Learning that
doesn’t stop.
The division’s long-horizon research programme into how learning systems assign credit, retain what they know, and pay for it in energy.
Backpropagation works.
The brain doesn’t use it.
Every part deep inside a network has to work out how it contributed to a global outcome. Backpropagation answers that exactly, by means no synapse could physically use: updates that depend on other layers, two locked phases, and an end-to-end differentiable loss. That has been the field’s standing objection for three decades. It is still open.
Vestigium takes its name from the trace a thing leaves behind. It asks whether biology settles credit assignment and continual learning by a principle our gradient methods structurally cannot express, or whether it is a constraint-bound approximation to something we already understand. The field is genuinely split. Either answer is worth the decade.
The stake is commercial as much as scientific. A learning rule that needs no backward pass through the whole network, and no second locked phase, can run where the data already is: on the platform, inside the power budget, after the system has shipped. Every deployed model that has to send its experience home before it can learn from it is waiting on this question.
Four lines, all worked against the same joint: what a system can change using only what is available at the point of change. They are pursued together because they fail together. A rule that assigns credit locally but forgets last week has solved nothing.
Credit Assignment
How a system works out which of its own parts deserve credit for an outcome, and whether that can be settled from information already present where the change has to be made.
Continual Learning
Taking on something new without destroying what was already learned. The failure mode is well understood; a general remedy is not.
Representation
How the shape of an internal representation governs what a system can hold at once, and what it gives up to hold it.
Efficiency
The one existence proof of general intelligence runs on roughly twenty watts. We treat that as a design constraint.
Findings, methods and results are held internally. We are glad to discuss the programme in general terms, and in detail under a suitable agreement.
A result counts here only if it clears three bars: a baseline trained under the same budget, a condition where the effect ought to disappear and does, and a protocol someone else can run without asking us what was meant. Most ideas do not get that far. The programme is built to find that out early, while it is still cheap.
Fix the claim
Hypotheses, arms, controls and compute budget are written down before anything runs, together with a statement, set in advance, of what would count as the idea failing.
Run to budget
A fixed allocation per question. When the budget is met the result stands as it is. No quiet extension until the numbers improve.
Record either way
A negative result is recorded with the same weight as a positive one. The programme has been redirected more than once by an idea that did not survive its own test.
Long-horizon does not mean open-ended. These are the four places a result gets put to work, and they are why the questions were taken in this order.
Systems that adapt in the field, on the platform, without a round trip to a datacentre or a retraining cycle.
Learning and inference under a hard power budget. Demonstrations can ignore it. Deployments cannot.
Taking on new information without catastrophic forgetting, and without freezing the weights on release day.
Fundamental learning research held in the UK. The alternative is licensing it back from a foreign frontier lab.
Interested in Vestigium.
Research collaboration, funding, or a learning problem that will not sit still. The Machine Intelligence division would like to hear from you.
hello@viamachina.ai