Feedback Feedforward Alignment

as a more biologically plausible learning algorithm alternative to backpropagation of errors, allows for flexible inference

Cortical feedback connections underlie a rich palette of perceptual phenomena—from imagination and hallucination to occlusion resolution—which motivated our framework. We introduce Feedback–Feedforward Alignment (FFA), a biologically grounded learning algorithm in which:

• Feedforward pathways optimize a discriminative objective (e.g., classification loss),
• Feedback pathways optimize an autoencoder-style reconstruction objective,
• Both streams serve as mutual credit-assignment graphs, aligning their updates through purely local synaptic rules and thus bypassing the weight-transport problem.

During inference, trained forward and feedback connections run in a loop—iteratively refining activations based on reconstruction errors—to produce emergent generative functions such as denoising, de-occlusion, hallucination, and vivid mental imagery. Across MNIST and CIFAR-10 benchmarks, FFA achieves strong performance , exhibits robustness to input noise, and replicates flexible visual inference without any backpropagation.

Related Publications

Talks

Poster Presentation