What Predictive Coding Means Here
Predictive coding is a family of ideas where systems learn by minimizing prediction errors between expected and actual activity. In Yudi Labs research, predictive coding is studied as a training direction for sparse recurrent architectures like CMP.
Why It Matters
Current AI systems are dominated by global backpropagation. Predictive coding offers a different learning lens: local error signals, layer-wise prediction, and potentially more brain-aligned update rules.
Two Interpretations
Architectural Predictive Coding
The model architecture explicitly contains prediction and error populations, with iterative inference and settling.
Training-Algorithm Predictive Coding
The architecture remains mostly unchanged, but training uses local prediction-error updates.
Current Priority: Yudi Labs currently prioritizes training-algorithm predictive coding as the lower-risk, scalable path before deeper architectural predictive coding.
How It Connects to CMP
CMP v1.9 uses predictive sparse-pattern output as part of the research direction. The goal is to test whether local predictive learning can train sparse representation systems at progressive scale.
Open Questions
- Can predictive coding train CMP at 10M-122M scale?
- Can learned sparse encoders develop semantic structure under PC?
- Can PC-CMP show continual learning advantages?
- Can PC-CMP support few-shot learning?
- Can it generalize across modalities?