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.

BOWLL: A Deceptively Simple Open World Lifelong Learner

The quest to improve scalar performance numbers on predetermined benchmarks seems to be deeply engraved in deep learning. However, the real world is seldom carefully curated and applications are seldom limited to excelling on test sets. A practical system is generally required to recognize novel concepts, refrain from actively including uninformative data, and retain previously acquired knowledge throughout its lifetime. Despite these key elements being rigorously researched individually, the study of their conjunction, open world lifelong learning, is only a recent trend. To accelerate this multifaceted field's exploration, we introduce its first monolithic and much-needed baseline. Leveraging the ubiquitous use of batch normalization across deep neural networks, we propose a deceptively simple yet highly effective way to repurpose standard models for open world lifelong learning. Through extensive empirical evaluation, we highlight why our approach should serve as a future standard for models that are able to effectively maintain their knowledge, selectively focus on informative data, and accelerate future learning.

Masked Autoencoders are Efficient Continual Federated Learners

Machine learning is typically framed from a perspective of i.i.d., and more importantly, isolated data. In parts, federated learning lifts this assumption, as it sets out to solve the real-world challenge of collaboratively learning a shared model from data distributed across clients. However, motivated primarily by privacy and computational constraints, the fact that data may change, distributions drift, or even tasks advance individually on clients, is seldom taken into account. The field of continual learning addresses this separate challenge and first steps have recently been taken to leverage synergies in distributed supervised settings, in which several clients learn to solve changing classification tasks over time without forgetting previously seen ones. Motivated by these prior works, we posit that such federated continual learning should be grounded in unsupervised learning of representations that are shared across clients; in the loose spirit of how humans can indirectly leverage others' experience without exposure to a specific task. For this purpose, we demonstrate that masked autoencoders for distribution estimation are particularly amenable to this setup. Specifically, their masking strategy can be seamlessly integrated with task attention mechanisms to enable selective knowledge transfer between clients. We empirically corroborate the latter statement through several continual federated scenarios on both image and binary datasets.