Project: Learn like humans: curriculum learning for continual learning
Description
In incremental learning, the learner is presented with a sequence of t learning tasks. These tasks are typically sampled randomly, provided in an arbitrary order, which does not align with the natural learning progression observed in lifelong human learners. In human learning, we typically start with the fundamentals of a topic and gradually increase the level of difficulty.
Consider the above figure. Our ultimate objective is to learn about dolphins, and we are presented with two possible learning sequences. The first sequence combines animals and vehicles, while the second sequence starts with animals and then progresses to sea animals before focusing on dolphins. Among these options, we argue that following sequence 2, which incorporates a more transferable progression of knowledge relevant to our target learning goal, is more suitable, highlighting the need for automatic design of learning sequences.
Automated machine learning (AutoML) can optimize task sequence design via Curriculum Learning. In curriculum learning, the goal is to find the optimal order of the data that will improve the performance on the downstream target task (i.e. dolphin classification). In case of incremental learning, curriculum learning can be repurposed to identify the order as well as the structure of the tasks, instead of the training data itself.
In their pioneering work, Bell and Lawrence first show that the order in which incremental learning tasks are presented to the learner highly influences the results. And then, they propose an automatic technique, that given the space of learning tasks, outputs the optimal order which will yield the highest incremental learning accuracy and reduce forgetting. However, the contribution is mostly executed on simplistic, synthetic datasets, and only considers the order and not structure, leaving room for real-life experiments combining both.
In this thesis, you'll be the first to explore how AutoML can be used to build systems that choose themselves which tasks to solve first in order to solve future problems better.
Reading:
Khan, F., Mutlu, B., and Zhu, J. (2011). How do humans teach: On curriculum learning and teaching 163 dimension. Advances in neural information processing systems, 24.
Soviany, P., Ionescu, R. T., Rota, P., and Sebe, N. (2022). Curriculum learning: A survey. International 189 Journal of Computer Vision, 130(6):1526–1565.
Bell, S. J. and Lawrence, N. D. (2022). The effect of task ordering in continual learning. arXiv preprint.
Details
- Supervisor
-
Joaquin Vanschoren
- Interested?
- Get in contact