Bio-inspired network research

MagicBrain

Research

A line of research into bio-inspired neural networks: spiking networks and local learning. The code is open (PyPI 0.8.0) and is not used in products.

The problem

Almost all applied AI today is built on large models that someone else trained. We want to understand other approaches too: what networks closer to biological ones can do, where their limits are and which of it can be checked.

Why the usual tools fall short

Working on top of someone else’s model does not explain how networks themselves work. How a network’s structure can form as it runs is still an open question. So we describe observable properties and engineering maturity, and make no claim that this approach beats existing ones.

How it works

What we show here is engineering maturity; there is no scientific result yet. For the classical Hopfield network, Lean 4 proves that its energy does not increase when neurons update one at a time. These proofs do not cover the TextBrain spiking network. There is also observable behaviour: the network recovers a stored pattern from a corrupted input. We make no claims about speed or text-generation quality: on the evaluation bench the spiking network does not yet beat a simple bigram model.

Demo

Both results run from the open repository: lake build compiles the Lean 4 proofs for the classical Hopfield network, and pytest -k pattern_memory checks the recovery of a stored pattern from a corrupted input.

Technology

All our products and libraries →

Where it fits

A checkable base for research

Formal Lean 4 proofs for the classical Hopfield network show how we check research code.

A long-term line of research

A direction we pursue to understand the field. It is not sold as a product.

What backs this up

Lean 4: 21 statements about the classical Hopfield network, with no gaps

It is proved that the energy of the classical Hopfield network does not increase when neurons update one at a time. This is a classical result, formalised anew: the proofs do not cover the TextBrain spiking network. We show that the proofs build.

Reproduce: lake build

Pattern memory: store a pattern, corrupt some bits, recover it

Observable behaviour of the memory on a corrupted input. Behaviour only, with no claims about effectiveness.

Reproduce: pytest -k pattern_memory

The code is open, so everything above can be checked in the repository. Open the code ↗

Work it out on your own figures

The AVA calculator works out whether automating a process pays off, on your figures, with no form and no price from us. If it does not pay off, it says so.

Estimate the payback

Tell us about your problem

Describe it in your own words. If it cannot be solved or will not pay off, we will tell you before any work starts.

In the reply we will say what we understood, what we need to clarify and whether it is worth doing.

or email us directly: hello@xteam.pro

New write-ups on Telegram