Delegation of computation for machine learning

点击翻译按钮获取中文摘要

This research falls into the areas of machine learning and quantum computing. Machine learning is ubiquitous in our daily lives, from Spotify recommendations to estimating how long your commute is going to take today. Delegating machine learning tasks to a more powerful third-party in the cloud (for example) offers the potential to make more difficult learning problems tractable, which is a very attractive prospect. At present, however, we blindly trust that the third-party in the cloud is doing a good job. This is fine for less sensitive tasks such as song recommendations, but is much less desirable in more sensitive situations. This research will aim to produce protocols that remove this need for blind trust,and instead allow us to verify that the the third-party has indeed done a sufficiently good job.We will aim to develop a theory of how to delegate machine learning, i.e. asking a powerful computer in the cloud to carry out a machine learning task for you. This will include the development of protocols which allow machine learning tasks to be delegated in a way which doesn't require trust, allowing us to verify that what the powerful computer in the cloud is telling us is in fact true. We will aim to produce protocols for this delegation which are transparent, privacy-preserving and post-quantum secure.This work is highly interdisciplinary, comprising the areas of cryptographic proofs and machine learning. This combination is a new direction of research. We will use deep technical tools from the theory of cryptographic proofs to approach the problem of delegating machine learning via novel methodologies.Delegating machine learning tasks to a more powerful computer is an attractive prospect. However, ideally, we would prefer not to have to trust that said powerful computer is doing a good job, and would rather have some way of verifying this fact for ourselves. The aim of this work is to produce protocols which remove this need for trust, and facilitate the delegation of machine learning tasks in a verifiable, transparent, privacy-preserving way. These protocols could immediately be applied to any delegated machine learning task in which it is important that the output is sufficiently high quality. Both sides stand to gain from this, as the side doing the verifying has a guarantee that what they are receiving is up to scratch, while the side doing the learning will now have the means to convince anyone that they are doing a good job.This work falls within the areas of machine learning and quantum information. It looks at these two areas with respect to the area of delegation of learning, which are priority areas for EPSRC, and it strives to make a significant industrial and societal impact.External Partner; StarkWare