cute is an evolutionary algorithm to evolve convolutional neural networks for image classification tasks, and eventually other tasks like object detection. cute stands for CNNs through asynchronoUs Training and Evolution
A list of things that need to be done:
- master and worker process logic
- design and implementation of the cnn genome
- implementation of island speciation strategy
- mutations and crossover operations (dependent on [2])
- handling program arguments
- dataset processing - currently only support MNIST and a debug dataset which is just a truncated MNIST
- intelligently handle multiple GPUS, train multiple networks on the same gpu if possible