Spiking neural networks communicate through precise spike timing, and the way input values are encoded into spike patterns can determine whether similar concepts end up far apart in representational space. This 'Hamming cliff' problem is the target of a new preprint by Davies et al. (2025), who propose preprocessing inputs with Gray, thermometer, Johnson, and hybrid codes before feeding them to spiking networks on SpiNNaker hardware.

These compact spatial encodings are designed to preserve semantic distance: inputs that are close in meaning should produce spike patterns that are close in Hamming distance. The work directly addresses the semantic-spatial disparity that arises when naive binary encodings are used in neuromorphic systems.

The abstract does not report specific experimental results, so the effectiveness of the different encodings relative to one another remains to be seen. The contribution at this stage is the encoding framework itself, aimed at making spike-based representations more faithful to the structure of the data.