Towards AI that is forever learning efficiently

点击翻译按钮获取中文摘要

This project falls within the EPSRC Statistics and Applied Probability research area. It will look at researching techniques for AI that allow a model to continually learn on different datasets which includes adapting to perform on current training data and not 'forgetting' how to perform well on previous data. If all the data a neural network has trained on up to a certain point in time is not independently identically distributed, we can run into a problem as the gradient updates used to train the network aren't on average changing the neural network to perform optimally on all the data. This problem arises in many areas including when we try to train a network on a sequence of tasks containing different data, in this scenario the network will optimize on the current data and loose performance on previous data which has been named catastrophic forgetting. The area of research named 'continual learning' looks at how to train models in such a scenario and one of the significant issues to overcome for this research area is mitigating catastrophic forgetting. Research in this area could lead to being able to continually update models with new data whilst its deployed, being able to teach models by interacting with them and more easily train on very large datasets. It may also look at researching techniques that allow for the efficient use of data. One way to more efficiently use data is being able to work with more data, ie. data with less requirements. A classic example of this is working with unlabelled data to be able to get good performance on an image classification task where we only have a small number of labelled images available. Such techniques are good at learning ways to represent data that can be useful for a myriad of downstream applications. The area that focuses on trying to train models that are good at extracting information from data in an unsupervised way that is useful for downstream applications is called representation learning. Advantages of (unsupervised) representation models are that they allow for training models on much larger amounts of data and therefore solve problems that due to data restrictions were previously unsolvable. The project may also look at the intersection of these two research areas, ie. continual learning from data with less requirements such as without the need for labels. This is an area not yet widely explored by researchers and could have a large impact on the kind of problems we can apply deep learning to. It could allow for models that whilst deployed can train continually on raw incoming data, forever improving and learning and without the need for lengthy data preparation. Aims and objectives of this project include: Develop new methods to facilitate learning in the continual setting. Develop new methods for the efficient use of data for representation learning which includes representation learning on data with less requirements (eg. images without labels). Develop new methods that facilitate learning in the continual setting that are efficient at using data. Apply these methods to problems to validate their performance empirically.