Continual Metalearning

点击翻译按钮获取中文摘要

Context of the research and potential impactMeta learning, also referred to as "learning to learn", studies techniques and machine learning models able to adapt and generalise to new tasks and environments quickly, without having to go through a long training process every time. Continual learning focuses on the development of methods that can deal with a continuous stream of data.Advancements in continual meta-learning would hugely influence the application of modern machine learning models in industry, economy and society, where data is often received in a continuous stream. For machine learning models to be deployed in the real-world, they need to be able to detect model/data mismatch, concept drift, anomalies and out of sample quantification. The models require metrics to quantify when, how and to what extent performance is degrading, to determine when their parameters need to be updated.Our goal is to create models that can distinguish between meaningful and not meaningful changes in the data, and adapt to the first. Adaptive models need a dynamical architecture and they are appropriate when working with sequential streams of data. Detecting and quickly adapting to meaningful changes is critical in the real-world, for models' predictions to stay reliable.Aims and ObjectivesThe aim of this project is to advance the field of continual meta-learning, to enable a meaningful integration of machine learning models in industry settings and to deal with real-world data. The first task of the project is to provide an overlook of existing techniques. In particular, we will focus on Neural Processes and on extending them to the continual learning domain. Neural Processes are hybrid models, that like Gaussian Processes define distributions over functions, and have computationally efficient training like Neural Networks, with the ability to adapt their priors to data [Garnelo et al., 2018b,a]. Subsequent research can be done on the inclusion of meaningful inductive biases for and from multiple domains, and on the transfer learning abilities of ensembles/reservoirs of Neural Processes.During the DPhil we will explore the development of dynamical models with the ability of detecting model-data mismatch, concept drift, anomalies and out of sample quantification. An approach would be to recognise which past training information are now obsolete and remove them, with the model updating its parameters accordingly.A good knowledge of existing performance metrics is needed, to understand what may still be missing and eventually develop other metrics useful to update the model and compensate for anomalous and drifting data.A long-term goal of this research is to make continual meta-learning models applicable on real-world data, such as time series data or spatio temporal data to tackle environmental problems.Novelty of the research methodologyNeural Processes are a relatively new architecture, initially proposed by Garnelo et al. [2018b,a]. Jha et al. [2022] presents a summary of the most recent applications of Neural Processes. To our knowledge, the application of Neural Processes to continual learning settings, or using an ensemble of Neural Pro-cesses for transfer learning, have not been yet explored. Moreover, as their development is fairly recent, work needs to be done to assess their applicability in real-world settings.As new dynamical and adaptive models are developed, it is likely that we will need additional metrics to assess their features and overall performance.Alignment to EPSRC's strategies and research areasThe research aligns with the EPSRC research area of Artificial Intelligence technologies, as the research is carried out in a machine learning group in partnership with the spin-out Mind Foundry, that applies cutting edge machine learning to deliver solutions to business and industry.