Spiking neural network PCL+ predicts visual sequences with synaptic delays

Predicting the future from sensory input is a critical capability for both biological and artificial intelligence, yet it requires maintaining a memory of recent events. A new paper from researchers at (affiliations likely) proposes Predictive Coding Light+ (PCL+), a spiking neural network (SNN) that learns to predict visual sequences by leveraging two biological mechanisms: spike timing-dependent plasticity (STDP) and synaptic delays. STDP adjusts connection strengths based on the precise timing of spikes, while delays allow information to be stored temporarily. Together, they enable PCL+ to retain a short-term record of past inputs and use that context to forecast upcoming frames. The network operates entirely unsupervised, making it a step toward more biologically plausible learning.

The team tested PCL+ on two tasks. First, it reproduced classic findings from visual cortex sequence learning, validating that the model aligns with known neural dynamics. Second, on a challenging gesture recognition dataset, the network learned to "fill in" missing input—predicting what the next frames should look like even when parts of the sequence were occluded. This mirrors how biological systems compensate for gaps in perception. The results show that STDP combined with synaptic delays is sufficient for learning temporal dependencies. For the AI community, PCL+ offers a blueprint for building low-power, event-driven systems that can handle time-series data without massive datasets or backpropagation, opening doors for neuromorphic hardware applications.