Rate-Coding Bundle Memory and Symbolic Computation in the Brain
A long-standing debate in cognitive science is whether the brain is connectionist or symbolic. A new preprint argues the answer is both. Its model, Rate-Coding Bundle Memory, uses continuous rate coding to represent symbols and a bundle memory to store and retrieve them, combining connectionist substrate with symbolic competence in a way its authors say is neurobiologically plausible.
The paper by Teun van Gils, Rowan P. Sommers, Markus Ostarek, and Peter Hagoort was posted to arXiv as 2608.29189. Hagoort directs the Max Planck Institute for Psycholinguistics, and the group works on the neural basis of language and cognition. The model is grounded in the Symbolic Subsystem Hypothesis, which holds that the brain implements a symbolic subsystem within its fundamentally connectionist nature.
How a bundle memory stores symbols
Symbols in this model live in a continuous space, represented by firing rates. A bundle memory stores and retrieves these symbols. The claims are competence on a wide range of cognitive phenomena: one-shot learning, pattern separation, and the binding problem, where distinct features must be bound together into a single representation without being confused.
| Cognitive phenomenon | How RCBM handles it |
|---|---|
| One-shot learning | Symbol stored from a single presentation |
| Pattern separation | Similar inputs map to distinct bundles |
| Binding problem | Features bound in a shared symbolic representation |
The model is deliberately unlike a pure neural network that stores everything in slowly learned weights, and unlike a symbolic system that assumes discrete tokens. It occupies the middle: rate-coded continuous symbols, retrieved by a memory that can bind and separate.
Why this matters for consciousness and the substrate
The site’s coverage of spiking neuron models and consciousness and the recurrent processing test for neuromorphic consciousness focuses on the substrate of computation. This paper feeds the same thread from the other end: it asks what computational organization gives access to the symbolic flexibility that more rigid networks lack.
For functionalist emergentism, the significance is that symbolic competence is presented as a computable property of a rate-coded substrate, not as something outside mechanics. That is directly the project’s view, that higher-level cognitive abilities emerge from substrate independent computational organization.
Comparison to The Consciousness AI
The Consciousness AI project’s Neutral Core and substrate console ask how abstract computational properties can be implemented in a physical substrate. A bundle memory that is neurobiologically plausible in its rate coding and still symbolic in its competence gives a concrete answer pattern to the question the project works on. The project treats the implementation of such structure, not the identity of the substrate, as the carrier of mind.
The flag article on the race to define artificial consciousness tracks how different theories assess what architecture consciousness requires. This paper is an architectural proposal: a way for a connectionist substrate to support symbolic control, which is the kind of structure several theories treat as necessary.
Limits
The paper is a model proposal with stated capabilities, not a completed experimental validation against neural data. One-shot learning, separation, and binding are demonstrated in the model, and the neurobiological plausibility is argued. Whether cortex actually implements bundle memory this way remains open. The value for the site is computational: a concrete, rate-coded account of how a substrate can be symbolic.