Sakana AI researchers Jeffrey Seely and Julian Gould introduced Augmented Lagrangian Predictive Coding (PC-ALM), a layer-local alternative to backpropagation. The method adds a Lagrange multiplier to constrain each layer.
According to the MarkTechPost synopsis, in linear networks PC-ALM recovers exact backpropagation gradients, and with an inference budget T = 2L it matches backpropagation for widths and depths from 8 to 128. In the specified reference configuration, the cosine similarity of gradients increased from 0,604 to 0,909.
On MNIST, the method trained residual multilayer perceptrons with a depth of 1000 layers, achieving a result approximately two points away from backpropagation. The source does not provide details of the experiments or the primary paper, so the scale of practical applicability remains unclear.
editorial commentary
Why it matters
The likely value of PC-ALM lies in testing the extent to which local learning rules can scale without full reliance on backpropagation. The next observable signal will be the publication of the primary work with reproducible experiments on a broader set of tasks. Significant uncertainty remains because currently only one synopsis is available without full context.