Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

• Conscious states are both rich and ineffable.
• Explanatory gap: consciousness cannot be reduced to physical processes.
• Information theoretic dynamical systems perspective on richness and ineffability of consciousness.
• Richness of conscious experience corresponds to amount of information in conscious state.
• Ineffability corresponds to amount of information lost at different stages of processing.
• Attractor dynamics in working memory induce impoverished recollections of original experiences.
• Discrete symbolic nature of language is insufficient for describing rich and high-dimensional structure of experiences.
• Similarity in cognitive function of two individuals relates to improved communicability of their experiences to each other.

Paper Content

Introduction

• Conscious states present apparent contradictions: they are both rich and ineffable.
• This paper provides an information theoretic dynamical systems perspective on why this is the case.
• Conscious experiences are derived from neural processes according to known physical laws.
• Illusionists argue that consciousness is an illusion, while realists deny this.
• The overflow debate is between those who hold that consciousness is rich and ineffable, and those who deny it.
• This paper proposes a formal description of how neural dynamics could give rise to the ordinary sense of richness and ineffability in the brain.
• Attractor dynamics in working memory would induce impoverished recollections of our original experiences.
• The discrete symbolic nature of language is insufficient for describing the rich and high-dimensional structure of experiences.
• Similarity in the cognitive function of two individuals relates to improved communicability of their experiences to each other.
• Shannon information theory and Kolmogorov complexity play a central role in making arguments precise.
• Various sources of ineffability in conscious experience are discussed.
• Implications of the model on the debate surrounding ‘phenomenal’ vs. ‘access’ consciousness are discussed.

Preliminaries: computation through neural dynamics

• Neural activation space: a space in which temporally evolving states of neural activity follow trajectories
• Attractors: emergent properties of dynamical systems that allow neural networks to solve computational problems that require persistent memory

Neural activation state space

• Brain activity can be quantified in several ways
• Brain activity can be represented as a vector in an N-dimensional state space
• Mathematical framework of dynamical systems theory can be used to reason about mental states
• Projecting the state onto lower-dimensional subspaces can reveal what a pattern of neural activity represents
• Probabilistic transition model can measure likelihood and information content of a state
• Similarities between states can be quantified according to some distance metric

Neural dynamics

• Neural activity unfolds in time
• Recurrent loops form between neurons
• Activity of one neuron influences future activity of surrounding neurons
• Vector field illustrates how neural state transitions in the absence of noise
• Inputs warp vector field, resulting in different trajectories

State attractors

• Neural dynamics often require persistent memory
• Dynamical systems can form regions in state space called “basins of attraction”
• Trajectories that enter these regions progress towards steady states
• Attractors can have complex and high-dimensional structure
• Attractors are mutually exclusive and can be identified with symbols
• Attractors can be used for efficient long-term memory
• Attractors can be modulated by external inputs
• Attractors have been used to explain neural computations
• Attractors have been observed empirically
• Attractors are mutually exclusive and create a partition of the state space
• Attractors have a dual discrete and continuous nature
• Introducing randomness in the dynamics can sample one of the attractors
• An RNN was trained to solve a 3-bit flip-flop task
• The RNN solved the task using fixed point attractors
• Each attractor corresponds to one of the 8 possible bit configurations
• Attractor dynamics can naturally emerge in neural networks to implement computations

A dynamical systems model of conscious experience

• Dynamical systems model of consciousness with state attractors can account for communicable, rich, and ineffable aspects of experience
• Information theoretic perspective used to characterize richness as information, ineffability as information loss, and communicability as information retention
• Model links ineffability to Global Workspace Theory
• Multiple forms of information loss discussed in intra-personal and inter-personal communication pathways

Motivating attractor dynamics as a model for conscious experience

• Working memory is typically considered to be the attended contents of short term memory
• Many theories of consciousness claim that what we are consciously aware of is the contents of working memory
• Global Workspace Theory states that information becomes conscious by gaining entry into a limited workspace
• Workspace serves as a hub for coordinating brain-wide activity
• Working memory and attractor dynamics are linked
• Attractor model for working memory postulates that working memory emerges from recurrently connected cortical neural networks
• Attractor dynamics can support suppression and integration of inputs
• Neural winner-take-all models implement hybrid analog-discrete computation
• Robustness, discreteness, and temporal integration of information are all traits apparent in working memory
• Attractor dynamics predict that our experience consists of a sequence of relatively stable states
• Conscious neural representations are more stable than unconscious ones
• Robustness to noise and reproducibility of states are core properties of attractor dynamics

Richness and ineffability

• Notation is used to describe random variables, probability distributions, and expectations
• Richness of experience is having a lot of detail, texture, nuance, or informational content
• Ineffability is too great for words and is measured by information loss when trying to express a conscious state in words
• Effability is the negation of ineffability and is measured by the amount of shared information between two variables
• Shannon entropy measures the average number of bits needed to specify a value of a variable
• Kolmogorov complexity measures the information carried by an individual state independently of probability distributions
• Richness corresponds to Kolmogorov complexity and ineffability corresponds to conditional Kolmogorov complexity
• Shannon entropy and Kolmogorov complexity are closely related metrics of richness
• Shannon entropy has drawbacks for characterizing ineffability
• Kolmogorov complexity should be used to characterize richness and ineffability of experiences

Intra-personal ineffability

• The functions f A φ, f S φ and f M φ are deterministic while f V φ and f X φ are generally stochastic.
• Entropy H φ is parameterized since it depends on p φ.
• The relation of time-varying states x to a smaller subset a of attractor states is a defining characteristic of attractor dynamics.
• The presence of attractors decreases the richness of transient states overall.
• Attractor states are strictly less rich than trajectory states.
• Selective release of attractor working memory states to downstream processing implements an information bottleneck that limits the richness of downstream inputs.
• Conscious experience is upper bounded by the richness of trajectories.
• Information loss from trajectory X to working memory output A is expected to be significantly large.
• Ineffability of experience is most obvious when attempting to put it into words.
• Information loss from attractor A or conscious experience S to verbal message M means the latter do not fully identify the former.
• Positive information loss H φ (S|M ) implies it is generally impossible to perfectly recover the conscious experience from the verbal message.
• Attractors can have more complex and high-dimensional structure than simple fixed points.
• The brain is hierarchical in nature with many levels of spatial and temporal organization.
• Attractor dynamics appear to be ubiquitous across organizational levels and cortical regions of the brain.
• The presence of multiple weakly coupled neocortical attractor networks yields benefits including specialization and increased memory capacity.

Inter-personal ineffability

• Communication between two individuals is modeled using a speaker (Alice) and a listener (Bob).
• Alice’s experience is communicated to Bob using a message m.
• Bob’s experience is a function of Alice’s cognitive parameters φ.
• Ineffability of Alice’s experience given the message is K(s|m).
• Ineffability increases when Bob’s cognitive parameters φ are not known.
• Ineffability of the attractor state is E p φ (a|m) K(a|m).
• Ineffability of the conscious experience is at least as great as the complexity of the working memory attractor.
• Ineffability of the conscious experience can only improve by conditioning on Bob’s parameters.
• The mismatch between Alice and Bob’s parameters corresponds to the difference between their cognitive functions.
• Inter-personal ineffability is greater when Alice and Bob are dissimilar.

Phenomenal and access consciousness

• Information theoretic dynamical systems perspective on richness and ineffability
• Ambiguities in the meaning of access
• Three notions of effability, accessibility, and reportability
• Global Workspace Theory of consciousness
• Three classes of representations: unconscious, conscious, and accessible but not amplified
• Richness of transient states may be consciously experienced but not directly accessible or reportable
• Debate between adherents of Global Workspace Theory on whether items from third class are conscious
• Attendedness may be used to delineate conscious and unconscious
• Attractor model of working memory suggests second explanation for self-report of phenomenal experience
• Rich phenomenal experience exists and is reportable

Conclusion

• Characterizes conscious experience from an information theoretic perspective
• Connects ineffability with information loss
• Attractor model for working memory
• Compressed nature of language increases information loss
• Attention mechanisms in machine learning
• Model of emitter-receptor communication
• Information bottlenecks in machine learning
• Benefits of information loss
• Visualization of neural state space
• Neural dynamics and trajectories in activation space
• Attractor dynamics in neural networks
• Shannon entropy and Kolmogorov complexity
• Model of intra-personal ineffability
• Cognitive dissimilarity and ineffability