A Comprehensive Survey of Few-shot Learning: Evolution, Applications, Challenges, and Opportunities

In plain words

Most learning systems need many labelled examples before they can recognise something new. Often that is not possible: a factory camera sees only a few defective parts, or a sensor on a small device collects very little data. Few-shot learning is the field that tries to learn from just a handful of examples.

This survey sorts the many ways researchers have tried to do that. It groups them by the problem each one attacks. One group makes more training examples from the few that exist. Another reuses knowledge from a larger dataset. A third trains a model on many small tasks so that it learns to adapt quickly. The last adds other kinds of information, such as text next to images. The authors also explain how few-shot learning differs from ordinary machine learning, transfer learning and meta-learning, and they show how the ideas are applied to images.

The survey does not propose a new method. It is a map of the field, with comparison tables and a list of open problems, so it is most useful for deciding where to look next.

TL;DR

• A survey of recent few-shot learning work. It reports no experiments of its own §1.
• It groups methods by challenge: data augmentation, transfer learning, meta-learning, and multimodal learning §1 fig. 6.
• Its computer-vision tables list author-reported accuracy for classification, detection, segmentation and instance segmentation tab. 9 tab. 10 tab. 11 tab. 12.
• Its strongest-looking numbers are copied from other papers under mixed protocols, so use them only as pointers to the original work.

Problem

Edge devices each produce very little data, and traditional data-driven methods do not perform well with so few samples §1. Data distributions in real settings also have long tails, and a model trained in one domain often does not generalise to another §1. The survey's industry examples are a quality inspection line that is sensitive to lighting and has few defect samples, and electronic component identification with many brands and few samples of each tab. 1.

Without a method for these cases, the core difficulty is that a few samples cannot show the true data distribution §3 §2.6. Transfer learning helps only when source and target domains are related §4 §2.3. Meta-learning helps only when training and target tasks are similar §2.6 §5.4.

Method

This is a survey, so the "method" is the taxonomy and the summaries of each branch. The paper defines the setting first, then covers each branch.

Setting §2.1. A support set holds N classes with K labelled samples each (N-way K-shot). The query set holds the samples to predict. N-way 1-shot is one-shot learning; N-way 0-shot is zero-shot learning §2.1. A cross-domain variant transfers knowledge from a source domain to a target domain with a domain gap §2.1 §2.3.

Taxonomy §2.6 fig. 6. The paper splits FSL into single-modal and multimodal learning. Single-modal learning has three branches: data augmentation, transfer learning and meta-learning. Each corresponds to one of three problems: estimating the true data distribution from few samples (data augmentation), building reusable data-to-label mappings (transfer learning) and deriving task-to-model mappings (meta-learning) §2.6. Multimodal learning adds information from other modalities §2.6 §6.

Episode sampling. The survey describes the N-way K-shot episode in words. The pseudocode below follows that description; the text does not give an algorithm.

# N-way K-shot episode  [§2.1]
# Input:  labelled pool: mapping from class to its samples
# Output: support set with N classes and K samples per class

def sample_support(pool, n_way, k_shot):
    # 1. Pick N classes for this task  [§2.1]
    classes = random_sample(keys(pool), n_way)
    # 2. Draw K labelled samples per class; these form the support set  [§2.1]
    support = [draw(pool[c], k_shot) for c in classes]
    # 3. Return the support set; query samples are predicted  [§2.1]
    return support

Data augmentation §3. The paper divides augmentation into hand-crafted rules and learned data processing §3.

• Hand-crafted rules need expert design. At the data level, examples include random erasing, cutout-style filling, random cropping and mixing §3.1 fig. 7. Feature-level methods map images to a latent space and augment there, for example by combining global and local features §3.1 fig. 8. tab. 4 summarises the methods in this group.
• Learned data processing uses meta-learning to choose or generate augmentation policies §3.2 fig. 9. The paper names augmentation policy search (DADA), conditional generators (AFHN, MetaGAN) and a delta-encoder that learns differences within a class §3.2 tab. 5.
• The survey places the shift to learned processing after meta-learning matured around 2018 §3.3.

Transfer learning §4. The baseline pre-trains a backbone on a large dataset and then adapts it to the few-shot task §4 fig. 10.

• Pre-training and fine-tuning §4.1. Most parameters are frozen and only the classification layer is trained at test time §4.1. The paper states that fine-tuning improves 5-way 1-shot accuracy by 2% to 7% over the baseline in works it cites §4.1.

# Transfer-learning baseline: pre-train, freeze, fit a head  [§4.1]
# Input:  backbone; large labelled dataset; support set
# Output: classifier head for the target classes

def fine_tune_baseline(backbone, large_dataset, support):
    # 1. Pre-train the backbone on a large labelled dataset  [§4.1]
    backbone = pretrain(backbone, large_dataset)
    # 2. Freeze most lower-level parameters  [§4.1]
    backbone = freeze_all_but_last(backbone)
    # 3. Train only the classification head on the support set  [§4.1]
    head = fit(new_head(), backbone, support)
    return head

• Cross-domain FSL §4.2. The survey points to the BSCD-FSL benchmark with four datasets: CropDiseases, EuroSAT, ISIC and ChestX §2.5 §4.2. Methods it lists include feature-space ridge regression (FRN), affine feature transformation (FWT), self-supervised training on unlabelled target data (STARTUP), and domain-specific feature alignment §4.2.
• Comparison §4.3. The survey states that meta-learning beats transfer learning in standard FSL, and that this is reversed in cross-domain FSL §4.3.

Meta-learning §5. The paper treats meta-learning as learning from a dual sampling of data and tasks, so that the learner gets a good initialisation or a good learning rule for new tasks §5 fig. 4.

• Learning model parameters §5.1 tab. 6. MAML learns an initialisation with an inner loop per task and an outer update. Reptile is a first-order variant. The survey also lists FOMAML, Meta-SGD, TAML, iMAML and iTMAML, each with a stated change (for example, iMAML proposes a new loss and a gradient method that does not depend on the specific inner optimisation method) §5.1. Learned optimisers (LSTM-based, trained with reinforcement learning, Bayesian inference or evolution) and meta-learned neural architecture search (MetaNAS, few-shot NAS) are also covered §5.1.
• Learning a metric §5.2 tab. 7 fig. 11. The survey covers Siamese networks, triplet losses, prototype networks (with margin and attention variants), matching networks, relational networks and their variants §5.2.
• Transmitting information over graphs §5.3 fig. 12. Graph neural networks pass information between support and query nodes. The paper lists EGNN, Meta-GCN, GFL, DPGN, GERN, HGNN and Frog-GNN §5.3.

Multimodal learning §6. The branch uses text, attributes or other signals to complement the few images §6 fig. 13.

• Multimodal embedding §6.1 tab. 8. Semantic information (labels, attributes, descriptions) is aligned with visual features, for example through VAEs or embedding losses §6.1.
• Generating images from semantics §6.2. Text is used to generate extra training features or images with GANs, stacked conditional GANs, or attribute-conditioned encoders §6.2.

Computer vision §7. The survey counts papers by task and year, in tables and charts §7 fig. 14 fig. 15 fig. 16 fig. 17.

• Image classification §7.1 tab. 9. The table lists models on mini-ImageNet under 5-way 1-shot and 5-way 5-shot.
• Object detection §7.2 tab. 10. Results are given as 10-shot and 30-shot AP.
• Semantic segmentation §7.3 tab. 11. Results are given as 1-shot and 5-shot mean IoU.
• Instance segmentation §7.4 tab. 12. Results are given as 1-shot MAP50.

Future directions §8. The paper lists four: a better way to estimate the true data distribution from few samples §8.1; more robust data-to-label mappings that combine fine-tuning and meta-learning §8.2; meta-learning that works across task types and has a causal theory §8.3; and pre-trained models that fuse three or more modalities §8.4.

Hyperparameters

The survey runs no training of its own, so it states no training hyperparameters. The settings it does state are below.

Results

Author-reported. The survey does not run experiments. Each table below copies numbers from the cited paper named in the first column. The survey's own columns (extra training data, approach, date, code) are copied as printed. Tables are not on a common protocol, so do not rank rows across tables.

Image classification, mini-ImageNet §7.1 tab. 9. The survey lists these in table order, which is not sorted by accuracy. Highest 5-way 1-shot in the table: PT+MAP, 82.92%. Highest 5-way 5-shot: Transductive CNAPS + FETI, 91.5% (extra data).

Few-shot object detection §7.2 tab. 10. The table does not name the detection dataset.

Few-shot semantic segmentation §7.3 tab. 11.

Few-shot instance segmentation §7.4 tab. 12.

Survey-level summary. The text names no single best method. It says the top three image-classification methods are based on feature augmentation and feature transformation of the backbone §7.1. The table is not sorted, so this claim is not checked against it here.

Paper vs code

No official code. The LaTeX source and the abstract name no repository for this survey. Papers With Code appears only as the data source for fig. 5. No GitHub search was run, because a survey has no code of its own (repo convention for survey cards). The "Code" columns in tab. 9 to tab. 12 refer to the cited papers, not to this one.

Pitfalls

• No method to implement. Nothing here is an algorithm with equations. MAML is described only as "identifying the global optimisation direction by calculating the direction for each task" §5.1. Read the original papers before implementing any method from this list.
• Figure cross-references point at the wrong figure. Sections 7.2, 7.3 and 7.4 cite the label image for their charts §7.2 §7.3 §7.4. That label is the image-classification chart fig. 14. The detection and segmentation charts are fig. 15, fig. 16 and fig. 17.
• Figure and table numbers come from the LaTeX source. They are counted in source order, which should match the compiled paper, but no compiled PDF was checked. Confirm before citing a number.
• Scope is inconsistent. The abstract and introduction describe the last three years [abstract] §1. The tables include 2017 and 2019 entries tab. 9 tab. 12. The selection rule and the search sources are not stated.
• Contradictory definitions. Section 2.1 treats zero-shot as the N-way 0-shot case §2.1. Section 2.3 says one-shot learning and FSL "can essentially be considered as special ZSL" §2.3. The direction of the relation is therefore unclear.
• Query classes. Section 2.1 says query-set classes never appear in the support set §2.1. It does not say whether query samples come from the same classes, so the episode definition is ambiguous.
• Metric learning and episodes. Section 5.2 says metric learning "no longer divides the model into training and testing stages" §5.2, then places it inside the meta-learning frame in the same section.
• Unsupported claims. The survey says KGBert "surpasses humans in the field of FSL for the first time" §2.2; no source or number is given. The claim that meta-learning beats transfer learning in standard FSL §4.3 §5.1 is made without a table or citation in those sections.
• Copy errors in the tables.
  • In tab. 9, the Transductive CNAPS row lists "Self-Organizing Maps" as its approach. The same text appears on the WRN row.
  • In tab. 9, the row labelled Rodríguez et al. has the citation key patacchiola2020bayesian.
  • In tab. 12, the row labelled He et al. has the citation key zhang2020sg.
  • In tab. 3, SUN397 is credited to "Oh et al." with the citation key xiao2010sun.
  • The date column in tab. 9 has "Jau 2020" and "Jau 2019" where "Jun" is probably meant.
• Mask R-CNN as a few-shot baseline. tab. 12 lists Mask R-CNN (Aug 2017) as a few-shot entry with the approach "Data augmentation". The approach label is the survey's own choice.
• Dataset licences. tab. 3 lists Paris-Lille-3D under "CC BY-NC-ND 3.0". A non-commercial, no-derivatives licence cannot be used under this wiki's rules. The other licence entries are the survey's statements and were not checked against the datasets.
• Dataset numbers are unsourced. The class counts and split sizes in tab. 3 are given without a source in the text. Check them against the dataset papers before use.
• Table 10 has no dataset column. The detection results do not say which benchmark they use, so they cannot be compared with each other directly.
• Numbers from mixed protocols. Rows differ in extra training data, backbone, shot protocol and evaluation. The "Extra data" column flags some of these. Within one table, a higher number does not necessarily mean a better method.
• Hidden costs and speculation. The outlook sections state predictions without evidence §8.1 §8.3 §8.4, for example that pre-training will become the mainstream algorithm §8.3.

Lineage

• Builds on: none; this is a survey. It cites earlier surveys as related work: a 2018 survey on experience and conception learning (cite key shu2018small), a 2020 survey on generative and discriminative models (cite key lu2020learning) and a 2020 survey with a data, model and algorithm taxonomy (cite key wang2020generalizing) §1.
• Compared against: the authors place their challenge-based taxonomy against these three surveys §1.
• Superseded by: not known from wiki content.