Conference Agenda

Consciousness science is a fundamentally interdisciplinary endeavor, combining neuroscience, philosophy, mathematics, and other disciplines. To get an overview of the current state of the field, we analyzed over 30000 papers from the last decades and created a map of consciousness science. This map was built from recent reviews of consciousness theories (Seth & Bayne, Kuhn 2024, Signorelli et al. 2021) to identify current trends and future directions. We used large language models to measure semantic similarities in the abstracts and automatically clustered the literature according to the topics discussed. This clustering revealed mostly isolated strands of research focused on developing separate theories and gathering evidence. We present an interactive version of our map of consciousness science that can be individually explored and searched.

From this analysis, we argue that more interdisciplinary and intertheoretical research is needed to strengthen the foundations of the field. We discuss the epistemic consequences of recent projects that aim to bridge this gap, such as adversarial collaborations (Melloni et al. 2023) and open science databases of experimental tests (Yaron et al. 2022). Based on perspectival realism (Massimi 2022), we conclude that consciousness science aims at theory convergence and a unified explanation of conscious phenomena. However, an alternative path to progress in the field may come from an acceptance of model pluralism and the possibility of identifying multiple modally robust phenomena that need to be explained.

In this paper, we consider how Cognitive Carrying Capacity (CCC) determines the extent to which artificial agents model the self within resource-constrained environments. We approach the question from an AI perspective, asking what cognitive mechanisms enable artificial agents to generate self-reports similar to those of their biological counterparts. We argue that the key mechanisms include attention, environmental modeling, and memory, all operating under the central constraint of energy limitations. Attention filters environmental stimuli through internal goals/costs and restricts the range of environmental models the agent can generate. This energy limit leads to the narrowed correlation of goals/costs with specific models of self as separated from the rest of the environment. Unlike more traditional accounts, ours sees the self/world separation as a function of energy constraints, aligning with Dennett’s view that the self is not a fixed entity but a dynamically evolving construct (Dennett, 1991). The Cognitive Carrying Capacity of an agent establishes a threshold that defines the boundaries of the self as one of a group of models the agent constructs. The larger the energy resources, the larger and more complex the self-representation becomes. By adjusting energy constraints in computational experiments, we explore how CCC predicts the self-representation threshold in multi-agent environments, allowing for a dynamic and adjustable scale of self-representation, resulting in agents "seeing" themselves in some cases as a sole agent, while in others as the whole universe. We argue that CCC limitations is the reason biological agents cannot see themselves as the whole Universe.

Recent brain imaging studies revealed complex representations of emotions in the human ventral visual stream and even in the early visual cortex. Do these findings mean that early perception may represent important mechanisms in the generation of the subjective experience of emotions, like fear? Using multi-voxel decoding of human neuroimaging data, we show that patterns of hemodynamic activity predictive of a specific “fear profile” (i.e., fear ratings reported by a given participant) can be observed in the ventral visual stream whether a participant reports being afraid of the stimuli or not. Similarly, we show that the same fear profiles can be decoded in pre-trained deep neural networks that were not trained specifically to recognize emotions, like fear. Further, we found that the subjective experience of fear was associated with the synchronization of multivariate information between ventral visual areas and prefrontal regions. Taken together, these findings support the view that the subjective experience of fear may depend on the relevant visual information triggering implicit metacognitive mechanisms in the prefrontal cortex.

We introduce phenomenomics, a novel research program related to computational (neuro)phenomenology that aims to be for consciousness science what genomics is for genetics. It proposes a large-scale, data-driven, computation-centric approach to everything experienceable by different (classes of) experiencers — their phenomenome. Phenomenomics models experiential worlds as ‘descriptions’ compressing regularities and invariances constructed by observers from their first-person perspective. Being about experience, descriptions encode latent geometries as law-governed as genomic sequences. The task is to reverse-engineer this high-dimensional mathematical space.

Phenomenomics works from two directions simultaneously. An experiencer-first trajectory mines descriptive corpora — phenomenological writings, large-scale surveys of perceptual diversity, quality-space datasets — to extract, parameterise, and model experiential regularities and invariances. Concurrently, a substrate-first trajectory builds models of neural dynamics, e.g., predictive processing models optimising for variational free energy minimisation. Phenomenomics aims to ultimately converge these trajectories, perhaps via hyper-model fitting.

An initial, tractable step on this path is extending existing work in computational phenomenology. For example, training and fine-tuning language models (LMs) can capture latent structures in corpora of experiential descriptions. Phenomenomics would see the LM architecture itself as discovering and encoding — via hidden layer activations, attention weights, etc. — the latent structure of phenomenomes.

These parameter spaces enable synthetic phenomenology: Training generative models (VAEs, GANs) to produce novel descriptions that aim to be experimentally indistinguishable from ground-truth human reports. In the bigger picture, these generative models are seen as priors for fitting substrate-first brain process models in a bidirectional hyper-model that reverse-engineers the dual-aspect geometry of experience.

Researchers use a variety of experimental tasks to study consciousness. However, it has been argued that different paradigms aimed at studying consciousness may in fact not manipulate and measure exactly the same construct. This could either result from measuring different aspects of consciousness, or occur because tasks engage confounding cognitive capacities such as attention, working memory or verbal report. Here, I present a novel ‘task ontologies’ approach to address this issue, aiming to improve the interpretation and synthesis of empirical studies.

Firstly, I argue that neural mechanisms cannot be used as objective arbiters for deciding whether two tasks measure the same construct. Subsequently, I discuss two requirements of (a) useful concept(s) of consciousness: (i) We need to have concepts that are neither too fine-grained (identifying a concept with one specific experimental task), nor too coarse-grained (erroneously lumping different aspects/capacities); (ii) We want a certain degree of flexibility allowing for tailoring concepts to explanatory needs. I explain how a novel ‘task ontologies’ approach can meet these requirements and demonstrate how this approach could be applied to the neuroscientific study of visual consciousness. In short, a categorization of task-induced behaviours is achieved by weighing different (empirical and theoretical) criteria to produce the best systematization of context-dependent behaviour. Finally, I discuss the main implications of this task ontologies approach, including a new conceptualization of validity for consciousness neuroscience–and cognitive neuroscience more generally.

In conclusion, the task ontologies approach provides an integrated empirical and theoretical methodology to advance the conceptual foundations of consciousness science.

Phenomenal structuralism claims that each phenomenal character is fully individuated in terms of its relations to other phenomenal characters. This challenges the widely held view that phenomenal character is an intrinsic affair, i.e., that what it is like to undergo a particular experience is solely a matter of that experience’s intrinsic properties. Rather than offering another direct argument for phenomenal structuralism, the aim of this talk is to develop an error theory about intrinsicalism that explains how we come to believe that phenomenal character is intrinsic, why we hold onto this belief, and why it is nevertheless mistaken.

To this end, I will appeal to several considerations and point out how they account for the persistence of intrinsicalist intuitions: First, introspection and memory suggest that experiences are independent and self-contained, fostering the impression that what it is like to undergo a particular experience is wholly independent of other experiences. Second, from an evolutionary perspective, treating phenomenal properties as non-relational properties is cognitively simpler and more efficient. Third, linguistic behavior further reinforces the impression that phenomenal character is intrinsic.

However, none of these factors provides substantial support for intrinsicalism, as I will show. The upshot is that phenomenal intrinsicalism is much less plausible than it is commonly taken to be. This paves the way for a less presuppositional approach to theorizing about the nature of phenomenal character in general and for phenomenal structuralism in particular.