Spiking neurons can discover predictive features by aggregate-label learning

The brain routinely discovers sensory clues that predict opportunitiesor dangers. However, it is unclear how neural learning processes canbridge the typically long delays between sensory clues and behavioraloutcomes. Here, I introduce a learning concept, aggregate-labellearning, that enables biologically plausible model neurons to solvethis temporal credit assignment problem. Aggregate-label learningmatches a neuron’s number of output spikes to a feedback signal that isproportional to the number of clues but carries no information abouttheir timing. Aggregate-label learning outperforms stochasticreinforcement learning at identifying predictive clues and is able tosolve unsegmented speech-recognition tasks. Furthermore, it allowsunsupervised neural networks to discover reoccurring constellations ofsensory features even when they are widely dispersed across space and time.