Skip to content
Amol Kelkar
Say hello
All projects

pattern-machine

Pattern-discovery engine - long-running personal research on spiking-NN pattern formation.

Status
active
Started
2022
Focus
Research
Topics
neuroscience, machine learning, deep learning, research

How do structural patterns emerge from raw temporal sensory streams when the only available signal is local - no labels, no backprop through time? Pattern Machine framed this as a spiking-neural-network architecture with local Hebbian-style learning, asking whether repeated temporal regularities could self-organize into reusable spatial primitives. A multi-year research thread, now in its third iteration.

The 2022 iteration built a full patternmachine Python package with layers, patch and pattern grids, signal-grid sets, local learning, and pattern-similarity primitives, alongside a documented experiment series running from single-layer regression through 1D bouncing-ball dynamics, temporal-to-spatial mapping, high-dimensional multi-patch patterns, and pattern hallucination. Each experiment left behind animated GIFs, pickled trained patterns, and reproducible drivers - the thinking is auditable. The 2023 rewrite (under the i3AI org) collapsed that into a smaller, cleaner computational-graph core with explicit operations modules and a LIF-with-adaptive-threshold demonstration on PyTorch. Neither of the first two iterations produced a definitive answer to the underlying question, but the experimental artifacts show what worked and what didn't.

The third iteration, started in 2026, builds the theory one small pre-registered experiment at a time, each tied to what is known about cortex. Its first passing result tests context: a spiking unit learns which other patterns usually appear with its own, through inhibitory interneurons that each learn one familiar context, and then keeps responding only when its pattern appears in a combination it has not seen. The result holds on a fresh set of seeds. A later experiment adds pattern completion - a unit fires early when the patterns that usually accompany its own appear without it - while keeping that context signal, and it too holds on fresh seeds. A small reserve of units, woken only by input the other units do not explain, then gives a rare pattern a code of its own without taking anything from the frequent ones.