Source-linked AI summary
The Consciousness Prior
Yoshua Bengio
TL;DR
The paper addresses how to learn high-level abstract representations of concepts manipulated through language. It proposes a consciousness-inspired prior based on attention-selected, broadcast conscious states and sparse dependencies among high-level variables. The paper argues that this framework connects representation learning with natural language and symbolic knowledge representation, while leaving some aspects of consciousness unstudied.
Problem
Learning high-level abstract representations and disentangling abstract factors remains a central representation-learning challenge addressed by the proposed prior.
Method
The paper formalizes consciousness as attention selecting a few elements from a broader representation and broadcasting them as a conscious state, with a sparse factor-graph prior over high-level variables.
Results
The paper presents the consciousness prior as a framework for discovering human-relevant high-level variables and connecting conscious states with natural language and symbolic knowledge representations.
Takeaways & Limitations
The framework suggests combining attention-based conscious states with uncertainty-aware symbolic facts, rules, and reasoning mechanisms.
Takeaways & Limitations
The paper does not study aspects of consciousness such as the notion of self or subjective perception.
Abstract
from arXiv · showhide
A new prior is proposed for learning representations of high-level concepts of the kind we manipulate with language. This prior can be combined with other priors in order to help disentangling abstract factors from each other. It is inspired by cognitive neuroscience theories of consciousness, seen as a bottleneck through which just a few elements, after having been selected by attention from a broader pool, are then broadcast and condition further processing, both in perception and decision-making. The set of recently selected elements one becomes aware of is seen as forming a low-dimensional conscious state. This conscious state is combining the few concepts constituting a conscious thought, i.e., what one is immediately conscious of at a particular moment. We claim that this architectural and information-processing constraint corresponds to assumptions about the joint distribution between high-level concepts. To the extent that these assumptions are generally true (and the form of natural language seems consistent with them), they can form a useful prior for representation learning. A low-dimensional thought or conscious state is analogous to a sentence: it involves only a few variables and yet can make a statement with very high probability of being true. This is consistent with a joint distribution (over high-level concepts) which has the form of a sparse factor graph, i.e., where the dependencies captured by each factor of the factor graph involve only very few variables while creating a strong dip in the overall energy function. The consciousness prior also makes it natural to map conscious states to natural language utterances or to express classical AI knowledge in a form similar to facts and rules, albeit capturing uncertainty as well as efficient search mechanisms implemented by attention mechanisms.
1 Introduction
The paper proposes a consciousness-inspired prior for learning top-level abstract representations of concepts humans manipulate through natural language.
- The proposed prior targets top-level abstract representations of concepts used in natural-language communication.It is presented as a new kind of prior for representation learning.
- The prior is inspired by modern theories of consciousness, including Global Workspace Theory.The paper frames consciousness as awareness of information in one’s own mind or of external objects.
- Its central contribution is a machine-learning justification for selecting elements of conscious thought through attention and broadcasting them to downstream brain processes.The proposed mechanism is described as influencing perception, action, and subsequent conscious thought.
2 System 2 Processing and Global Workspace Theory of Consciousness
The paper operationalizes conscious cognition as language-reportable, System 2-like processing and relates it to Global Workspace Theory. It focuses on a bottleneck in which selected information becomes globally available, while excluding other aspects of consciousness from study.
- The paper defines conscious aspects of cognition as those humans can report through language because consciousness lacks a generally accepted definition.It associates conscious processing with Kahneman’s System 2 cognitive abilities.
- Global Workspace Theory describes conscious thought as selected information that temporarily forms working memory and becomes globally available to other brain processes.The theory is the dominant family of theories anchoring the paper’s account of consciousness.
- The paper studies machine-learning formalizations of Global Workspace Theory rather than the notions of self or subjective perception.Experiments are intended to evaluate effects such as sample efficiency and out-of-distribution generalization.
3 Consciousness Prior Theory
The consciousness prior constrains high-level representation learning through a low-dimensional attention-selected state and a sparse factor-graph assumption over abstract concepts. It also connects these representations to language, sequential reasoning, and symbolic knowledge while preserving uncertainty and context.
- 3.1 Extracting a Conscious State: The high-level state is intended to disentangle abstract explanatory factors so simple transformations can select information about individual factors.Elements of the representation are treated as variables whose dependencies must be modeled jointly, including dependencies across time.
- 3.1 Extracting a Conscious State: The framework derives a low-dimensional conscious state from a high-level representation using attention, prior conscious state, memory, and stochastic selection.The state represents a small subset of information made consciously accessible from a broader unconscious representation.
- 3.2 Sparse Factor Graphs: The consciousness prior assumes a sparse factor graph in which each factor depends on only a few variables, motivated by the structure of natural-language statements.Strong factors correspond to concise statements that involve few concepts and are highly likely to be true; attention can selectively explore relevant graph regions.
- 3.3 Learning Objectives: The proposed learning objective trains representation and attention mechanisms so attended elements are useful for predictions or actions, alongside reconstruction, supervised, reinforcement-learning, or unsupervised objectives.The framework separates selecting and combining a few elements into conscious substates from using sequences of those substates for predictions or actions.
- 3.5 Connection to Language and Symbolic Knowledge Representation: Language can regularize encoders because conscious states should map bidirectionally to utterances, encouraging representation elements to correspond to concepts expressible in words or phrases.The paper presents this as a possible weak form of supervision that may sharpen disentangled representations and support collaborative task-solving.
- 3.5 Connection to Language and Symbolic Knowledge Representation: Conscious states offer a possible bridge between deep learning, symbolic knowledge representation, and higher-level cognition by representing facts or rules while retaining uncertainty and goal-dependent context.Sequentially combining such states is presented as a basis for more complex predictions, inferences, or actions.
4 Considerations for Experimenting with the Consciousness Prior
The paper recommends starting with simple, fast-to-evaluate experiments to test distinct approaches to the consciousness prior. Experiments should isolate the training objective, use meaningful abstractions with high predictive power, and initially avoid linguistic input.
- Simple toy experiments enable rapid comparison of qualitatively different approaches and make learned representations easier to analyze.The proposed theory is novel and may develop in multiple ways, so short turnaround and interpretable concepts are emphasized.
- Experiments without linguistic input can test whether the training objective and framework alone discover appropriate high-level concepts.The text notes that babies can learn intuitive physics without linguistic guidance.
- Meaningful abstractions with high predictive power are preferable to pixel-level prediction when detailed future states are difficult to predict.Predicting whether a pile of blocks falls on or off a table illustrates a discrete outcome that is easier to forecast than exact landing positions.