Research
Building intelligence the hard way — and the only way that scales
Information is the persistence of deviation from expectation under degradation.
— The founding principle behind every NeoSynaptics system
Active Research
A prediction-error-driven 2D grid with passive depth stack. Measures how long deviations persist across blur-and-leak layers. The core engine for bio-inspired feature extraction from raw video.
View live demo →Surprise-driven spatial gating patterns. A coordination field modulates substrate activity based on local surprise magnitude — amplifying novel regions and suppressing expected ones.
EMA-baseline deviation readout for novelty detection. Ghost neurons track exponential moving averages and fire when the substrate deviates significantly from its own history.
Z-score bandpass and novelty filtering inspired by biological retinal processing. Combines depth-stack readouts with integrated residual signals for hierarchical feature selection.
Frontiers
Can a substrate with no learned weights extract features competitive with trained neural networks on standard benchmarks?
What is the optimal leak rate for maximizing information persistence while maintaining temporal sensitivity?
How does the coordination field interact with depth propagation — does surprise gating improve or interfere with multi-scale feature extraction?
Can ghost neuron readouts serve as a self-supervised signal for downstream learning without explicit labels?
Is there a theoretical upper bound on the information density achievable through passive depth propagation in finite-layer substrates?
Code & Papers
First publication forthcoming. All research code is open source.
View on GitHub →