Deep learning performs speech recognition by training millions of weights through backpropagation on massive datasets. A biologically inspired alternative — Assembly Calculus — now performs the same task with no weight training at all.
Correction note (2026-11-07): the abstract of arXiv:2603.16923 (Adelson, Sethu and Dang) introduces an Assembly Calculus speech framework with probabilistic mel binarisation, population-coded MFCCs, a multi-area architecture and cross-area update schemes, and says it detects phone (F1 = 0.69) and word (F1 = 0.61) boundaries without any weight training, and reaches 47.5 percent accuracy on phone recognition and 45.1 percent on command recognition. The no-weight-training statement is made for boundary detection, not for the recognition accuracies, and the cross-area update schemes used for those tasks are not described in the abstract. The essay says the system performs the same task as deep learning with no weight training at all, and adds 'no dataset larger than the signal itself', a dynamical system with no learned parameters, and the sufficiency of Hebbian association and winner-take-all; those statements go beyond the abstract. I compared this essay with the abstract only.
The system converts speech into spike patterns via probabilistic mel binarization and population-coded MFCCs, then processes them through hierarchical neuronal assemblies governed by Hebbian plasticity and winner-take-all competition. No backpropagation. No gradient descent. No dataset larger than the signal itself. Phone boundaries detected at F1=0.69; word boundaries at F1=0.61; phone recognition at 47.5%.
The numbers aren't competitive with state-of-the-art deep learning. That's not the point. The point is that they're nonzero. A dynamical system with no learned parameters, using only local plasticity rules and competitive inhibition, can segment continuous speech into phonemes and words. The capacity to parse speech is not in the weights — it's in the architecture. Hebbian association and winner-take-all are sufficient mechanisms.
This inverts the standard assumption about what learning contributes. Deep learning treats architecture as scaffolding and weights as the intelligence. Assembly Calculus treats architecture as the intelligence and weights as unnecessary. Neither extreme is the full story, but the existence of a working zero-training baseline reveals that the architectural contribution is larger than the deep learning paradigm suggests.