A personal note: this one has fascinated me for decades. Long before it was practical to build anything from it, the question of how biological systems do so much with so little was the thing I kept circling back to. The November 2024 date marks when it became formal research here — not when the interest started.
The Efficiency Gap
Modern AI has achieved extraordinary results, but much of that progress has depended on increasingly large models, dense computation, extensive training data, and substantial energy consumption.
The human brain presents a striking contrast. It supports perception, memory, learning, language, and adaptive behavior on roughly the power consumed by a household light bulb. This is not a direct benchmark — brains and AI systems perform different tasks on fundamentally different substrates — but it is strong evidence that useful cognition does not necessarily require continuous, uniform computation at enormous scale.
We're investigating whether selected principles from biological cognition can inform more efficient and adaptive AI architectures.
Beyond Simply Imitating Neurons
The objective is not to construct a literal digital brain, or to assume every biological mechanism is computationally optimal. Biology evolved under constraints that differ sharply from those of digital hardware.
Instead, the work focuses on extracting mechanisms that can be stated as testable computational hypotheses:
- Sparse activation. Engaging only the components needed for the current situation.
- Event-driven processing. Doing work when meaningful changes occur, rather than continuously recomputing everything.
- Predictive processing. Allocating attention and computation to unexpected or poorly predicted signals.
- Gated working memory. Selectively deciding what should enter, remain in, or leave active memory.
- Specialized processing. Coordinating multiple capabilities rather than applying one uniform mechanism to every problem.
- Complementary learning systems. Combining fast episodic learning with slower concept formation and consolidation.
- Local adaptation. Updating relevant parts of a system without repeatedly retraining the whole architecture.
Current Lines of Inquiry
Three questions have taken up most of our attention in this area.
Does anything like backpropagation happen in a neuron?
Backpropagation drives modern deep learning, and it demands a precise global backward pass — symmetric weights, a separate error channel, gradients computed across the entire network. Biology has no obvious mechanism for any of that.
So the question is whether real neurons accomplish credit assignment some other way: through local dendritic error signals, feedback that doesn't require weight symmetry, or prediction mismatch resolved layer by layer. If a biologically plausible approximation exists and holds up under test, it changes what learning has to cost.
Can neural activity be modeled as a queue of events?
A neuron that fires when it crosses threshold is an event, not a matrix multiply. Treating activity as a queue of discrete events — rather than dense synchronous updates across a whole tensor — makes sparsity and asynchrony native to the model instead of bolted on afterward.
This connects directly to event-driven processing above, but as an implementation strategy rather than an abstraction. The open question is where queue and scheduling overhead begins to consume the savings it was meant to produce.
How much does three-dimensional structure matter?
Biological cognition happens in space. Dendritic trees have geometry, signals take real time to cross physical distance, and wiring cost constrains what can connect to what. Most artificial architectures discard all of it — a layer is a flat set of units with no position.
Whether that spatial embedding is incidental to biology or load-bearing for its efficiency is one of the more interesting open questions in the field, and one we keep returning to.
Research Questions
- Can sparse activation reduce computation without significantly reducing task performance?
- Can prediction error serve as an effective attention and routing signal?
- Can credit assignment occur through biologically plausible mechanisms — local dendritic error signals, feedback without weight symmetry, layer-local prediction mismatch — closely enough to substitute for backpropagation?
- At what scale does event-queue scheduling overhead overtake the savings from asynchronous, sparse activation?
- Is the spatial embedding of biological structure incidental, or load-bearing for its efficiency?
- Does selective memory updating improve long-horizon reasoning and control?
- Can separate fast and slow learning mechanisms reduce catastrophic forgetting?
- Do heterogeneous systems outperform monolithic models when measured by useful performance per unit of energy?
- Which efficiencies survive implementation on conventional hardware?
- When does the overhead of routing and sparsity eliminate the expected benefit?
Experimental Perspective
Claims of biological inspiration should rest on controlled experiments, not analogy. Each proposed mechanism is compared with an appropriate conventional baseline and evaluated independently before being folded into a larger architecture.
Relevant measurements include task accuracy, adaptability, sample efficiency, retention, active computation, memory movement, latency, and actual system-level energy consumption.
Negative findings matter just as much. A biological abstraction that adds complexity without measurable benefit should not be retained simply because it sounds scientifically compelling.
Why It Matters
Current AI development often treats larger models and additional computation as the most reliable path to greater capability. That approach will remain important, but it shouldn't be the only one.
The brain's efficiency may not reduce to a single transferable discovery. It may arise from the interaction of sparsity, prediction, selective memory, specialization, and learning across multiple time scales.
Our goal is to determine which of these principles produce practical improvements in artificial systems — more useful cognition with less computation, less data, and greater adaptability.