Is Hinton's F-F Algorithm A Model for the Brain?

The Forward-Forward Algorithm

The Forward-Forward algorithm is a new learning procedure for neural networks. It replaces the forward and backward passes of backpropagation with two forward passes, one with positive (real) data and the other with negative data that could be generated by the network itself.

Each layer in this model has its own objective function which is to have: high goodness for positive data && low goodness for negative data.

The goodness can be measured by the sum of the squared activities in a layer, but there are other possibilities, including minus the sum of the squared activities. If the positive and negative passes could be separated in time, the negative passes could be done offline, which would simplify the learning process in the positive pass and allow video to be pipelined through the network without ever storing activities or stopping to propagate derivatives​¹​.

This new algorithm was proposed in response to the limitations of backpropagation. Backpropagation, although widely used, remains implausible as a model of how the cortex learns. It requires perfect knowledge of the computation performed in the forward pass to compute the correct derivatives. This limitation becomes evident when we insert a black box into the forward pass, rendering backpropagation impossible unless we learn a differentiable model of the black box. The Forward-Forward Algorithm, however, does not require backpropagation through the black box, making it a feasible solution in such cases​¹​.

The Forward-Forward algorithm (FF) is comparable in speed to backpropagation, and it can be used when the precise details of the forward computation are unknown. It also has the advantage that it can learn while pipelining sequential data through a neural network without ever storing the neural activities or stopping to propagate error derivatives. However, it is somewhat slower than backpropagation and does not generalize quite as well on several of the toy problems investigated, so it is unlikely to replace backpropagation for applications where power is not an issue. The two areas in which the forward-forward algorithm may be superior to backpropagation are as a model of learning in cortex and as a way of making use of very low-power analog hardware without resorting to reinforcement learning​¹​.

The Forward-Forward algorithm is a greedy multi-layer learning procedure inspired by Boltzmann machines and Noise Contrastive Estimation. It operates by having two forward passes that operate in the same way but on different data and with opposite objectives. The positive pass operates on real data and adjusts the weights to increase the goodness in every hidden layer. The negative pass operates on "negative data" and adjusts the weights to decrease the goodness in every hidden layer​¹​.

The algorithm employs a simple layer-wise goodness function. It normalizes the length of the hidden vector before using it as input to the next layer. This removes all of the information that was used to determine the goodness in the first hidden layer and forces the next hidden layer to use information in the relative activities of the neurons in the first hidden layer. These relative activities are unaffected by the layer-normalization​¹​.

The aim of the original paper is to introduce the FF algorithm and to show that it works in relatively small neural networks containing a few million connections. A subsequent paper will investigate how well it scales to large neural networks containing orders of magnitude more connections​¹​.

Bibliography

1. Ren, M., Kornblith, S., Liao, R., and Hinton, G. (2022). Scaling forward gradient with local losses. arXiv preprint arXiv:2210.03310​¹​.

AFTERWARD: The forward-forward algorithm exemplifies Hinton's commitment to understanding the brain and subsequently applying those insights to enhance AI. Hinton's work on this algorithm emanates from his enduring skepticism about backpropagation's congruency with brain operations. He suggests that the human brain, the most complex system known to us, is unlikely to adhere to the rigorous requirements of backpropagation. This assumption has driven Hinton's pursuit of an algorithm that mirrors plausible brain processes more closely.

Conclusion

The forward-forward algorithm promises an exciting frontier in the field of AI and deep learning. It emerges from the critique of the established norms, thereby opening new avenues for exploration and development in AI research. Hinton's commitment to understanding the human brain's working and applying those insights to develop more advanced AI algorithms is commendable. As the forward-forward algorithm continues to be examined and refined, it holds the potential to revolutionize our understanding of how learning might occur within the cerebral cortex, and by extension, how AI can be made more effective, adaptable, and 'brain-like.'