Continual Learning for Recurrent Neural Networks: an Empirical Evaluation

Abstract

Learning continuously during all model lifetime is fundamental to deploy machine learning solutions robust to drifts in the data distribution. Advances in Continual Learning (CL) with recurrent neural networks could pave the way to a large number of applications where incoming data is non stationary, like natural language processing and robotics. However, the existing body of work on the topic is still fragmented, with approaches which are application-specific and whose assessment is based on heterogeneous learning protocols and datasets. In this paper, we organize the literature on CL for sequential data processing by providing a categorization of the contributions and a review of the benchmarks. We propose two new benchmarks for CL with sequential data based on existing datasets, whose characteristics resemble real-world applications. We also provide a broad empirical evaluation of CL and Recurrent Neural Networks in class-incremental scenario, by testing their ability to mitigate forgetting with a number of different strategies which are not specific to sequential data processing. Our results highlight the key role played by the sequence length and the importance of a clear specification of the CL scenario.

Continual learning of longitudinal health records

“…However, as noted by Cossu et al [11], this does not capture the full breadth of potential scenarios. For example, newly encountered datasets may introduce a mix of domain shifted instances of old classes, new classes, or novel combinations of classes.…”

“…Strategy specific hyperparameters were tuned independently for each method. Hyperparameters were sampled from a range of reasonable values determined from the literature [36,11]. Where methods shared identical or analogous parameters, the search-space was also shared to ensure fair comparison (for example, regularization strength in EWC, SI, and LwF).…”

“…Related work: Cossu et al [11] present a comprehensive evaluation of methods on a set of proposed benchmarks for sequence data. We extend on this work by evaluating such methods on real-world clinical scenarios, over a broader array of model architectures (including Transformers and alternative recurrent networks).…”

“…Maltoni and Lomonaco [18] divide scenarios into multi-task, single-incremental task, and multiple-incremental tasks, along with a secondary classification for new examples containing new instances of old classes, new classes, or both. However, as discussed by [11], several of the proposed categories are unrealistic or rare. For simplicity we use the terminology of Van de Ven and Tolias [17].…”

Avalanche: an End-to-End Library for Continual Learning

“…to Robotics [29,39,41], from Reinforcement Learning [26,32] to Sequence Learning [8], among others. However, continual learning algorithms today are often designed and implemented from scratch with different assumptions, settings, and benchmarks that make them diffi-3595 cult to compare among each other or even port to slightly different contexts.…”