From Local Learning to Global Prediction Through Layered Surprise Cascades

What the math gives to ML

The paper proposes a biologically local alternative to backpropagation in which each layer learns from paired and mismatched input-label examples, while recurrent bottom-up, top-down, and lateral signals create a layered surprise state. Its transferable asset is the inverted Forward-Forward objective: correctly paired examples are driven toward low layer activity, whereas mismatched examples are driven toward high activity, making activity itself a local compatibility or surprise score. A practical neural-network translation is a recurrent multilayer classifier whose layers train independently with stop-gradient interfaces and whose inference-time prediction is obtained by comparing the total surprise of candidate labels. This could reduce activation storage and enable asynchronous or continual layerwise training, but it must be tested against backpropagation because local objectives may sacrifice representation quality.