PathNet: Evolution Channels Gradient Descent in Super Neural Networks
-
Resources
-
Introduction
- Single neural network for multiple tasks = reusing knowledge
- Each user of giant network gets agents which learn task as efficiently as possible
- Agents learn to reuse parameters in environment by executing actions within neural net
- Agents work with other agents, sharing parameters where possible, or updating disjoint parameters if sharing isn’t possible
- Conventional neural nets: cost scales quadratically with model width
- PathNet: (theoretically) constant computation with respect to network width
- Fixed subset of larger network used for forward and backward passes
- PathNet
- Supports transfer, continual, and multitask learning
- Starts with all paths
- Tournament style genetic algorithm used to evolve paths
- Path trained for a few episodes using gradient descent with an RL algorithm
- Learning + evolution occur simultaneously
- As perfect performance is achieved, population converges to single path
- Optimal path for the corresponding task gets fixed
- Process is repeated with new task
- Evolves population of paths to a subset of network and fixes this path so that functionality can’t be lost
- Similar to progressive nets
- Progressive nets = fixed transfer learning topology
- PathNet = learn transfer learning topology
- Allows relationship between earlier and later tasks to be evolved
-
Methods
-
PathNet Architecture
- $L$ Layers with $M$ Modules
- Each module is a convolutional or linear layer followed by transfer function (i.e., ReLU)
- In each layer, outputs of layer are summed before passed into active modules of next layer
- Active: module present in path of genotype being evaluated
- Maximum of $N$ modules per layer allowed in a pathway
- Final layer is unique (unshared) for each task
-
Pathway Evolution: Serial and Parallel
- Transfer Learning Paradigm
- Binary MNIST classification tasks
- CIFAR and SVHN classification tasks
- Atari games
- Labyrinth Games
-
Results
- Binary MNIST Classification
- CIFAR and SVHN
- Atari Games
- Labyrinth Games
· research