One interesting pre-print just caught my eye. We often think of DL in computer vision as a pattern recognition methodology. Authors here refevers the problem and try to identify patterns most difficult to detect. On top of that they use genetic algorithm with DL.
How can you use genetic algorithm for deep learning?
Genetic algorithms (GA) is yet another great example how biology can inspire computer science concepts. On a high level it can be though of optimization or a search algorithm.
This blogs gives a great breakdown and a simple implementation of how GA works:
Similar to gradient descent or ascent algorithms widely used in deep learning to train the deep neural networks, GA is heuristic and is based on a Monte Carlo (i.e. random) approach. So in principle there could be a way of training based on GA. In the paper above it seems however, GA was used to narrow down the dimensions of the target (Y): “Targets in our main experiments were constructed using nine dimensions (three for each of the two colours and three for texture), resulting in a parameter space containing a total of 6.18 × 10^17 possible patterns. Since our parameter space was so large, we were unable to exhaustively or randomly select targets with sufficient diversity. Therefore, we implemented a genetic algorithm (GA) to optimise the colour and texture parameters, based on participants’ responses trial by trial, for hardest or easiest to see stimuli”.
This is not unique, in a paper below they seem to use GA for deep reinforcement learning: https://arxiv.org/pdf/1905.04100.pdf
Here’s an example of how GA can be used for hyperparameters search for a neural network: https://blog.coast.ai/lets-evolve-a-neural-network-with-a-genetic-algorithm-code-included-8809bece164