Source-linked AI summary
Inductive Biases for Deep Learning of Higher-Level Cognition
Anirudh Goyal, Yoshua Bengio
TL;DR
Deep learning systems can perform well on familiar tasks, but their reasoning and generalization remain inconsistent under changing or out-of-distribution conditions. This paper examines higher-level and sequential inductive biases, including causal, semantic, attentional, and sequential structure, to clarify principles that may improve such capabilities.
Problem
Deep learning performance can be excellent yet reasoning, judgment, and generalization vary with tasks, architectures, training data, and distribution shift.
Method
The paper proposes and clarifies higher-level inductive biases involving causal and semantic variables, dynamic attention, grounded language, and serial-parallel computation.
Results
The paper concludes that additional structure and inductive biases may substantially improve system-2 tasks, natural language understanding, out-of-distribution systematic generalization, and efficient transfer learning.
Takeaways & Limitations
Studying human and animal inductive biases may clarify principles of intelligence and inspire neural architectures and training frameworks for more flexible generalization.
Takeaways & Limitations
The proposed attention framework still lacks discrete types, hard argument choices, and more powerful inference machinery beyond type matching.
Abstract
from arXiv · showhide
A fascinating hypothesis is that human and animal intelligence could be explained by a few principles (rather than an encyclopedic list of heuristics). If that hypothesis was correct, we could more easily both understand our own intelligence and build intelligent machines. Just like in physics, the principles themselves would not be sufficient to predict the behavior of complex systems like brains, and substantial computation might be needed to simulate human-like intelligence. This hypothesis would suggest that studying the kind of inductive biases that humans and animals exploit could help both clarify these principles and provide inspiration for AI research and neuroscience theories. Deep learning already exploits several key inductive biases, and this work considers a larger list, focusing on those which concern mostly higher-level and sequential conscious processing. The objective of clarifying these particular principles is that they could potentially help us build AI systems benefiting from humans' abilities in terms of flexible out-of-distribution and systematic generalization, which is currently an area where a large gap exists between state-of-the-art machine learning and human intelligence.
1. Has Deep Learning Converged?
Deep learning achieves strong performance in supervised and reward-rich settings, but its reasoning, judgment, and transfer remain inconsistent across tasks and distribution shifts. The paper proposes studying human and animal inductive biases, including causal and semantic representations, to address these gaps.
- High test accuracy does not establish correct reasoning or judgment, which vary with task, architecture, training data, and train–test alignment.
- The paper reviews high-level cognitive inductive priors involving sequential conscious processing and variables that are generally verbalizable.
- Current systems often excel on narrow tasks with abundant labels or rewards, whereas humans reuse knowledge to generalize from few examples.
- Structural-mechanistic models aim to predict across distributions generated by the same causal system rather than fitting one particular distribution.
- Causal understanding concerns how interventions change distributions and may require moving beyond independent samples toward non-stationary real-world processes.
- The paper argues that efficiently implementing and combining these inductive biases in one system remains an open question.
2. About Inductive Biases
Inductive biases are necessary for generalization, but standard deep learning often assumes fixed distributions and can overfit them. The paper emphasizes transfer, continual learning, systematic compositionality, and non-stationary environments as targets for additional biases.
- No learning algorithm generalizes best across all distributions, making inductive preferences necessary for selecting among equally good training solutions.
- Agency, Sequential Decision Making and Non-Stationary Data Streams: Agents encounter non-stationary observations because they, other agents, and exploration can change which parts of the environment are observed.
- Transfer Learning and Continual Learning: Transfer and continual learning require extracting information from past tasks to improve learning speed across potentially infinite task streams.
- Systematic generalization derives meanings for novel compositions from existing concepts and can apply to combinations absent from training data.
- Humans generalize from finite combinations to larger concept combinations, while proposed systematicity extends compositionality beyond distributed representations and network depth.
- Current deep learning methods tend to overfit the training distribution, motivating evaluations that test systematic out-of-distribution generalization.
3. Inductive biases based on higher-level cognition as a path towards systems that generalize better OOD
The paper draws on conscious cognition to motivate inductive biases for flexible reasoning and out-of-distribution generalization. It highlights a division between fast habitual processing and slower, sequential conscious processing, implemented through shared representations and global communication.
- Conscious vs Unconscious Processing in Brains: Humans commonly recruit conscious attention when novel situations require focused reasoning, new information, or inhibition of habitual reactions.
- System 1 and System 2: System 1 processing is rapid and parallel, whereas system 2 processing is controlled, sequential, and usually verbalizable.
- System 1 and System 2: Knowledge can exist as explicit, verbally communicable system 2 representations or implicit, intuitive system 1 representations that may emerge with practice.
- The Global Workspace Theory: Global Workspace Theory proposes a shared representation through which specialist components interact.
- The Global Workspace Theory: Broadcasting selected information enables modules to synchronize and exchange modality-independent semantic representations.
- The Global Workspace Theory: These shared representations are proposed to support flexible recombination of knowledge for systematic out-of-distribution generalization.
3.2 Attention as dynamic information flow.
Attention dynamically selects information for downstream computation, enabling variable binding and routing between modules. The paper contrasts soft selection with stochastic or deterministic hard attention.
- Soft attention selects elements by combining their value vectors using query-conditioned convex weights.The weights are produced by a softmax based on key–query matches.
- Stochastic hard attention samples one value, while deterministic hard attention selects one value when the softmax is saturated.
- Attention creates dynamic connections between computational modules rather than relying solely on fixed neural-network connections.
- Key-value attention implements variable binding by selecting which entity or object supplies each computational argument.Soft attention performs this binding differentiably while mixing multiple possibilities.
- Current attention mechanisms lack discrete types, hard argument choices, and inference that composes modules according to context.
3.3 Blend of Serial and Parallel Computations.
The proposed system-2 perspective combines parallel module activity with sequential, attention-controlled communication through a global workspace. This supports working-memory operations and is intended to improve generalization to unfamiliar settings.
- Blend of Serial and Parallel Computations: Modules can operate in parallel, but arbitrary communication passes through an attention-controlled global-workspace bottleneck.
- Blend of Serial and Parallel Computations: Because few elements enter coherence at each selection step, system-2 inference becomes highly sequential.
- Blend of Serial and Parallel Computations: Selected contents can enter short-term memory and influence subsequent thought, action, perception, and learning.
- Blend of Serial and Parallel Computations: The global workspace is proposed to hold only a partial state, including selected or imagined future elements rather than the entire environment.
- Blend of Serial and Parallel Computations: System-2 cognition is associated with handling unfamiliar settings and powerful out-of-distribution generalization.
- Blend of Serial and Parallel Computations: Exchangeable information broadcast through an inter-module communication route is hypothesized to support systematic generalization.Working memory also acts as a blackboard where active modules can improve coherence.
3.4 Semantic Representations Describing Verbalizable Concepts
The paper proposes that high-level conscious representations are generally verbalizable and often correspond to semantic variables that support language grounding and causal reasoning. These representations are linked to modular mechanisms intended to remain useful across distribution changes.
- Semantic Representations Describing Verbalizable Concepts: High-level variables manipulated consciously are proposed to be generally verbalizable through words or phrases.The mapping from semantic representations to language is described as simple but lossy.
- Semantic Representations Describing Verbalizable Concepts: System-1 knowledge includes brain content that is implicit and difficult to translate into natural language, unlike system-2 knowledge.
- Semantic Representations Describing Verbalizable Concepts: Grounded language learning couples language with an agent’s observations and actions to shape top-level representations toward language-expressed concepts.The paper presents this as a form of weak supervision rather than forcing human-specified labels.
- Semantic Representations Describing Verbalizable Concepts: Semantic variables are hypothesized to often be causal variables, including agents, objects, actions, and their properties or modalities.
- Semantic Representations Describing Verbalizable Concepts: Structural causal models organize causal influences into independent mechanisms, with each mechanism relating direct causes to their direct effects.
- Semantic Representations Describing Verbalizable Concepts: The causal structure, mechanisms, and definitions of high-level causal variables are described as stable across distribution changes.
3.6 Local Changes in Distribution in Semantic Space
The paper proposes that distribution changes are usually localized in a semantic space: only a few variables or mechanisms need modification to explain them. For interventions, this locality follows from the limited number of variables directly affected by actions.
- Local Changes in Distribution in Semantic Space: Observation distributions can change because of environmental dynamics and because of the learner’s actions or policy.
- Local Changes in Distribution in Semantic Space: Humans often explain distribution changes briefly, motivating the assumption that their sources are localized in semantic space.
- Local Changes in Distribution in Semantic Space: Only one or a few variables or mechanisms are assumed to need modification to account for most distribution changes.
- Local Changes in Distribution in Semantic Space: New words may be introduced when existing vocabulary cannot concisely explain a change, corresponding to newly named latent variables.
- Local Changes in Distribution in Semantic Space: For system-2 interventions, changes begin at one or a few causal-graph nodes, while downstream effects follow from those initial interventions.
3.7 Stable Properties of the World
The section proposes learning stable aspects of the world slowly while adapting quickly to non-stationary or novel aspects. This can make represented knowledge more stable and reduce adaptation when distributions change.
- Stable properties should be learned more slowly than non-stationary or novel properties, which should be learned faster.
- The learning process should pressure the agent to discover stable aspects among quickly changing observations.
- More stable knowledge would require less adaptation when the data distribution changes.
- New high-level concepts composable from existing concepts can be learned as quickly as from a single example or definition.
- Causal interventions may temporarily break causal links while most of the causal graph remains stable.
3.8 Sparse Factor Graph in the Space of Semantic Variables
The section proposes representing semantic high-level variables with a sparse, structured factor graph rather than independent marginals. This sparsity is linked to natural-language knowledge and the limited write access of conscious processing.
- The joint distribution of high-level concepts should be represented by a sparse factor graph.
- A factor graph uses variable nodes and factor nodes, with factors representing dependencies among connected variables.
- Natural-language knowledge is argued to be sparse because sentences typically relate only a few concepts.
- Sparse factor graphs allow only a few variables to be synchronized during an inference step, matching the conscious-processing bottleneck.
- Sparse dependencies are preferred to complete independence because semantic concepts can have strong but sparse relationships.
- Units in slightly lower layers of generative adversarial networks more often have semantic interpretations than top-level units.
3.9 Variables, Instances and Reusable Knowledge Pieces
The section proposes reusable schemas or factor templates that encode generic relations and can be instantiated across different arguments. Inference can then generate only the graph pieces needed for reasoning.
- Parameter sharing allows general patterns to be instantiated into multiple factors instead of using separate parameters for every factor.
- Generic factors or schemas are reusable probabilistic relations whose argument variables can be bound to specific instances.
- Natural language and knowledge graphs motivate representing knowledge as relations that can be instantiated on different instances.
- Inference need not store the full graph of all instances; it can generate needed graph pieces and reason at an abstract level.
- The framework distinguishes verbalizable declarative knowledge from the inference machinery used to reason with it.
3.10 Relevant causal chains (for learning or inference) can be approximated as very short chains
The section proposes approximating relevant causal dependencies with short chains of events linked through a sparse semantic factor graph. This bias is motivated by conscious-processing limits and can mitigate long-term dependency learning problems.
- Event-based segmentation represents sensory streams as variable-length episodes and events rather than equally spaced time steps.
- Event segmentation supports temporal reasoning by enabling selective retrieval of context-relevant past information.
- Relevant causal chains are proposed to be sparse and decomposed into short event chains linked by a top-level semantic factor graph.
- These chains support counterfactual learning, credit assignment, explanations, and plans for achieving goals.
- Human short-term memory and conscious-processing limits motivate assuming that relevant temporal dependencies involve short chains or shallow dependency graphs.
- The short-dependency prior is described as the strongest ingredient for mitigating vanishing gradients in long-term dependency learning.
3.11 Context-dependent processing involving goals, top-down influence, and bottom-up competition
Human perception combines top-down context and priors with bottom-up sensory signals, with their relative relevance changing by context. The section favors dynamically integrating these signals throughout the processing hierarchy.
- Top-down signals encode context, priors, and preconceptions, while bottom-up signals represent what is directly observed through sensation.
- The appropriate top-down signals can change with context, motivating dynamic weighting of top-down and bottom-up information.Attention is offered as one possible mechanism for dynamically weighting the signals.
- Dynamic combination of these signals could improve robustness to distractions and noisy data.
- Dynamic Integration of Bottom-up and Top-Down Information is presented as an architectural inductive bias.
- The proposed architecture combines top-down contextual information with bottom-up sensory signals at every computational level of the hierarchy.
4. Declarative Knowledge of Causal Structure
The section presents causal structure, independent mechanisms, interventions, and agent actions as forms of declarative knowledge that can support out-of-distribution generalization and rapid adaptation. It connects causal models with sparse distribution changes, transfer learning, and affordances.
- A causal model represents a family of joint distributions corresponding to different interventions, unlike a statistical model's single joint distribution.Interventions modify the default distribution, for example by removing a node's parents and setting its value.
- A Structural Causal Model assigns each variable a deterministic function of its direct causes and an independently modeled noise term.The causal graph encodes parent relationships through the DAG configuration C.
- The Independent Causal Mechanisms principle treats a complex generative model as a composition of mechanisms that do not inform or influence one another.The principle includes separate intervenability, modularity, autonomy, and invariance.
- Observational data from one training distribution cannot distinguish A → B from B → A when both graphs are Markov equivalent.The example therefore uses transfer-distribution samples to disambiguate the causal hypotheses.
- Under covariate shift caused by intervention on A, p(B | A) remains unchanged while the marginal distribution of A changes.
- The causal and anti-causal models are compared by how quickly they adapt to transfer-distribution samples through fine-tuning.Adaptation speed is measured by log-likelihood after multiple stochastic gradient-ascent steps.
- Changes between related domains can be attributed to changes in as few causal mechanisms as possible, enabling agents to recognize sparse changes and adapt existing knowledge.
- Affordance representations encode how agents can cause environmental changes by controlling objects or influencing other agents.A useful causal variable is suggested to be one that can be selectively controlled through actions or skills.
5. Conclusions
The conclusion argues for moving from deep statistical models toward deep structural models with additional inductive biases for dynamic, changing conditions. These biases are intended to organize knowledge into stable, reusable parts for new settings and tasks.
- Deep structural models are proposed to extend deep statistical models from system 1 tasks toward system 2 tasks in changing conditions.
- Additional structure and inductive biases may improve natural language understanding, out-of-distribution systematic generalization, and efficient transfer learning.
- The proposed biases are motivated by expected and observed gains in out-of-distribution generalization and fast adaptation rather than same-distribution test performance.
- The general insight is to organize knowledge into stable reusable parts likely to remain useful across new settings and tasks.