Learn Continually, Generalize Rapidly: Lifelong Knowledge Accumulation for Few-shot Learning

Abstract

The ability to continuously expand knowledge over time and utilize it to rapidly generalize to new tasks is a key feature of human linguistic intelligence. Existing models that pursue rapid generalization to new tasks (e.g., fewshot learning methods), however, are mostly trained in a single shot on fixed datasets, unable to dynamically expand their knowledge; while continual learning algorithms are not specifically designed for rapid generalization. We present a new learning setup, Continual Learning of Few-Shot Learners (CLIF), to address the challenges of both learning settings in a unified setup. CLIF assumes a model learns from a sequence of diverse NLP tasks arriving sequentially, accumulating knowledge for improved generalization to new tasks, while also retaining performance on the tasks learned earlier. We examine how the generalization ability is affected in the continual learning setup, evaluate a number of continual learning algorithms, and propose a novel regularized adapter generation approach. We find that catastrophic forgetting affects generalization ability to a lesser degree than performance on seen tasks; while continual learning algorithms can still bring considerable benefit to the generalization ability 1 .

Robust internal representations for domain generalization

“…To improve upon these methods, we introduce a Continual Learning of Few‐Shot Learners (CLIF) that aims to tackle the challenges posed by CL and FSL within a unified framework using a base transformer (Jin et al. 2021). CLIF is designed to address the learning process where a model sequentially learns from a diverse range of NLP tasks.…”

“…Transformers are much larger compared to prior models such as convolutional neural networks and using weight regularization or experience replay on the full model may not lead to desired results. To improve upon these methods, we introduce a Continual Learning of Few-Shot Learners (CLIF) that aims to tackle the challenges posed by CL and FSL within a unified framework using a base transformer (Jin et al 2021). CLIF is designed to address the learning process where a model sequentially learns from a diverse range of NLP tasks.…”

ConTinTin: Continual Learning from Task Instructions

“…Systems in (Sun et al, 2020a,b) incrementally learned among five disparate NLP tasks. Jin et al (2021) further extended the size of the task stream (one benchmark has 26 tasks, the other covers 55) and studied TCL in a few-shot scenario. It is worth mentioning that all the listed work in TCL consistently transformed all tasks into question answering format (as pointed out in (McCann et al, 2018), many NLP tasks can be formulated as question answering), thus TCL in these literature was actually converted into DCL.…”

“…Similar with (Xia et al, 2021;Jin et al, 2021), our work also focuses on low-resource continual learning; in contrast, our learning problem belongs and Cohen, 1989), and never-ending learning (Carlson et al, 2010). to TCL while each task in our formulation is expressed by instructions instead of labeled examples.…”

Variational Autoencoder with Disentanglement Priors for Low-Resource Task-Specific Natural Language Generation

“…The primary challenge addressed in continual learning literature is overcoming catastrophic forgetting (French, 1999;Biesialska et al, 2020;Wu et al, 2022), Various approaches have been proposed to tackle the forgetting problem, e.g., rehearsal-based methods (Han et al, 2020;de Masson d'Autume et al, 2019;, regularization-based methods (Li et al, 2019;Huang et al, 2021), and dynamic architecture methods (Ke et al, 2021;. Continual few-shot learning is an even more challenging yet realistic setting which encourages learners the quick adaptation ability during learning (Jin et al, 2021;Yoon et al, 2020). Comparing to the numerous researches out of NLP applications (Yap et al, 2021;Yoon et al, 2020;Dong et al, 2021), continual few-shot language learning is still an under-explored area (Jin et al, 2021).…”