Spiking cognitive maps: overview

Public summary. Detail is withheld until submission under the disclosure policy.

Cognitive Map Formation and Energy Efficiency in Spiking Neural Networks with Local Plasticity

Stage 2 of Psyche, an open research project that builds brain-inspired components in stages with artificial general intelligence (AGI) as its long-term aim, and the project’s active line of research since 2026-10-23.

Now: P0 Novelty review and setup (2026-10-23 to 2026-11-15) · Next gate: A Novelty, 2026-11-15

The phase, gate and deadline tables and every date on these pages are generated from one file, tools/research_plan.json; dates are never typed by hand. The line builds on stage 1, the substrate, whose code and full record stay in this repository, summarised in PSYCHE.md. How the substrate led to this line, and what carries over from it, is told in Origins.

Motivation

Edward Tolman’s maze experiments with rats, summarised in Tolman (1948), are the classic evidence that animals learn more than the responses they are rewarded for. In latent-learning experiments, rats that explored a maze without reward showed little sign of having learned it; once food was placed at the goal, their performance soon matched that of rats rewarded from the start. In shortcut experiments, when the familiar route was blocked, rats chose routes they had never taken that led roughly toward the goal. Tolman argued that such behaviour is better explained by an internal map of the environment, a cognitive map, than by chains of learned stimulus-response associations.

Later work found cells whose activity is tied to space. Place cells in the hippocampus fire when an animal is in a particular part of its environment (O’Keefe and Dostrovsky, 1971). Grid cells in the medial entorhinal cortex fire at many locations that together form a regular triangular lattice across the environment (Hafting et al., 2005). These findings motivate brain-inspired models of map learning. They do not by themselves say how a map is learned.

The open question this line takes up is narrower: whether behaviours that an internal map would explain, and memorised responses would not, can arise in a network of spiking neurons that learns only through local synaptic plasticity (see the glossary), and what such behaviour costs in estimated synaptic operations.

Aim

The line is framed as a test. It asks:

• whether a spiking network, loosely inspired by hippocampal circuitry and trained only with local synaptic plasticity, shows on classic cognitive-map tests the behaviours that an internal map would explain and memorised responses would not;
• whether other families of agents, given the same experience, show them too;
• how many synaptic operations the spiking network needs, estimated analytically.

The plan does not depend on a positive answer. The environments, the battery, the comparisons and the analysis form a safe base that stands on its own. If the spiking model falls short, a documented account of where and why is kept as a result in its own right (roadmap.md).

Research questions

Questions only. Hypotheses, pass criteria and the method behind each answer are withheld until submission.

#	Question
RQ1	Can a network of spiking neurons that learns only through local synaptic plasticity form a cognitive map of its environment?
RQ2	Does it show behaviours that memorised stimulus-response associations would not explain?
RQ3	Do place-like representations emerge in the network without being built in?
RQ4	What does the behaviour cost, expressed as an analytic estimate of synaptic operations?

Planned contributions

These are plans, not results.

1. A spiking model that learns only through local plasticity, evaluated on the public battery below. What it does and does not show will be reported at release, whichever way it falls. The model stays private until its release under the disclosure policy.
2. An open cognitive-map test battery derived from classic animal-behaviour paradigms (latent learning, detour, shortcut, structural transfer, transitive inference), with comparisons across families of agents. The paradigms are published; how each is operationalised for an agent, and its pass criterion, are the project’s own and are pre-registered privately. See test-battery.md.
3. A measurement framework that reports every agent on three axes: behaviour on the battery, the structure of its internal representations, and an analytic estimate of its synaptic operations.

What this line does not claim

• No general claim about cognition. “Cognitive map” names a measurable class of behaviour: behaviour that memorised stimulus-response associations would not explain and an internal model of the environment would. Passing a test is evidence of that behaviour and nothing more. Words for mental states are not applied to the model; results describe behaviour and internal structure.
• Brain-inspired, not a statement about biology. The model is described as brain-inspired and, where it can be argued, biologically plausible. Matching an animal’s behaviour on a test is not evidence that animals solve the task the same way.
• Energy only as an estimate. Energy enters only as an analytic estimate of synaptic operations, read as a lower bound under stated assumptions. Nothing is measured on hardware. There is no general claim that spiking networks cost less than conventional ones; any comparison holds for this task, this metric and these assumptions.
• No priority claim. Whether the combination of aims is new is the question of a decision gate (gates.md). These pages make no priority claim, and a decided gate does not change that before release.
• No claim of general intelligence. AGI is the long-term aim of the whole project, named only as an aim. This line studies one measurable class of behaviour, and nothing on these pages presents its model as general intelligence or as a measured step toward it.
• No results yet. These pages describe a plan. Any number reported later will come from a captured results file.

The full discipline, including scope and writing rules, is in claims-and-scope.md.

Evaluation pipeline

The evaluation is model-agnostic. Every agent receives the same pre-generated exploration of the same environments and is scored by the same battery and the same metrics.

Read from top to bottom: seeded environments produce one shared exploration; the baseline agents and the spiking model all learn from it; the same battery tests every agent; and each agent is reported on the same three metrics, which together form the results.

The line’s code, the battery included, is developed privately and reaches this repository only as lagged releases, under cogmap/ and the disclosure policy. The spiking model stays private until its release.

Pages

Page	What it covers
Roadmap	Phases, layers and deadlines, generated from the plan file
Gates	The decision gates, the question each one answers, and what happens if one is missed
Test battery	The classic paradigms behind the battery, the environment families and the fairness rules
Claims and scope	What the line may and may not claim, what is in and out of scope, and the writing rules
Engineering	Coding standards, reproducibility, determinism, statistics and figures
Glossary	Terms used on these pages
Reading list	Classic background reading
Origins	How stage 1, the substrate, led to this line: what carries over, what does not, and where it points
Disclosure policy	What is public, what is withheld, and when it is released

Back to the repository README.