Spiking neural networks require hidden layers to solve certain classification tasks. Chemical reaction networks solve the same tasks without them.
Jaffard and Sbalzarini (arXiv:2603.12060) formalize learning in mass-action chemical kinetics and prove that CRNs without hidden layers achieve classification performance that SNNs provably cannot match without adding architectural complexity. On handwritten digit classification, the CRN outperforms the equivalent spiking network while using a structurally simpler architecture.
The mechanism is in what counts as a “layer.” A chemical reaction network operating under mass-action kinetics implements polynomial transformations of its inputs — concentrations multiply and combine according to stoichiometric coefficients. These polynomial interactions are richer per unit of architecture than the weighted sums and threshold activations of neural networks. What requires a hidden layer in a neural network to compute — a nonlinear combination of inputs — is available for free in a CRN because the reaction kinetics are inherently nonlinear.
The implication extends past computational efficiency. If raw biochemistry is more expressive per unit of architecture than neural circuits, then the universal assumption that information processing requires neurons may be backwards. Cells without nervous systems — bacteria, protists, plants — may not be performing simpler computation than neurons. They may be performing the same computation through a more expressive substrate, one that never needed the layered architecture we assumed was necessary.
The hierarchy was assumed, not demonstrated. Neural computation was placed above chemical computation because nervous systems are more complex, and complexity was equated with capability. The equivalence runs the other way.