On the Binding Problem in Artificial Neural Networks

In plain words

Neural networks are very good at finding patterns in data, but they often fail to apply what they learned to new situations that are a little different. A person understands a scene as a set of things (a cup, a hand, a table) and can rearrange those things in endless ways. The authors argue that today's networks have no reliable way to keep the pieces of a scene separate, label them as things, and then reuse them in new combinations. They call this missing ability the binding problem.

The paper does not present a new model. It splits the problem into three jobs: first, forming separate "object" pieces out of raw input; second, storing those pieces so they do not blend into each other; third, combining them to reason and predict. For each job it reviews ideas from neuroscience, psychology and machine learning, and it argues that the right answer is to build these abilities into neural networks directly, learned end to end, rather than bolting on a separate symbolic system.

Its value for an engineer is as a map of the design space and a list of open questions, such as how to learn objects without labels and how to measure whether a system generalizes systematically.

TL;DR

A position-and-survey paper, not a method paper. The central claim is that the inability of neural networks to dynamically bind information that is spread over the network is the root cause of their weak systematic generalization §1 §2.3. The binding problem is divided into representation (keeping object representations separate and in a common format) §3, segregation (forming objects from raw input, mostly unsupervised) §4, and composition (relating objects to build structured models) §5. The paper gives no equations, algorithms, hyperparameters or experiments; implementable content is limited to the design requirements and the cited mechanisms.

Problem

Existing networks need large amounts of data, transfer poorly to new tasks, and are fragile under distribution shift §1. Figure 1 collects examples: CNNs biased toward texture over shape, agents that fail when a game's paddle is shifted, and networks that struggle with same-or-different visual relations fig. 1 §2.1. The authors attribute these failures to the binding problem: the network cannot dynamically route information distributed across units into symbol-like entities §2.3. Architecture and weights are fixed at training time, so routing cannot change with context §2.3.

The problem is stated in its general form. Perceptual binding (for example colour and shape of one object) and cognitive binding (for example variable binding in reasoning) are treated as instances of one problem §2.3.

Method

The paper's framework has three parts. The text gives no concrete functions for them, so the schematic below only encodes the division the authors make §2.3 fig. 2:

# Schematic of the three aspects of the binding problem  [§2.3] [fig. 2]
# Input:  raw sensory input x
# Output: prediction, inference or behavior y

def binding_pipeline(x):
    # 1. Segregation: form object representations from raw input  [§4]
    objects = segregate(x)
    # 2. Representation: keep object features separate and in a common format  [§3]
    objects = separate_and_format(objects)
    # 3. Composition: relate objects and build a structured model over them  [§5]
    y = compose(objects)
    return y

Representation §3. Object representations must be self-contained, so that they can be related without losing integrity, while keeping the distributed features of neural codes §2.3. Four desiderata are named §3.1:

• Separation §3.1.1: features of different objects must not interfere, even when composed, and separation is an independent degree of freedom rather than a function of content.
• Common format §3.1.2: objects must share an interface so that relations and learned skills transfer across objects.
• Disentanglement §3.1.3: factors of variation should map to features in a local, independent way.
• Temporal dynamics and reliability §3.2.1 §3.2.2: object representations need recurrent updates with a stable identity, and they need to represent uncertainty, including objects that are temporarily unobserved.

Mechanisms reviewed §3.3: slot-based representations (instance, sequential, spatial and category slots) §3.3.1; augmentation of distributed codes with separation §3.3.2; tensor product representations §3.3.3; and attractor dynamics for multi-stable inputs §3.3.4. The authors judge that slots are the simplest approach, while augmentation and tensor product representations are harder to build but support more sophisticated use §3.4.

Segregation §4. Objects are defined by the task and context, so the authors argue that segregation must be learned largely without supervision and later refined with task information §4.4. The notion of an object is broad and spans abstraction levels, from physical objects to categories and behaviours §2.3 §7. Reviewed properties: modularity §4.1.1, hierarchy §4.1.2 and multiple domains §4.1.3. Dynamics: multistability §4.2.1, top-down feedback §4.2.2 and consistency over time §4.2.3. Methods: clustering approaches to segmentation §4.3.1, neural segmentation §4.3.2, sequential attention §4.3.3 and probabilistic generative models §4.3.4. The authors stress that segregation should be fully differentiable so that it can interact with the other two problems §4.4.

Composition §5. Composition needs learned abstract relations that can be applied recursively to object representations, and a mechanism that restructures information flow when the needed structure is not known in advance §2.3. Relations and variable binding are discussed in §5.1.1 and §5.1.2, and relational frames, in which entailment rules such as transitivity or symmetry apply to relations between objects, in §5.1.3. Relational responding and structure inference are in §5.2.1 and §5.2.2. Methods: graph neural networks §5.3.1, which the authors describe as the promising approach for relational responding §5.4, and neural computers §5.3.2. Attention is named as a mechanism for dynamic variable binding, and self-attention is discussed among the methods §5.3.1 §5.4.

Design principles that the discussion makes explicit §7:

• Objects are the key to compositionality, and compositionality is key to systematic generalization.
• Object learning should be unsupervised, integrated into the network, and learned end to end rather than handed to a separate symbolic system.
• Specialized inductive biases may help, but learning should take priority where possible.
• Scaling alone is judged unlikely to produce object representations, and the authors are pessimistic about scaling on less structured inputs such as raw perception §7.

Hyperparameters

The paper has no trained model and no tuned parameters. The only numeric parameter is a cited capacity estimate.

Results

No experiments, no benchmarks and no results are reported. The paper is a survey and position paper. Its evaluation section §3.4 §4.4 §5.4 lists measurement approaches proposed in other work, without numbers from this paper:

The conclusion calls for benchmarks that bridge toy datasets and real sensory data, and for unsupervised metrics analogous to FID, without defining such a metric §8.

Paper vs code

No official code.

No official code was found after these steps: the paper and its abstract contain no URL, and the source has none; a GitHub repository search for "binding problem neural" and "greff binding problem" returned no repository owned by an author or their lab.

Pitfalls

• Nothing to implement directly. The framework is conceptual. Requirements such as "separation is an independent degree of freedom" §3.1.1 do not say how to build it. Start from the primary methods the paper cites for each mechanism.
• Object is used in a broad sense. The paper deliberately includes abstract entities such as categories and behaviours §2.3 §7. It states that segregation resembles instance segmentation but concerns objects in their most general form §2.3, so do not map "object" directly onto segmentation masks.
• Strong positions are not tested here. The claims that the binding problem is the root cause §1 §2.3, that objects should be learned unsupervised §7, and that scaling will not suffice §7 are argued, not measured in this paper.
• Scope of the comparisons. Comparisons with hybrid neuro-symbolic work and with causal, graph-network and large-language-model approaches §7 reflect the authors' view and are not benchmarked.
• Capacity figure is from cited psychology. The 3 to 9 object limit §3.1.1 is presented as "generally believed" and is not a result of this paper.
• Figures are described, not reproduced. The key diagrams are fig. 1 (failure examples), fig. 2 (three aspects) and fig. 3 (superposition catastrophe). Later figures are not numbered here; cite them by section.
• Dated snapshot. arXiv v1 from December 2020 is the only version in the record. Later work on object-centric and compositional models is not covered here.
• Section numbering. Numbers follow the source order (§2 binding problem, §3 representation, §4 segregation, §5 composition, §6 related disciplines, §7 discussion, §8 conclusion). Related-discipline material (Gestalt, feature integration, relational frame theory, linguistic compositionality) is in §6 and is background only.

Lineage

• Builds on: the connectionist critique of symbolic reasoning in neural networks (Fodor 1988; Marcus 2003), cited as earlier perspectives §1 §8.
• Compared against: Lake et al. (2017), whom the authors see as using neural networks as a means for structured models rather than implementing them directly §7.
• Related frameworks discussed: causal structural models §7, graph neural networks §5.3.1 §7, and working-memory accounts of consciousness (Bengio 2019) §7.
• Superseded by: not known from content in this wiki.