grid CNN baseline, then ST-ResNet or STDN | Single-step TaxiBJ RMSE 23.550 for the CNN, against 18.702 for ST-ResNet and 17.826 for STDN [arxiv:2108.09091 tab. 4]. The authors find no single dominant grid model [arxiv:2108.09091 §5.2] |
| Matching street or drone images to satellite tiles with one encoder | Shared ConvNeXt-B encoder with symmetric InfoNCE | R@1 75.81 against 69.89 for the authors' ViT variant on VIGOR same-area [arxiv:2303.11851 tab. 4]. Hard-negative sampling is a separate factor in the same paper, and its gain is not isolated here [arxiv:2303.11851 §3.4] |
| Very deep plain stack that will not train, with no teacher network or two-stage schedule | Highway (gated skip) layers | A 19-layer convolutional highway net reaches 92.24% on CIFAR-10, against 91.61% for the 19-layer FitNet 4 that used a teacher [arxiv:1505.00387 tab. 1]. No seed variance is reported [arxiv:1505.00387 §3.1] |
| Only image-level labels, with lesion-level patches available for a first stage | Patch-trained CNN converted to a whole-image classifier | Per-image AUC 0.91 for a four-model ensemble on DDSM test [arxiv:1708.09427 §3.1.4]. The results rest on 376 test images and an older film-digitised set [arxiv:1708.09427 §3.1.1] |
| Road or binary masks in aerial images, with a small model wanted | Residual U-Net without cropping | Break-even 0.9187 against 0.9053 for U-Net on one dataset [arxiv:1711.10684 tab. II], with about 7.8M parameters against 30.6M [arxiv:1711.10684 §3.4]. One dataset and one value per model |
| Undersampled MRI where the measured k-space must be kept | Cascade of CNN blocks with data-consistency layers | MSE 0.89 x 10^-3 against 2.12 x 10^-3 for DLMRI at 3-fold [arxiv:1704.02422 tab. 1]. The timings are measured on the authors' GPU [arxiv:1704.02422 §7.2.3] |
| Small character images, benchmark-style classification | PreActResNet-18 with mixup if the training budget allows; small CNN otherwise | Kuzushiji-MNIST: 98.83% with manifold mixup against 95.12% for the small CNN [arxiv:1812.01718 tab. 1]. Baseline training settings are not given in the paper [arxiv:1812.01718 §3.1] |
| Multi-label chest radiographs with uncertain report phrases | Model uncertain labels explicitly, choosing the method per finding | Ignoring uncertain labels is biased where they are common: Consolidation has 12.78% uncertain against 6.78% positive [arxiv:1901.07031 tab. 1]. Validation AUROC differences are mostly not significant [arxiv:1901.07031 § Results] |
| Real-time detection in a family of model sizes, for orientation only | YOLO-style detector with C3k2 and C2PSA blocks | An overview with no measurements of its own, and no layer settings to build from [arxiv:2410.17725 §4] [arxiv:2410.17725 §6] |
| Large dense-prediction backbone, with a large pre-training budget | DCNv3 deformable 3 x 3 operator (InternImage) | 83.5 top-1 at 30M parameters, against 82.1 for ConvNeXt-T [arxiv:2211.05778 tab. 2]. Slower than ConvNeXt-B [arxiv:2211.05778 app. D]. The top result depends on a joint pre-training set that is not listed [arxiv:2211.05778 §4.1] |
| Few hundred labelled photos, fixed set of classes, pretrained AlexNet at hand | Fine-tune the last layers with SGD with momentum at 3e-4 | 98.35 ± 2.02% against 95.26 ± 3.04% for Adam [arxiv:2606.23286 tab. 1]. Scored on the validation split only [arxiv:2606.23286 §4] |
| Binary segmentation with few masks, and ImageNet classifier weights are available | U-Net with an ImageNet-pretrained VGG11 encoder | Validation IoU 0.686 against 0.593 from random initialisation, on 30 validation images with one run each [arxiv:1801.05746 §3]. Only one dataset is reported [arxiv:1801.05746 §3] |
| Classification where local detail and global layout both matter, and extra compute is acceptable | Parallel CNN and transformer branches with feature coupling | 83.4% top-1 for Conformer-S against 81.8% for DeiT-B [arxiv:2105.03889 tab. 2]. Detection runs slower than a ResNet-101 baseline [arxiv:2105.03889 app. C] |
| Phone or edge classifier where latency must be measured on the device | Pure CNN with separated token and channel mixers (RepViT-M1.0) | 80.0% at 1.0 ms with distillation, against 79.0% at 1.1 ms for EfficientFormerV2-S1 [arxiv:2307.09283 tab. 1]. Latency comes from one phone and Core ML, and the full architecture tables sit in supplementary material that was not read [arxiv:2307.09283 §3.5] |
| Second-stage false-positive filter for a 3-D detector, with few labelled patients and minutes per volume affordable | Random-view 2.5D classifier with mean aggregation | At 3 false positives per volume, lymph-node sensitivity is 70% (mediastinum) and 83% (abdomen), against prior results of 55% and 30% [arxiv:1505.03046 §III-E]. Reported sensitivities exclude candidate misses, because the training candidate list is made complete [arxiv:1505.03046 §III-E] |
| Dense per-pixel saliency from one labelled dataset, with a VGG-style backbone at hand | Encoder-decoder with side-output supervision | Best on most metrics for four of five benchmarks [arxiv:1705.02544 §4.3]. Side-output supervision adds 0.05 s-AUC on TORONTO [arxiv:1705.02544 tab. 9]. Not best on MIT300 [arxiv:1705.02544 tab. 4] |
| Labelled source data, unlabelled target data with the same classes, and a shift between operating conditions | CNN with joint distribution adaptation | 97.5% average against 87.4% for marginal-only adaptation and 80.0% for a plain CNN [arxiv:1804.07265 tab. III]. The data used for the penalty-weight search is not stated [arxiv:1804.07265 §IV-A] |
| Estimating one image's blur kernel for super-resolution, with no training set and minutes per image | Deep linear generator (KernelGAN) | With ZSSR, 30.363 dB at ×2 against 29.368 dB for the earlier estimator [arxiv:1909.06581 tab. 1]. 61 s per image on a V100 [arxiv:1909.06581 §5.3]. The benchmark uses synthetic blur [arxiv:1909.06581 §5.2] |
| Noisy correspondences between two 3-D scans, with a GPU available | 6-D convolution over correspondences with weighted Procrustes | 91.3% recall on 3DMatch against 74.9% for RANSAC with 8M iterations [arxiv:2004.11540 tab. 1]. On KITTI, FCGF alone has higher recall (98.2% against 96.9%) [arxiv:2004.11540 tab. 3] |
| Sentence-level sentiment with word embeddings, where training time matters | 1-D text CNN with global average pooling | Sentiment140 accuracy 0.8006 against 0.8282 for the LSTM, with 7 min against 64 min of processing time [arxiv:2006.03541 tab. 2] [arxiv:2006.03541 tab. 7] |
| Emotion from face and voice on the same clip, with a check for mismatch | Three-block CNNs with a cross-modal agreement check | 95.7% reported, against 70.14% to 95.6% for compared models [arxiv:2605.30099 tab. 3]. One run, and the sarcasm claim rests on one video [arxiv:2605.30099 §III.F] |
| Per-pixel classification of UAV RGB orthomosaics, where the site or season changes | Small CNN on RGB patches, or a linear network on 9 × 9 patches | On the second site, MCC 0.82 for the linear network and 0.79 for the CNN, against 0.61 for the Excess Blue threshold [arxiv:2609.40212 tab. 3] [arxiv:2609.40212 tab. 2]. Two sites only [arxiv:2609.40212 §4] |
| Time-series classification on UCR-style benchmarks, with budget for a per-dataset grid search | Gramian angular and Markov transition images with a tiled CNN | Lowest error on 9 of 20 datasets, author-reported [arxiv:1506.00327 tab. 1]. Each dataset needs 20 pretraining runs over the grid [arxiv:1506.00327 §3.2]. Baselines are copied from earlier papers and each result is one run [arxiv:1506.00327 §3] |
| Two-finger grasps from a depth point cloud, with no object model and novel objects | 15-channel grasp image with a LeNet-style scorer | 93% grasp success over 288 clutter attempts with the three-channel form, and 53% without the classifier [arxiv:1706.09911 tab. 3]. Three channels are 2 to 2.5 points less accurate than 15 and about 4 times faster [arxiv:1706.09911 §7.2] [arxiv:1706.09911 §8.4]. The selection cost is not in the released code [arxiv:1706.09911 §8.5] |
| Dense 3-D face shape and alignment from one image, with no fitting at test time | Residual encoder-decoder regressing a 256 x 256 position map | Mean NME 3.62 on AFLW2000-3D, against 4.49 for 3DSTN [arxiv:1803.07835 tab. 1]. 9.8 ms per image on a GTX 1080 [arxiv:1803.07835 §4.4]. Training code is not released, and the test-time crop is not stated [arxiv:1803.07835 §4] |
| Image backbone that should match a hierarchical Transformer without attention | ConvNeXt, a pure ConvNet with large depthwise kernels | ConvNeXt-T 82.1% against Swin-T 81.3% on ImageNet-1K at 224 px [arxiv:2201.03545 tab. 1]. Box AP ties Swin-T on COCO Cascade Mask R-CNN 3x [arxiv:2201.03545 tab. 3]. The top results use ImageNet-22K pretraining, and Swin's numbers come from Swin's own reports [arxiv:2201.03545 tab. 1] |
| Sentence or utterance sequences where a label depends on the preceding texts | Sentence CNN with a second-stage history window | Headline 65.5% on DSTC 4, 84.6% on MRDA and 73.1% on SwDA [arxiv:1603.03827 tab. 4]. The LSTM is ahead on DSTC 4 in the same table [arxiv:1603.03827 tab. 4]. The paper does not say how a validation split was made for two of the datasets [arxiv:1603.03827 §3.1] |
| Video action recognition where motion matters and per-video cost must stay near a 2D network's | 2D ResNet with short- and long-term temporal difference modules (TDN) | 52.3% top-1 on Something-Something V1 at 36 GFLOPs with 8 frames [arxiv:2012.10071 tab. 2]. The released code differs from the paper, so follow the code if you need to match the reported numbers [arxiv:2012.10071 §3.2] |
| Scattered 3-D points, output needed at arbitrary positions | Parametric continuous convolution (PCCN) | 58.27 mIOU on indoor scans against 48.92 for SEGCloud [arxiv:2101.06742 tab. 1]. The driving data is in-house and no official code was found [arxiv:2101.06742 §4.2] |
| Image backbone that needs attention at local cost, with non-centred windows acceptable | Local 2-D self-attention with haloing (HaloNet) | HaloNet-50 at 80.5% top-1 against 79.5% for ResNet-50 under the same recipe [arxiv:2103.12731 tab. 3]. Pure attention trains more slowly than the matched EfficientNets [arxiv:2103.12731 §5] |
| Short text with a sentiment or emotion model available to reuse, and only a few labelled examples | Pre-trained CNN features from related tasks, merged with a baseline CNN and classified by an SVM | With all feature groups, 97.71% F1 on Dataset 1 against 97.60% for the baseline CNN-SVM, and gains of 2.5 and 1.1 points on Datasets 2 and 3 [arxiv:1610.08815 tab. 2]. Transfer across datasets is weak in one direction [arxiv:1610.08815 §5.6]. Dataset sizes are in [arxiv:1610.08815 §5.1] |
| Raw, cloud-affected satellite series with no cloud masking or gap filling | A Transformer or LSTM rather than a 1-D CNN | The gap is clear on the 12-class raw set (kappa 0.82 to 0.83 against 0.71 to 0.78 for the CNNs). On 23 classes it is within one standard deviation [arxiv:1910.10536 tab. 1]. Preprocessing raised every model, so the model choice matters less on clean data [arxiv:1910.10536 §8.1] |
| Offensive-language tweets in a language with its own BERT model | BERT fine-tuned, with a CNN over its last four layers | Macro-F1 0.897, 0.843 and 0.814 for Arabic, Greek and Turkish, against 0.884, 0.822 and 0.816 for BERT alone [arxiv:2007.13184 tab. 2]. The average gain is 0.010, and Turkish is slightly lower. Single runs with no variance reported [arxiv:2007.13184 §5] |
| Private inference, where the server must not see the input and hours per image are acceptable | ResNet-20 under RNS-CKKS with bootstrapping and polynomial ReLU | 98.67% agreement with the plaintext network on 75 images [arxiv:2106.07229 tab. 4]. About 4 hours per image on a 112-core CPU [arxiv:2106.07229 tab. 5]. The authors name GPU, FPGA or ASIC accelerators as essential future work for run time, and report no measurements with them [arxiv:2106.07229 §5] |
| Plant or weed images from fields, few labels, and crop and weed look alike | Pre-trained CNN as a starting point, with a run without pre-training as a check | One study finds ImageNet weights still cut training iterations, another builds a network without pre-training. The survey does not settle this [arxiv:2103.01415 §10.1.1] [arxiv:2103.01415 §10.1.2]. Accuracies come from different setups and are not comparable [arxiv:2103.01415 §11] |
| Overfitting on a small labelled image set, with the choice of regulariser still open | Try one method from the input, internal or label family, and compare on your own data | The survey treats regularisation as a way to lower error on unseen data [arxiv:2201.03299 §1.2]. Its results come from original papers under different setups [arxiv:2201.03299 tab. 2] [arxiv:2201.03299 §7.2], so use them to pick a family, then read the original paper for the recipe [arxiv:2201.03299 §7] |
| Lithology or mineral mapping from satellite, airborne or ground-based spectra, where a CNN is one option | Compare a CNN with classical classifiers such as SVM and random forest on your own samples | The review names few deep-learning studies in this field and gives no CNN benchmark [arxiv:2103.07678 §5.5]. Its Table 2 lists sample studies with no selection rule [arxiv:2103.07678 tab. 2] |

On small datasets, the same authors report that a ResNet's built-in inductive
bias gives lower test error than a ViT at equal training loss [arxiv:2202.06709 app. C].
That is a single-study result, so check it on your own data.

## Implementation sketch

The shared core is a stack of convolution, nonlinearity and down-sampling,
followed by a classifier head. The variants add skips, branches or recurrence
inside the block. Kernel sizes and pool sizes differ by paper and are cited
where stated.

```pseudocode
# Common CNN core: convolution, nonlinearity, down-sampling  [arxiv:2303.14126 fig. 5]
# Input: images (B, H, W, C); output: one score per class (B, K)

def cnn_classifier(images, filters, dense_units, num_classes, k):
    x = images
    for f in filters:
        # Shared-weight convolution over a k-sized neighbourhood
        # k is paper-specific: 1 x 3 [arxiv:1804.02386 §3.2.2],
        # 1 x 1 and 3 x 3 [arxiv:1811.04241 §3], 3 x 3 x 3 in 3-D [arxiv:1807.03361 §2.3]
        x = relu(conv2d(x, filters=f, kernel=k))
        # Down-sample; pool size is not stated in arxiv:2303.14126
        x = max_pool(x)
    x = flatten(x)
    x = relu(dense(x, units=dense_units))
    # Softmax or sigmoid is applied by the loss  [arxiv:2303.14126 §3.3]
    return dense(x, units=num_classes)
```

```pseudocode
# Residual block: add the block input back to its output  [arxiv:1811.04241 eq. 4]
# Input and output have the same shape (B, C, H, W)

def residual_block(x, body):
    # Shortcut: the input passes through unchanged and is added
    return x + body(x)
```

```pseudocode
# Encoder-decoder skip joins  [arxiv:1612.05360 §3.1]
# The decoder adds the encoder map of the same level instead of concatenating

def decode_level(x, encoder_map):
    x = deconvolve(x, scale=2)
    # Pixel-wise sum; the paper contrasts this with U-Net's concatenation
    return x + encoder_map
```

```pseudocode
# Depthwise-separable convolution  [arxiv:2007.04417 §3.2]
# Per-channel spatial filter, then a 1 x 1 convolution that mixes channels

def dsc(x, channels_out, kernels_per_layer):
    # One spatial filter per output of each input channel
    x = depthwise_conv3x3(x, channels_out=kernels_per_layer * channels_in(x))
    # Pointwise 1 x 1 convolution to the target width
    return conv1x1(x, channels_out=channels_out)
```

```pseudocode
# Highway layer: gated mix of a transform and the input  [arxiv:1505.00387 §2, eq. 3]
# Input and output have the same shape  [arxiv:1505.00387 §2.1]

def highway(x, transform):
    # Gate in (0, 1); its bias starts negative so the layer carries x  [arxiv:1505.00387 §2.2]
    gate = sigmoid(gate_conv(x))
    # Blend the transform with the untouched input  [arxiv:1505.00387 §2, eq. 3]
    return transform(x) * gate + x * (1 - gate)
```

## Contradictions

None known between the cited papers on the same benchmark. Several cards flag
numbers that conflict inside a single paper, and these should be read with care:

- [CIFAKE](/papers/2303.14126) quotes 92.98% as its headline, while its final
  model scores 93.55% [arxiv:2303.14126 tab. 2] [arxiv:2303.14126 tab. 7].
- [Transportation modes from GPS](/papers/1804.02386) reports a confusion
  matrix whose printed recall and totals do not match its counts [arxiv:1804.02386 tab. 5].
- [IRRCNN](/papers/1811.04241) gives 3 trials in one section and 5 in another,
  and 50 epochs in the text against about 150 in its figures [arxiv:1811.04241 §4].
- [Advanced Steel Microstructural Classification](/papers/1706.06480) says
  augmentation adds about 2 points, while its own table shows a gain of about
  3 points [arxiv:1706.06480 §6] [arxiv:1706.06480 tab. 2].
- [YOLO-LITE](/papers/1811.05588) gives three VOC mAP values for its chosen
  network, 33.81%, 33.77% and 33.57%, in the abstract, a results table and the
  trial table [arxiv:1811.05588 tab. III] [arxiv:1811.05588 tab. V]. The
  trial table is the only place where the variant sits beside the others, so
  this page uses 33.57%.
- [A Taxonomy of Deep Convolutional Neural Nets](/papers/1601.06615) describes
  dropout as dropping neurons with probability p while scaling activations by p
  at test time. Those are two conventions, and the survey does not resolve them
  [arxiv:1601.06615 §2.3.2]. Check the original dropout paper before using it.
- [Highway Networks](/papers/1505.00387) numbers its layers inconsistently.
  The text maps Highway 1 and Highway 4 to FitNet 1 and FitNet 4, but the table
  names the FitNet 4 architecture "Highway 2" [arxiv:1505.00387 §3.2]
  [arxiv:1505.00387 tab. 1]. Use the table's labels.
- [Deep Learning to Improve Breast Cancer Early Detection](/papers/1708.09427)
  gives 0.88 in its abstract, which is the augmented value. The non-augmented
  value for the same model is 0.87 [arxiv:1708.09427 tab. 2].
- [InternImage](/papers/2211.05778) gives the stage-width search grid as 48, 64
  and 80 in the main text, but as 16, 32 and 64 in its appendix
  [arxiv:2211.05778 §3.2] [arxiv:2211.05778 app. B].
- [Transfer learning for ewaste recycling](/papers/2606.23286) describes the
  frozen layers two ways: all layers except the last three, and the fully
  connected layer. The two conflict [arxiv:2606.23286 §4].
- [TernausNet](/papers/1801.05746) gives the encoder as seven convolutions in
  its text and eight in its released code [arxiv:1801.05746 §2]. Implement
  the one you need, and check it against your own training run.
- [Conformer](/papers/2105.03889) gives Conformer-S as 83.4% in its ImageNet
  table but 83.5% with positional embeddings in its ablation table, and as
  10.6 GMACs in some tables and 10.3 in others [arxiv:2105.03889 tab. 2]
  [arxiv:2105.03889 tab. 6] [arxiv:2105.03889 tab. 8].
- [A Survey of Modern Deep Learning based Object Detection Models](/papers/2104.11892)
  gives AlexNet eight learnable layers in its text and seven in its backbone
  table [arxiv:2104.11892 §4.1] [arxiv:2104.11892 tab. 2].
- [RepViT](/papers/2307.09283) gives a stage ratio of 1:1:7:1 and depth
  2:2:14:2 in its text [arxiv:2307.09283 §3.3]. The released code builds
  3, 4, 16 and 3 blocks per stage for M0.9, and the paper does not say how it
  counts depth.
- [Improving Computer-aided Detection using Convolutional Neural Networks and Random View Aggregation](/papers/1505.03046) gives N = 100 views per candidate in total [arxiv:1505.03046 §III-A], and also 100 views at each scale [arxiv:1505.03046 §III-E]. The first reading gives 25 views per scale, and the second gives 400 across the four scales.
- [Deep Transfer Network with Joint Distribution Adaptation](/papers/1804.07265) says the results are almost all over 95% [arxiv:1804.07265 §IV-B], but its table lists 94.3% for F→E [arxiv:1804.07265 tab. III].
- [Deep Global Registration](/papers/2004.11540) says its method is about twice as fast as RANSAC with 2M iterations [arxiv:2004.11540 §5.1], while its table gives 1.21 s against 1.39 s [arxiv:2004.11540 tab. 1].
- [Sentiment Analysis Based on Deep Learning](/papers/2006.03541) says the CNN outperforms other models [arxiv:2006.03541 §6], but its table has the LSTM with word embeddings ahead on six of eight datasets [arxiv:2006.03541 tab. 2].
- [CvT](/papers/2103.15808) gives the convolutional-embedding ablation drop as 0.8 points in its text, while table 6 gives 0.9 [arxiv:2103.15808 §4.4] [arxiv:2103.15808 tab. 6]. Its text gives the key/value-stride drop as 0.3 points, while table 7 gives 0.7 [arxiv:2103.15808 §4.4] [arxiv:2103.15808 tab. 7].
- [Evaluation of Conversational Agents](/papers/2605.30099) gives 83.18% for Anger in its CNN-only table [arxiv:2605.30099 tab. 1] and 93.18% after the agreement step [arxiv:2605.30099 tab. 2], with no reason given for the change.
- [Imaging Time-Series to Improve Classification and Imputation](/papers/1506.00327) states a grid of image sizes 16, 24, 32, 40 and 48 and quantile counts 8, 16, 32 and 64, but Table 1 reports image sizes of 8 and 64 and quantile counts of 24, 40 and 48 [arxiv:1506.00327 §3.2] [arxiv:1506.00327 tab. 1]. This page does not choose between them.
- [A ConvNet for the 2020s](/papers/2201.03545) says the ResNet-200 regime stops gaining at 7 x 7 kernels [arxiv:2201.03545 §2.5], while its appendix says 5 x 5 [arxiv:2201.03545 app. C]. Table 11 gives 82.32 for 5 x 5 and 82.30 for 7 x 7 [arxiv:2201.03545 tab. 11].
- [Grasp Pose Detection in Point Clouds](/papers/1706.09911) says no selection strategy gives "18% more grasp failures" [arxiv:1706.09911 §8.5], while Table 3 gives 93% against 75% success [arxiv:1706.09911 tab. 3]. Those two statements do not give the same number.
- [Sequential Short-Text Classification](/papers/1603.03827) says the CNN beats the LSTM on every dataset [arxiv:1603.03827 §4]. Its grid ties them at 65.8 on DSTC 4, and its headline table gives the LSTM 66.2 against the CNN's 65.5 [arxiv:1603.03827 tab. 3] [arxiv:1603.03827 tab. 4].
- [The History Began from AlexNet](/papers/1803.01164) gives GoogLeNet's cost as 1.53G MACs in its text and 1.43G in its table [arxiv:1803.01164 §III.B.6] [arxiv:1803.01164 tab. II].
- [Towards Semi-Autonomous Robotic Arm Manipulation Operator Intention Detection from Force Data](/papers/2402.10220): the abstract gives F1 89% for task classification, but §IV-C reports the same 89% as accuracy for the quadruplet model. §V-A calls it F1 [arxiv:2402.10220 p. 1] [arxiv:2402.10220 §IV-C] [arxiv:2402.10220 §V-A]. The paper also gives the forecasting loss as RMSE in §III-G and as MSE in §IV-E [arxiv:2402.10220 §III-G] [arxiv:2402.10220 §IV-E].
- [TDN](/papers/2012.10071): the paper's long-term gate is an average of two sigmoid attentions, in (0, 1) [arxiv:2012.10071 eq. 2] [arxiv:2012.10071 eq. 6]. The released code shifts each sigmoid by −0.5, which gives a gate in (−0.5, 0.5), and reverses the sign of the temporal difference [code: ops/base_module.py#L89] [code: ops/base_module.py#L112-L116].
- [Scaling Local Self-Attention](/papers/2103.12731) gives a base learning rate of 1.6 in §4.1 and 0.1 in appendix C.3 [arxiv:2103.12731 §4.1] [arxiv:2103.12731 app. C.3].
- [A Deeper Look into Sarcastic Tweets](/papers/1610.08815) gives the Dataset 2 sentiment-feature F1 as 87.00% in §5.4, but its table gives 86.50% for that cell and 87.00% for Dataset 1 [arxiv:1610.08815 §5.4] [arxiv:1610.08815 tab. 2]. Its text also calls sentiment the best pre-trained model on Dataset 3, while the table places personality (74.41%) above sentiment (73.50%) [arxiv:1610.08815 §5.5] [arxiv:1610.08815 tab. 2]. The Ptacek et al. Dataset 3 F1 is 62.37% in the text and 63.37% in the table [arxiv:1610.08815 §5.5] [arxiv:1610.08815 tab. 3]. This page uses the tables.
- [Self-attention for raw optical Satellite Time Series Classification](/papers/1910.10536) divides attention scores by the square root of the hidden size in its equation, while its code divides by the square root of the per-head key size [arxiv:1910.10536 §3.4] [arxiv:1910.10536 eq. 4]. With eight heads the two differ by a factor of about 2.8 at a hidden size of 256.
- [KUISAIL at SemEval-2020 Task 12](/papers/2007.13184) gives 10 training epochs in the paper and 6 in its released code [arxiv:2007.13184 §4.4]. Reported numbers may come from either setting.
- [A Review on Deep Learning in UAV Remote Sensing](/papers/2101.10861) gives 190 Web of Science records in its text, while its per-type counts (136 articles, 46 proceedings and 10 reviews) sum to 192 [arxiv:2101.10861 §3]. The paper's totals for analysed studies are consistent with the per-type counts.

## Papers

- [Places](/papers/1610.02055): a scene-centric image dataset and the AlexNet, GoogLeNet and VGG scene CNNs trained on it.
- [FusionNet](/papers/1612.05360): residual encoder-decoder with summation skips for electron-microscopy segmentation.
- [Advanced Steel Microstructural Classification](/papers/1706.06480): fully convolutional segmentation of steel micrographs with per-object majority voting.
- [Transportation modes from GPS](/papers/1804.02386): one-dimensional CNN over GPS-derived motion channels, with a bootstrap ensemble.
- [IRRCNN](/papers/1811.04241): inception-style recurrent residual blocks for breast histopathology.
- [Weakly-Supervised Convolutional Neural Networks for Multimodal Image Registration](/papers/1807.03361): three-dimensional CNN that predicts a displacement field from sparse label pairs.
- [DeepPurpose](/papers/2004.08919): drug-target interaction library with CNN and other encoders for molecules and proteins.
- [SmaAt-UNet](/papers/2007.04417): small attention U-Net with depthwise-separable convolutions for precipitation nowcasting.
- [Collaborative federated learning for healthcare](/papers/2101.07511): VGG16-based COVID-19 classifier trained with clustered federated averaging.
- [How Do Vision Transformers Work?](/papers/2202.06709): analysis of self-attention and the AlterNet hybrid of CNN and local attention.
- [CIFAKE](/papers/2303.14126): small CNN that separates real CIFAR-10 images from Stable Diffusion outputs.
- [YOLO-LITE](/papers/1811.05588): seven-layer detector built from Tiny-YOLOv2 by trial and error, aimed at computers without a GPU.
- [Implicit Bias of Gradient Descent on Linear Convolutional Networks](/papers/1806.00468): theory of which separator gradient descent finds in a linear convolutional network of depth L.
- [Exploring the Regularity of Sparse Structure in Convolutional Neural Networks](/papers/1705.08922): magnitude pruning of CNN weights at four grain sizes, compared on ImageNet and in a dataflow simulation.
- [A Transfer Learning Evaluation of Deep Neural Networks for Image Classification](/papers/2605.11989): eleven pretrained CNNs compared by accuracy density on five small image sets.
- [A Taxonomy of Deep Convolutional Neural Nets for Computer Vision](/papers/1601.06615): survey of CNN building blocks, the AlexNet recipe and its main variants, with no experiments.
- [Medical Text Classification using Convolutional Neural Networks](/papers/1704.06841): text CNN over word2vec vectors that labels medical sentences into 26 categories.
- [SMHD](/papers/1806.05258): Reddit dataset of self-reported mental health diagnoses, with a text CNN baseline beside FastText and bag-of-words models.
- [Greedy Layerwise Learning Can Scale to ImageNet](/papers/1812.11446): CNN trained one block at a time with shallow auxiliary classifiers, without end-to-end backpropagation.
- [DL-Traff](/papers/2108.09091): survey and benchmark of urban traffic models, with a plain CNN and ST-ResNet among the grid baselines.
- [Sample4Geo](/papers/2303.11851): cross-view geo-localisation with one shared ConvNeXt-B encoder and hard-negative sampling.
- [Highway Networks](/papers/1505.00387): gated skip connections that let very deep plain stacks train.
- [Cascaded CNN for dynamic cardiac MRI](/papers/1704.02422): CNN blocks interleaved with data-consistency layers for undersampled cardiac MRI.
- [Brain Intelligence: Go Beyond Artificial Intelligence](/papers/1706.01040): position paper on brain-inspired AI, with no convolutional design content.
- [Deep Learning to Improve Breast Cancer Early Detection on Screening Mammography](/papers/1708.09427): patch-trained CNN converted to an all-convolutional whole-image mammogram classifier.
- [Road Extraction by Deep Residual U-Net](/papers/1711.10684): residual U-Net-style encoder-decoder for road masks in aerial images.
- [Deep Learning for Classical Japanese Literature](/papers/1812.01718): character datasets from cursive Japanese books, with CNN and PreActResNet-18 baselines.
- [CheXpert](/papers/1901.07031): chest radiograph dataset with uncertainty labels, and DenseNet121 models trained with four ways of handling them.
- [InternImage](/papers/2211.05778): CNN backbone built on DCNv3, a deformable 3 x 3 operator with shared projections.
- [YOLOv11: An Overview of the Key Architectural Enhancements](/papers/2410.17725): architecture review of a real-time detector with C3k2 and C2PSA blocks.
- [Transfer learning-based method for automated ewaste recycling in smart cities](/papers/2606.23286): AlexNet with its output layers fine-tuned on 650 smartphone-back photos.
- [Deep learning in agriculture: A survey](/papers/1807.11809): survey of 40 agricultural deep-learning studies and their CNN and transfer-learning choices.
- [TernausNet](/papers/1801.05746): U-Net with an ImageNet-pretrained VGG11 encoder for building masks in aerial images.
- [A Survey of Modern Deep Learning based Object Detection Models](/papers/2104.11892): survey of detectors and their backbones, with author-reported AP and FPS copied from source papers.
- [Conformer](/papers/2105.03889): parallel CNN and transformer branches joined by feature coupling units in every block.
- [Dive into Deep Learning](/papers/2106.11342): open textbook with a chapter on convolutional networks and runnable code in several frameworks.
- [RepViT](/papers/2307.09283): pure CNN built from MobileNetV3 with lightweight-ViT layout choices, latency measured on an iPhone 12.
- [Improving Computer-aided Detection using Convolutional Neural Networks and Random View Aggregation](/papers/1505.03046): random-view 2.5D CNN that scores detector candidates by averaging over views.
- [Deep Visual Attention Prediction](/papers/1705.02544): VGG16 encoder with three deconvolution decoders and side-output supervision for saliency.
- [Deep Transfer Network with Joint Distribution Adaptation](/papers/1804.07265): CNN adapted to unlabelled fault data with a joint distribution penalty.
- [Blind Super-Resolution Kernel Estimation using an Internal-GAN](/papers/1909.06581): deep linear convolutional generator that estimates one image's blur kernel.
- [Deep Global Registration](/papers/2004.11540): 6-D convolutional network that scores point-cloud correspondences for registration.
- [Sentiment Analysis Based on Deep Learning: A Comparative Study](/papers/2006.03541): DNN, CNN and LSTM sentiment classifiers compared on eight datasets.
- [Reading Race: AI Recognises Patient's Racial Identity In Medical Images](/papers/2107.10356): standard CNNs predict self-reported race from medical images, and probes find no single cause.
- [Evaluation of Conversational Agents: Understanding Culture, Context and Environment in Emotion Detection](/papers/2605.30099): three-block CNNs on face and speech features, with a mismatch check for sarcasm.
- [Recognition of Urbanized Areas in UAV-Derived Very-High-Resolution Visible-Light Imagery](/papers/2609.40212): vegetation-index thresholds compared with linear, MLP and CNN pixel classifiers.
- [Recent Trends in Deep Learning Based Natural Language Processing](/papers/1708.02709): survey of deep NLP, with a section on CNN sentence models and copied benchmark tables.
- [CvT: Introducing Convolutions to Vision Transformers](/papers/2103.15808): Vision Transformer with convolutional token embedding and depthwise convolutional projections.
- [RFAConv: Receptive-Field Attention Convolution for Improving Convolutional Neural Networks](/papers/2304.03198): drop-in k x k convolution with a per-window softmax over its k² slots.
- [Imaging Time-Series to Improve Classification and Imputation](/papers/1506.00327): GASF, GADF and Markov transition images of a time series, classified by a tiled CNN and used for imputation.
- [Joint 3D Face Reconstruction and Dense Alignment with Position Map Regression Network](/papers/1803.07835): residual encoder-decoder that regresses a 256 x 256 face position map in one pass.
- [Grasp Pose Detection in Point Clouds](/papers/1706.09911): samples two-finger grasp candidates from point clouds and scores 15-channel grasp images with a small CNN.
- [A ConvNet for the 2020s](/papers/2201.03545): ConvNeXt, a pure convolutional backbone built by modernising a ResNet-50.
- [Sequential Short-Text Classification with Recurrent and Convolutional Neural Networks](/papers/1603.03827): LSTM or CNN text encoder with a window over earlier texts, for dialog-act labels.
- [The History Began from AlexNet: A Comprehensive Survey on Deep Learning Approaches](/papers/1803.01164): survey of deep learning with a section on CNN building blocks and designs from LeNet to DenseNet.
- [TDN: Temporal Difference Networks for Efficient Action Recognition](/papers/2012.10071): 2D ResNet with short- and long-term temporal difference modules for video.
- [Deep Parametric Continuous Convolutional Neural Networks](/papers/2101.06742): continuous convolution whose kernel is an MLP over relative 3D positions, for scattered points.
- [Scaling Local Self-Attention for Parameter Efficient Visual Backbones](/papers/2103.12731): HaloNet, an image backbone built from blocked local self-attention with haloing.
- [Optimizing Neural Network for Computer Vision task in Edge Device](/papers/2110.00791): small gesture CNN at four input sizes, converted with TensorFlow Lite for a Raspberry Pi.
- [Towards Semi-Autonomous Robotic Arm Manipulation Operator Intention Detection from Force Data](/papers/2402.10220): CNN classifiers and a CNN+LSTM forecaster on glovebox teleoperation force and joint data.
- [UIEC^2-Net: CNN-based Underwater Image Enhancement Using Two Color Space](/papers/2103.07138): three-block CNN with RGB, HSV and attention branches for underwater images.
- [Machine learning based disease diagnosis: A comprehensive review](/papers/2112.15538): systematic review of ML and DL for diagnosis, 2012 to 2021, which reports CNNs as the most used model.
- [A Deeper Look into Sarcastic Tweets Using Deep Convolutional Neural Networks](/papers/1610.08815): three CNNs pre-trained on sentiment, emotion and personality, merged with a sarcasm CNN for an SVM.
- [Self-attention for raw optical Satellite Time Series Classification](/papers/1910.10536): Transformer compared with an LSTM, two 1-D CNNs and a random forest on raw and preprocessed Sentinel-2 series.
- [KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media](/papers/2007.13184): language-specific BERT with a CNN over its last four layers, for offensive tweets.
- [A Review on Deep Learning in UAV Remote Sensing](/papers/2101.10861): literature review of deep learning on UAV imagery, coding 232 studies by network, task, sensor and context.
- [A Survey of Deep Learning Techniques for Weed Detection from Images](/papers/2103.01415): survey of 70 deep-learning weed-detection papers, with their architectures, datasets and metrics.
- [Encrypted ResNet-20 inference with RNS-CKKS](/papers/2106.07229): ResNet-20 evaluated on encrypted CIFAR-10 images with RNS-CKKS bootstrapping and a polynomial ReLU.
- [High precision grasp pose detection in dense clutter](/papers/1603.01564): conference version of the grasp classifier, with occlusion channels and CAD pretraining.
- [A review of machine learning in processing remote sensing data for mineral exploration](/papers/2103.07678): review of machine learning for mineral mapping, with convolutional networks in its deep-learning group.
- [Avoiding Overfitting: A Survey on Regularization Methods for Convolutional Neural Networks](/papers/2201.03299): survey of input, internal and label regularisers for CNNs, with a compiled results table.