Backpropagated plasticity: learning to learn with gradient descent in large plastic neural networks

T Miconi, J Clune, KO Stanley
How can we build agents that keep learning from experience, quickly and efficiently, after their initial training? Here we take inspiration from the main mechanism of learning in biological brains: synaptic plasticity carefully designed by evolution to produce efficient lifelong learning. We show that plasticity, just like connection weights, can be optimized by gradient descent in large (millions of parameters) recurrent networks with Hebbian plastic connections. Applied to the task of arbitrary natural image memorization, recurrent plastic networks with more than two million parameters can be trained to memorize and reconstruct sets of novel, high-dimensional (1,000+ pixels) natural images not seen during training. Surprisingly, the trained networks exhibit highly structured plasticity, in contrast with traditional, homogenous auto-associative networks. Crucially, traditional non-plastic recurrent networks fail to solve this task, and require orders of magnitude more training to partially solve a considerably simpler version of it. In conclusion, backpropagated plasticity may provide a powerful novel approach to the learning-to-learn problem.
Pub. Info: 
NIPS 2017