DeepMind just published a mind blowing paper: PathNet

When working through Michael Nielsen's deep learning tutorial[1], this was exactly the question I had. I used to think deep learning involved 1000s of layers. Turns out that training a network a few layers deep can already be a real pain.

Currently we seem to be training one-off networks that are very good at a single thing (e.g. recognising hand written digits in 28x28 pixels), but we haven't been able to really connect multiple networks together to get to more generalised applications.

I didn't see this model coming, but it show DeepMind is taking an interesting direction here, potentially leading to more generally applicable ML capabilities.

[1]: http://neuralnetworksanddeeplearning.com

PathNet paper: https://arxiv.org/abs/1701.08734

For artificial general intelligence (AGI) it would be efficient if multiple users trained the same giant neural network, permitting parameter reuse, without catastrophic forgetting. PathNet is a first step in this direction. It is a neural network algorithm that uses agents embedded in the neural network whose task is to discover which parts of the network to re-use for new tasks. Agents are pathways (views) through the network which determine the subset of parameters that are used and updated by the forwards and backwards passes of the backpropogation algorithm. During learning, a tournament selection genetic algorithm is used to select pathways through the neural network for replication and mutation. Pathway fitness is the performance of that pathway measured according to a cost function. We demonstrate successful transfer learning; fixing the parameters along a path learned on task A and re-evolving a new population of paths for task B, allows task B to be learned faster than it could be learned from scratch or after fine-tuning. Paths evolved on task B re-use parts of the optimal path evolved on task A. Positive transfer was demonstrated for binary MNIST, CIFAR, and SVHN supervised learning classification tasks, and a set of Atari and Labyrinth reinforcement learning tasks, suggesting PathNets have general applicability for neural network training. Finally, PathNet also significantly improves the robustness to hyperparameter choices of a parallel asynchronous reinforcement learning algorithm (A3C).

Biology has figured this out by separating developmental evolution from learning. Developmental evolution evolves the genotype-phenotype map. For neurogeneisis evolution has evolved genes that generate the brain. Learning is then a fine tuning on hundreds of millions of years evolutionary algorithms. We can't get an Ape to speak but almost no human is impaired. Interestingly, the process of neurogenesis is controlled by genes which together form gene regulatory networks and when you represent those mathematically you discover that they're basically a special kind of neural network. Basically neural networks are a universal computational design that have been functionlly reproduced at least twice in nature.

> A plausible requirement for artificial general intelligence is that many users will be required to train the same giant neural network on a multitude of tasks. This is the most efficient way for the network to gain experience, because such a network can reuse existing knowledge instead of learning from scratch for each task.

Is this the hypothesis or already well-known? I'm not so sure if one giant DDN is better than a multitude of localized and specialized DDNs.

Pathnet is super interesting, but the authors of the paper don't make claims like this blog post does about being general AI, and this blog post doesn't back up those claims well either.