A new recurrent neural network model learns complex sequences the way real brains might, without relying on backpropagation.
Researchers built ELiSe (Efficient Learning of Sequences), a model designed to learn long temporal patterns, like the syllables in a birdsong, using only synaptic plasticity rules that are local, always-on, and don't require a separate error-correction phase. The model borrows two features from real cortical networks: a scaffold of connections already in place before learning starts, and dendritic compartments that let single neurons store and process more information than a standard point neuron. In tests modeling how birds learn songs, ELiSe reproduced the tangled, non-Markovian timing patterns in the training data and kept working after researchers introduced disturbances meant to throw it off. The authors also report the model holds up across different parameter settings, not just one tuned configuration.
Standard ways of training recurrent networks, including FORCE, RTRL, and backpropagation through time, work well but lean on error signals no real neuron has been shown to compute. Reservoir computing avoids that problem by only training the output weights, but it struggles as tasks get harder. ELiSe's pitch is a middle path: learning rules a real neuron could plausibly run, applied to sequences more complex than reservoir methods can handle.
This is a birdsong simulation, not a chatbot, so the near-term payoff is insight into how brains might actually learn, not a new tool for your app. The efficiency in the name is measured against biology, not against a GPU cluster.