The Consciousness Prior

By Yoshua Bengio. 2017

267 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Representation & Self-Supervised Learning, Cognitive Science & Psychology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Representation & Self-Supervised Learning

In Cognitive Science & Psychology

Citations in the corpus

Underspecification Presents Challenges for Credibility in Modern Machine Learning (2020, background): “Although some general-purpose heuristics have been proposed (e.g., Bengio, 2017), proposals for expressing application-specific requirements with flexible but unified frameworks may be a promising middle ground solution.”
Disentanglement via Mechanism Sparsity Regularization: A New Principle for Nonlinear ICA (2021, background): “While there is a significant amount of interest in learning probabilistic or causal graphs between high-level latent variables extracted from low-level observations (Bengio, 2019; Schölkopf, 2019; Schölkopf et al., 2021; Goyal and Bengio, 2021; Ke et al., 2021), there have been comparatively few pra”
Sever: A Robust Meta-Algorithm for Stochastic Optimization (2018, background): “This could be done, for instance, by training a neural network with some cost function that encourages independent features (some ideas along these general lines are discussed in [Ben17]).”
On the Binding Problem in Artificial Neural Networks (2020, background): “Alternatively, Bengio (2019) argues for the importance of a low-dimensional ‘conscious state’ (working memory) composed of largely independent units of abstraction that can be selected via attention (perhaps reminiscient of Schmidhuber, 1992b).”
Weakly-Supervised Disentanglement Without Compromises (2020, background): “In this paper, we consider learning disentangled representations from pairs of observations which differ by a few factors of variation (Bengio, 2017; Schmidt et al., 2007; Bengio et al., 2019) as in Figure 1.”
Recurrent Independent Mechanisms (2019, background): “Our positive experimental results lend support to the consciousness prior (Bengio, 2017), i.”
A Meta-Transfer Objective for Learning to Disentangle Causal Mechanisms (2019, background): “…kind of assumption has been proposed to help discover such linguistically relevant high-level representations from raw observations, such as the consciousness prior (Bengio, 2017), with the idea that humans focus at any particular time on just a few concepts that are present to our consciousness.”
Task-Aware Information Routing from Common Representation Space in Lifelong Learning (2023, background): “Such a system enables efficient CL in humans with systematic generalization across tasks (Bengio, 2017).”
Measuring the tendency of CNNs to Learn Surface Statistical Regularities (2017, background): “More generally, new proposals such as [2] aim to transition away from making predictions in the perceptual space and instead operate in the higher order abstract space.”
Troubling Trends in Machine Learning Scholarship (2018, neutral): ““thought vectors” [36] and the “consciousness prior” [4].”
Improving Generalization for Abstract Reasoning Tasks Using Disentangled Feature Representations (2018, background): “…of a generally useful latent space (disentanglement, linearity, sparsity, ...) and explore new objective functions that can be leveraged for unsupervised representation learning, such as various representation losses [8, 9], GAN-inspired discriminator networks [10] and predictive capacity [11, 12].”
Learning Awareness Models (2018, background): “The consciousness prior (Bengio, 2017) considers recurrent latent dynamics models similar to ours and suggests mapping from their hidden states to other spaces that aren’t directly modelled.”
Learning by Abstraction: The Neural State Machine (2019, methodology): “Drawing inspiration from Bengio’s consciousness prior [12], we further deﬁne a set of semantic embedded concepts that describe different entities and aspects of the domain, such as various kinds of objects, attributes and relations.”
On the link between conscious function and general intelligence in humans and machines (2022, methodology): “Given the limited capacity and focus of the global workspace, as well as the role of verbalization in reporting awareness, Bengio (2017) interprets the workspace as an area which contains and manipulates sparse factor graphs that represent the relationships between a small set of high-level…”
An Introduction to Deep Reinforcement Learning (2018, unknown): “Indeed, a conscious thought at a particular time instant can be seen as a low-dimensional combination of a few concepts in order to take decisions (Bengio, 2017).”
D’ya Like DAGs? A Survey on Structure Learning and Causal Discovery (2021, background): “It is also a burgeoning topic in machine learning and artificial intelligence [6, 16, 65, 75, 144, 214, 255, 265], where it has been argued that a consideration for causality is crucial for reasoning about the world.”
State-Denoised Recurrent Neural Networks (2018, background): “Related to language, the consciousness prior of Bengio [2017] suggests a potential role of operating in a reduced or simpliﬁed representational space.”
Evaluating Machines by their Real-World Language Use (2020, background): “Other ideas include different modeling choices (Bengio, 2017) and learning paradigms (Mao et al.”