Researchers from Oxford University have introduced a new algorithm for training neural networks that eliminates the need for both the traditional forward pass and backpropagation. It is called Mono-forward algorithm — a new approach to neural training by leveraging local errors for learning. This new algorithm is inspired by Geoffrey Hinton’s Forward-Forward framework. Conventional training relies on backpropagation. Mono-forward optimizes each layer using locally available information. It removes dependence on global error signals. It enhances biological plausibility and computational efficiency.

The researchers examined Mono-forward on multi-layer perceptrons and CNNs across benchmarks such as MINIST, Fashion- MINIST, CIFAR-10 and CIFAR-100. The results match or surpass the accuracy 0f traditional backpropagation methods. In addition, it provides reduced or more uniform memory usage, improved parallelization and comparable convergence rate.

It has improved efficiency of neural network training methods. This approach could pave the way for advances in neuromorphic computing and energy-efficient hardware. One can refer to the original paper: Mono-forward — Backpropagation Free Algorithm for Efficient Neural Network Training Harnessing Local Errors.