Efficient online learning with predictive coding networks
PCN-TA, a predictive-coding online learner that keeps latent states across consecutive frames of a sensory stream, cutting the cost that usually makes predictive coding expensive.
- 10%fewer weight updates than backprop
- 50%fewer inference steps than baseline PC
Problem
Robots at the edge need online learning from streaming sensory data. Predictive coding offers local, Hebbian-like updates that suit neuromorphic hardware, but it pays for them with many inference iterations per training step.
Approach
PCN-TA preserves latent states across temporal frames, so inference on each new frame starts from the previous frame's state instead of from scratch. It exploits the temporal correlation that a robot's input stream has anyway.
Evaluation
Online learning on the COIL-20 robotic perception dataset, against backpropagation and baseline predictive-coding networks.
Result
Accuracy comparable to backpropagation with 10% fewer weight updates, and 50% fewer inference steps than baseline predictive-coding networks.
Paper
Efficient Online Learning with Predictive Coding Networks: Exploiting Temporal Correlations
IROS 2025