Computational Consciousness
Science

Arrival and Coffee

Introduction

David Chalmers (Philosophy Keynote)

New York University

Is the J-Space a Global Workspace?

In a recent report from Anthropic, Gurnee et al define a special component of a language model's activation space: the Jacobian space, or J-space for short. Representations in the J-space are verbalizable in that they are especially likely to influence the words used in the language model's outputs. Gurnee et al argue that the J-space functions as a mechanism for conscious access, or access consciousness. They also argue that it closely resembles a well-known proposed mechanism for access consciousness: the global workspace. I will argue that while the J-space is interesting and powerful, its connections to the global workspace and to access consciousness have been somewhat overstated.

Srijani Saha

Harvard University

A Unified Account of Lightness Illusions via Edge-Based Reconstruction of Natural Images

The human visual system transforms patterns of light into rich perceptual experiences, where what we see is a construction that goes beyond simple measurement. Lightness illusions—where identical parts of an image can appear dramatically different depending on context—provide a window into these processes. Here we leverage a deep learning framework to investigate the constructive processes that give rise to lightness illusions, introducing the core computational goal of edge-based image reconstruction. Specifically, we demonstrate that autoencoder models trained to reconstruct natural images based only on an edge-based image representation naturally recapitulate a wide range of lightness illusions, which were previously assumed to require distinct mechanisms, inference over lighting sources, and explicit three-dimensional scene representation. These results offer a simpler, unified account of diverse lightness phenomena as emerging naturally from surface filling-in mechanisms, and broadly provide a framework for understanding the computational principles that underlie our perception of the visual world.

Paul Linton

Columbia University

Experience Before Inference: A Computational Account of Visual Experience

What are the building blocks of a computational model of visual experience? First, I argue that a model of visual experience should distinguish between three levels: (1) our low-level, non-inferential visual experience, (2) our inferences about our low-level visual experience, and (3) our inferences about the external world. The problem is that we often mistake the second level (our inferences about our low-level visual experience) for the first level (our visual experience itself), particularly in the context of visual constancies (depth, size, and color constancy). Second, I introduce QualiaNet as a computational model of stereo visual experience, combining two traditions in vision science: a simple, hand-crafted model of low-level visual experience feeding into a learned neural network responsible for higher-level inference. On this account, conscious experience is the input to a modern neural network model of vision, not something we should expect to emerge as the output of a sufficiently complex model.

Break (20 mins)

Sharif Kronemer

National Institutes of Health

The feeling of seeing: Insights into the neural origins of conscious sight from a case study of cortical blindness

Healthy conscious vision is associated with distinct perceptual features, including color, shape, and movement. After injury to the primary visual pathway, conscious vision can be impaired. However, some people with cortical blindness retain residual, degraded visual conscious perception. I will present an ongoing case study of a cortically blind person (right homonymous superior quadrantanopia) who experiences a rarely reported form of conscious awareness for images presented in his blind field: a non-visual “sensation” or “feeling of seeing”. Using fMRI and MEG, this investigation highlights: (1) secondary pathways and compensatory mechanisms involved in processing visual input following injury, and (2) the neural mechanisms underlying the conscious experience of sight. This work also explores how eye behaviors, such as pupil size and blinking, can inform the subjective experiences of conscious vision.

Marisa Carrasco (Psychology Keynote)

New York University

Acting Without Seeing: Eye Movements Reveal Visual Processing Without Awareness

Distinguishing between conscious and unconscious processing remains a central objective in cognitive psychology and neuroscience. Experimental protocols probing the qualitative dissociations between visual perception and action offer crucial insights into the boundaries of awareness. Because humans exhibit continuous, highly accurate eye movements, oculomotor responses provide an appealing model system for inferring underlying visual and cognitive processes. In this talk, I present three studies utilizing eye movements as a sensitive metric of unconscious processing across motion discrimination, feature-based attention, and emotional face perception. Methodologically, the first two studies combined binocular rivalry flash suppression with monocular adaptation, whereas the third utilized continuous flash suppression validated by objective and subjective measures. Finally, I relate these findings to brain pathways for perception-action dissociations and leverage them as a strong cautionary tale about the risks of inferring conscious perception solely from oculomotor and pupillary responses when relying on no-report protocols.

Lunch (On Your Own)

Biyu He (Neuroscience Keynote)

New York University

Mario Belledonne

MIT

Awareness through Goal-Conditioned Abstraction: Approaching the Visual Frame Problem with Multigranular Optimization

Human vision has the incredible ability to solve the frame problem: to pick out relevant items in the world in near real-time. In contrast to modern AI approaches, which utilize implicit, black-box solutions, here we introduce Multigranular Optimization (MO), an explicit and interpretable online algorithm that produces goal-conditioned abstractions over scene elements. Instead of encoding fixed resolution representations, MO flexibly distributes high-granularity to task-relevant items while coarsening irrelevant items to minimize wasteful computation. MO's goal-conditioned abstractions not only halve wall-clock runtime and error rate but, crucially, induce human-like awareness patterns. Coarse representations “explain away” the sensory signal of similar objects entering the scene, precluding their awareness; thus, MO provides the first trial-level computational account of inattentional blindness — one of the most foundational phenomena in psychology. This work begins to ground the algorithmic efficiency of perception with its rich experience.

Gal Vishne

Columbia University

Break (20 mins)

Matthias Michel

MIT

The Computational Architecture of Unconscious Processing and Why Studying It Matters

To identify the neural correlates and cognitive functions of consciousness, researchers compare conditions in which a stimulus is consciously perceived against conditions in which it is not. But this strategy faces a confound. Manipulations that make a stimulus unconscious, such as masking, binocular rivalry, and so on, do not disable just conscious processing. They also interrupt unconscious processing well before conscious processing could arise, as well as unconscious processing that runs parallel to consciousness. Observed differences between conscious and unconscious conditions may therefore reflect the disruption of unconscious processing rather than the contribution of consciousness. Solving this problem requires knowing where in the processing stream each consciousness-suppression method acts. I show that the architecture of unconscious processing can be empirically determined by testing which suppression methods block the effects of others. Applying this logic reveals that the most popular suppression methods used in the field interrupt processing early. This undermines many of the results of consciousness research over the past thirty years.

Megan Peters

UC Irvine + UCL

An inference machine for discovering consciousness in non-human systems

Much hullaballoo recently has focused on the signs, indicators, or empirical evidence that AI systems (or octopuses, or honeybees, or…) have consciousness. All of this assumes that we can interpret such indicators or tests the same in non-human systems as in the systems those tests were developed on — us — and that the “thing” those tests indicate (consciousness) is a well-formed, coherent, stable, real, and singular natural kind. Are these assumptions valid? In this talk I discuss new work extending and formalizing the iterative natural kinds strategy through combining epistemology and philosophy of science and belief updating with formal models borrowed from statistics and cognitive science to ask: how can we really test for consciousness in non-human systems, including discovering what we are even testing for?

Panel Discussion

David Chalmers, Marisa Carrasco, Biyu He, Megan Peters