This post is a summary of our recent preprint titled “Towards full-stack functionalist theories of consciousness: Identifying the functional profile”.
What is consciousness? An answer to this question is a theory of consciousness, and we aren’t going to add to the list of 500+ existing theories (today) 🙂 Instead, as one probably should when there’s such a proliferation of theories, that apparently don’t talk sufficiently well to each other, we’re going to try and think from scratch on how a theory of consciousness should be built in the first place. Of course, no one can think in a vacuum so we state upfront that this approach is heavily influenced by the zeitgeist of cognitive (neuro)science.
"She regained consciousness", "I'm conscious of the noise", "I'm conscious of making a mistake", "I'm self-conscious", "We need greater environmental consciousness", "I'm conscious of a red apple" — the word “consciousness” is used in many contexts, e.g., to talk about wakefulness, signal-acquisition, self-model, and understanding. Many of these concepts are dealt with by cognitive (neuro)scientists. However, there’s one aspect — characterized by “it feeling like something”, that “it feels like a movie playing”, that “an apple is right there with me in the world”, that “the seen color red is more than its wavelength” — that dominates the narrative when people currently brood on the topic of consciousness. This aspect is not the bread-and-butter of cognitive (neuro)scientists, as moving away from subjective experience is a central part of objective science practice. However, these reports of conscious experience (which can be interpreted in smart ways) are data that need explanation a theory of consciousness should provide.
What kind of explanation are we seeking? Descriptions of mental phenomena are what we are considering. Mental here does not imply an absence of a bodily correlate (e.g. pain feels like it is in the arm), but the idea being that the brain (or more broadly, the nervous system, for being general) is processing those signals leading to the experience and description thereof — get rid of the brain, the notion of pain is gone; get rid of the arm, you can still experience phantom pain! So ok, somehow neural activity has to give rise to conscious experience — that’s the explanation we’re seeking, i.e., requiring an understanding of what neural processes (cognitive or otherwise) lead to conscious experience. More precisely, we need to outline the measurements of conscious experience (reports and inferences thereof, and properly-calibrated behavioral and neural metrics), termed the explanatory profile. Then we need to outline the processes that consciousness is associated with, termed the functional profile. Finally, we have to explain how the computations in those processes are linked to the said measurements (the Attention Schema Theory description below is an example). This final step is critical in tying the stack together — we don’t just want theories that only deal with phenomenology or ones that only deal with neural processes.
As a first step, we expand on how to construct the functional profile, and present our initial findings. The aim here is not to outline the functions of consciousness as much as it is to find processes that are tightly coupled with measurements of consciousness. Formally, we want to identify processes Y that are found to engage entity X (external stimuli or internal information) differently when the participants indicated they were aware of X as opposed to when they were unaware of X. The idea is that these processes Y could then serve as building blocks towards a computational model that we could find the signatures of consciousness in (this will become clearer in the Attention Schema Theory example below). Note that finding processes Y where such a consciousness-related difference is not found is useful too — it constrains the computational modeling to not associate consciousness with every process in an agent! A final point to note re not associating consciousness with every process is that simpler explanations related to X not being signal-matched across the aware/unaware conditions are counterproductive. We thereby require studies to match the objective performance on discriminating awareness-related features of X to be matched across conditions, such that less of X being registered by one’s sensory cortex is not a reason for the difference seen in outcomes of process Y engaging X. Note that there are smart ways of getting around the X-signal-matching requirement and still not falling prey to the simpler explanations due to not signal-matching, as best demonstrated in Tsushima et al. 2006. Our paper provides more details on that front.
A characteristic example of studies showcasing a process Y being tightly-coupled with consciousness as outlined above is Persaud et al. 2011. Patient GY with blindsight was studied. GY’s left primary visual cortex is extensively damaged. Although he can discriminate between stimuli in his right visual field, he reports no awareness of those stimuli! The process Y that Persaud et al. were interested in studying was metacognitive evaluation; confidence judgements in simpler terms. They showed GY two gratings in the left or right visual field and asked him to indicate if a specific orientation was relatively up or down in that configuration. So, the X here is orientation information. They adjusted the contrast of the gratings in the two visual fields such that GY’s orientation judgement performance was matched across two fields (the conditions; right visual field would have a higher contrast stimuli than the left). They found that GY said his orientation judgement was guesswork when the stimuli were presented in the blind/right visual field and based on awareness (seen) in the sighted/left field. In the trials, he was also asked if he would wager money on his decisions — he could get 50p with 50% if he chose “Coin flip” or he could get 50p depending on the accuracy of his judgement if he chose “Response”. This measures his confidence in his judgments. What Persaud et al. found was he chose “Response” ~75% of the time in both conditions. This indicates that GY is confident that although he does not see what’s there he can judge properties of objects there, possibly due to his life experience being of a similar kind. What was surprising was, his blind visual field led to less calibrated confidence judgments — his “Response” judgement happened far more times when his orientation judgment was correct, in the sighted visual field than in the blind visual field. We can thus conclude that the process of confidence judgements — especially its calibration — is associated with being conscious of the properties being judged!
We ran a manual sweep, as described in our paper, to find papers in the cognitive (neuro)science literature which adhere to the above-mentioned experimental structure. We identified three primary sources which qualify — [consciousness associated with] Tsushima et al. 2006 (attentional control), Persaud et al. 2011 (metacognitive evaluation), and [consciousness not associated with] Keizer et al. 2015 (simple feature-location binding). We also identified other papers, which lack one of the conditions (e.g. signal-matching could still be a confound) that are indicative of associations with error correction, object-file updating, and semantic priming (described in our paper). These studies are the starting points towards building the functional profile.
In terms of how to move from this functional profile to the explanatory profile, we take inspiration from a worked out example — the Attention Schema Theory (AST) — that starts from attentional control and attempts to provide an explanation for a particular type of problem reports about experience (related to the meta-problem of consciousness), e.g., why does the experience of seeing an apple feel magical or non-physical? In AST, Graziano proposes that attentional control requires an attention schema (similar to how body control requires a body schema). This is a typical computational modeling move. The surprising, abductive, jump comes in when he proposes that the attention schema is related to the content of awareness! For example, when you look at the apple, in addition to your brain having information about the apple, the attentional control system — which is locked on to the apple, undisturbed by other features in the world — also contains a model of your attention being focussed on the apple and not on the other features, modeled as you interacting with the apple! There’s beautiful work showing that we model others’ gaze in such a situation as if rays are emanating from their eyes towards the object. However, this interaction does not have a physical correlate! When you grab an object with your hand, the interaction has a physical correlate, i.e., your hand. In this case, “grabbing the object” or “being grabbed by the object” via attention thereby feels magical / non-physical! These abductive jumps — proposing experience could be addressed via a consciousness-coupled process being modeled, and that representations emerging in that process could be linked to the structure of experience — are key in our framework towards building full-stack theories of consciousness.
AST provides a nice starting point on how to construct full-stack theories. However, there’s at least three fronts on which we need to make progress. One, we need to expand the set of processes to the entire “functional profile” in terms of building computational models wherein we’d look for correlates of experience. Currently, we’d need to jointly model attentional control and metacognitive evaluation to get started. Two, we need to expand the “explanatory profile” to account for rich experience structures, beyond just the sense of magic and non-physicality. This is a hard task, but putting together a set of informational structures that many phenomenological investigations — microphenomenology, the study of pain, and of course, problem reports themselves — could prove to be a good start. Ideally, these structures need to be translated into dynamical systems / representational analysis language such that above-mentioned computational models can interface with them. Three, finding the right associations between information found in the computational model and the rich experience structures would be critical to round up the stack; this we foresee as highly non-trivial but whose complexity would only become apparent once the computational model is built and can be parsed. Finally, once such full-stack theories are ready, we need to think about what new predictions it makes in terms of the information the brain might have in conducting these processes and what that might mean for predicting experience in novel situations.
We’ll be working on these fronts in the coming months. If you want to attack these problems seriously, please reach out to us for collaborations, or come work with us at Lossfunk! Also note that building a theory of human consciousness is just a starting point. Its utility would be in thinking about altered states of consciousness, other species’ consciousness, and AI consciousness, in terms of their welfare. We will also be thinking of these problems in the future.
Full paper link: https://arxiv.org/abs/2609.33366





