Spike-based temporal coding boosts object recognition accuracy by 50%

The Thousand Brains Theory (TBT) models object recognition via active sensor movement, building evidence contact by contact. Current implementations use dense floating-point vectors for each contact, but discard the order in which features are encountered — a critical loss of spatial context. Joy Bose proposes replacing these vectors with rank-order spike packets: each contact triggers a burst where the most strongly activated neuron fires first. The time gap between successive bursts implicitly encodes sensor displacement, while a biologically plausible STDP learning rule encodes traversal direction into synaptic weights. A dynamic parameter lambda adjusts reliance on earlier versus recent contacts, adapting to each object's geometry. Implemented in ~450 lines of NumPy, this temporal coding method preserves the sequence of features without explicit coordinate calculations.

Three synthetic experiments validate the approach. First, on objects with identical features arranged differently, temporal coding achieves perfect discrimination while dense vector methods perform at chance. Second, across all tested noise levels, the spike-based approach maintains a 30-50 percentage point advantage. Third, the adaptive lambda converges to distinct values that reflect the geometric complexity of each object. These results suggest that borrowing principles from biological spike-timing-dependent plasticity can make sensorimotor inference both more accurate and more robust. End-to-end evaluation on Monty's YCB benchmark remains future work, but the framework offers a principled path toward integrating temporal dynamics into reference-frame-based object recognition—an important step for robotics and neuromorphic computing.