Core Tokensets for Data-efficient Sequential Training of Transformers

Deep networks are frequently tuned to novel tasks and continue learning from ongoing data streams. Such sequential training requires consolidation of new and past information, a challenge predominantly addressed by retaining the most important data points - formally known as coresets. Traditionally, these coresets consist of entire samples, such as images or sentences. However, recent transformer architectures operate on tokens, leading to the famous assertion that an image is worth 16x16 words. Intuitively, not all of these tokens are equally informative or memorable. Going beyond coresets, we thus propose to construct a deeper-level data summary on the level of tokens. Our respectively named core tokensets both select the most informative data points and leverage feature attribution to store only their most relevant features. We demonstrate that core tokensets yield significant performance retention in incremental image classification, open-ended visual question answering, and continual image captioning with significantly reduced memory. In fact, we empirically find that a core tokenset of 1% of the data performs comparably to at least a twice as large and up to 10 times larger coreset. Deep learning models are rarely deployed in static environments and instead need to be continuously adjusted to their ever-changing surroundings. As distribution shifts occur over time, we require models to retain their performance on previous settings while accommodating new examples (Hadsell et al., 2020; Kudithipudi et al., 2022; Mundt et al., 2023). Naively, re-training a model from scratch for every change in the environment quickly becomes unfeasible for complex architectures with billions of parameters (Touvron et al., 2023; Brown et al., 2020). Consequently, research has focused on identifying small, representative subsets of the training data (Lopez-Paz & Ranzato, 2017), which can later be replayed continuously to avoid (catastrophic) forgetting (McCloskey & Cohen, 1989) of previously acquired information (Graham et al., 2021; Hassani et al., 2021; Xu et al., 2021).