Deep Learning in Dynamic, Constrained Systems
Modern machine learning methods are poorly targeted at many industry-relevant settings.
The simple objective functions they attempt to optimize make some very heavy assumptions:
- Computation, storage and time factors are not relevant;
- There is a single problem to be solved by a single monolithic deep learning method, not a distributed set of interrelated problems, or online, changing problems;
- The system is the same between training and deployment. Sensors are fixed, environments are fixed; users are fixed, bandwidth is fixed; computational availability is limitless, the training data was fully representative;
- The system is stationary at test time – for the most part the need for compute-constrained continuous learning is ignored;
- All parts of the system are always on.
The aim of this research is to develop and test a framework for distributed, computation-constrained non-stationary deep learning. Two potential tasks within this space are:
- Large-scale compression of deep neural networks given memory, computation and communication constraints.
We can consider compression of networks at a given performance level on standard neural network benchmarks. By presenting Pareto frontiers for each compression method (showing payoffs of computational costs and performance), we can score each method against the benchmark performance and costs.
- Lightweight online learning or online reinforcement learning: learning to learn efficiently.
In online learning setting we need to be able to learn well from small number of examples. At the same time we need to learn with minimal computational cost: we can not afford a long training run for each example. We also need to generalise well. We will sent up a competition form of this scenario, and through recent improved understanding of neural network learning works, we will develop efficient adaptive methods for meta-learning in this context.