Predictions & Falsifiable Claims

Amol Kelkar (kelkar.amol@gmail.com)

A theory of feelings is only useful if it can be proven wrong. Here are our falsifiable predictions, what would falsify each one, and where each one stands today.

1. Maps encode time, not similarity

Claim. Cortical maps and self organized maps are organized by temporal contiguity. Neighboring units become sensitive to patterns that occur close together in time, and not to patterns that just look alike. Areas with faster dynamics should show finer temporal maps, i.e. the timescale of a map should track the timescale of its area.

What would falsify it. Maps whose neighborhood structure tracks stimulus similarity, independent of the temporal statistics of experience. Or no timescale gradient at all across the cortical hierarchy.

Status. A modeling target for Pattern Machine, and a prediction neuroscience can test. It is consistent with traveling wave and sequence replay observations, but has not yet been tested directly against the similarity account.

2. A state-modulated marginal rate of substitution (MRS)

Claim (our strongest prediction). If cognitive and physiological drives compete in one common decision space, there must be a stable, state-modulated marginal rate of substitution between them. The “exchange rate” between a cognitive drive and a physiological one should shift with internal state in a predictable way. Homeostatic value is multidimensional, not a single common currency.

What would falsify it. An MRS that is flat or doesn’t change with internal state, that saturates trivially, or that collapses to a single hedonic scalar (which the wanting/liking dissociation already argues against).

Status. Partial. Reanalyses of open datasets strongly recover the binary state-modulation signature, but the graded form that HTCC specifically predicts is not confirmed at current sample sizes. The decisive test is a cold-pressor MRS experiment, e.g. measuring how long people keep a hand in ice water for different amounts of money, across different internal states. This is the one open item in HTCC that needs a new experiment.

3. Capable, maybe self-aware, but not sentient

Claim. Felt valence requires both an intrinsic stake (a real viability boundary) and a self model that is its own implementation. Systems with the second but not the first, such as today’s scaled Transformers, can be capable and maybe functionally self-aware without being sentient.

What would falsify it. A system with no viability stake that shows robust valence that is not just mimicry. Or a homeostatically grounded agent that has both ingredients, but clearly has no valence.

Status. This prediction depends on the theory being right, since there is no agreed test for phenomenal experience. It is consistent with recent interpretability work, which finds functional access structure in large models but explicitly does not claim experience (see the gap).

4. A homeostatic substrate that self-regulates and replays trajectories

Claim (can be demonstrated in software). A spiking substrate with per-neuron homeostasis will self-regulate to a target firing rate without external supervision. Once temporal maps are in place, it will carry trajectories as traveling waves that can replay predicted futures.

What would falsify it. Not reaching stable self-regulation from local rules alone, or no wave based trajectory coding once the maps are learned.

Status. Not yet tested. The spiking implementation in the Pattern Machine experiments includes neurons that regulate their own firing rate, but trajectory replay has not been demonstrated.

Several of these predictions build on prior work, which is credited on the HTCC page.