Smith et al. debut recurrent Forward-Forward algorithm for predictive coding

A team led by Andrew L. Smith (with Linxing Preston Jiang, Jason K. Eshraghian, Matthew S. Bull, and Stefano Recanatesi) has published a new paper on arXiv (2608.05481) that could reshape how AI learns from cortex-inspired principles. The paper, "From Local Learning to Global Prediction Through Layered Surprise Cascades," presents a recurrent variant of Geoffrey Hinton's Forward-Forward (FF) algorithm. Unlike standard FF, which increases activity for positive data, this variant uses an inverted objective that boosts activity for negative data. This simple tweak, combined with local contrastive learning and activity cancellation, lets the network build layered predictions that minimize surprise—without relying on error-coding neurons or explicit generative models.

The results are striking: the recurrent FF setup naturally produces predictive representations across layers, and it exhibits hallmark features of cortical computation, including top-down modulation and surprise signaling. This suggests that key principles of predictive coding can emerge from simple, local learning rules, offering a new theoretical bridge between neuroscience and machine learning. For AI engineers, this points toward training paradigms that are more biologically plausible and potentially more energy-efficient than backpropagation, since updates stay local. It also opens a path for neuromorphic hardware implementations that leverage surprise-driven dynamics. While the work is still theoretical, it adds momentum to the growing field of brain-inspired learning algorithms.