RESEARCH

Stochastic Optimization for Distributed and Constrained Machine Learning

Machine Learning is now ubiquitous in almost every engineering field. One of the cornerstones of obtaining a model with good accuracy is knowing how to properly optimize the training objective function. This is where stochastic optimization comes in.

Distributed optimization concerns itself with minimization of a sum of functions, where each component is located at the node of a network. Methods whcih aim to solve it generally consists of two key components: communication of information exchange with the neighboring node of the network and computations of the algorithm being used to solve the evaluating optimization related quantities like gradients.

Manifold optimization deals with problems where the constraints can be posed as Riemannian manifolds are considered. Manifold optimization has become ubiquitous in applied mathematics and more broadly, in statistics and engineering. One of the main challenges that can be overcome using the framework of Riemannian manifolds is that a problem that is non-convex in the Euclidean space may be geodesically convex when posed as manifold constraints. Another important aspect of exploiting the geometry of the problem at hand is that a large class of constrained optimization problems can be viewed as unconstrained optimization problems on manifold.

Papers

Wireless Machine Learning

In modern Iot based applications, wireless ML has paved the way to train on unlimited data being generated on edge devices. The potential to harness this data in a safe and privacy preserving manner is crucial fr the success of many AI systems.

Federated learning (FL) or collaborative learning is a machine learning technique that trains models across multiple decentralized edge devices (like cellular devices) or servers holding local data samples, without exchanging them. It thus alleviates the need to exchange information among the participating entities thereby allowing to address critical issues such as data privacy, data security, data access rights and access to heterogeneous data. Its applications are spread over a number of industries including defense, telecommunications, IoT, and pharmaceutics.

Neural network based models have, and will for the foreseeable future, achieved state-of-the-art accuracies in computer vision, natural language processing and reinforcement learning tasks. A large model size can potentially require infeasible amounts of data to be transmitted between the server and clients for training in FL. This makes it imperative to manage communication overhead in a proper manner which reduces the burden on the client.

Papers

Reinforcement Learning and Bandit Optimization

Reinforcement learning is is one of three basic learning frameworks in machine learning and considered to be the most versatile and generalizing. However, it also more complex and potentially quite difficult to solve. Originally inspired from the theory of operant conditioning, it can roughly be considered to be its mathematical analog where the learning uses positive rewards to steer the agent towards optimal behaviour and negative rewards to modify non-optimal behaviour. The problem is usually in the form of a Markov decision proves and many algorithms have been proposed within the Deep learning framework to solve it.

Bandit optimization can broadly be considered a simplified version of reinforcement learning. It aims to demonstrate the exploration/exploitation trade-off capabilities of algorithm or heuristic. The multi armed bandit problem considers a fixed set of resources that have to be allocated between competing choices with associated rewards. Since the rewards are modelled as a probability distribution, the perfect policy which would maximize the expected rewards is not known aprioiri.

Papers