Few-Shot Lifelong Learning

Abstract

Many real-world classification problems often have classes with very few labeled training samples. Moreover, all possible classes may not be initially available for training, and may be given incrementally. Deep learning models need to deal with this two-fold problem in order to perform well in real-life situations. In this paper, we propose a novel Few-Shot Lifelong Learning (FSLL) method that enables deep learning models to perform lifelong/continual learning on few-shot data. Our method selects very few parameters from the model for training every new set of classes instead of training the full model. This helps in preventing overfitting. We choose the few parameters from the model in such a way that only the currently unimportant parameters get selected. By keeping the important parameters in the model intact, our approach minimizes catastrophic forgetting. Furthermore, we minimize the cosine similarity between the new and the old class prototypes in order to maximize their separation, thereby improving the classification performance. We also show that integrating our method with self-supervision improves the model performance significantly. We experimentally show that our method significantly outperforms existing methods on the miniImageNet, CIFAR-100, and CUB-200 datasets. Specifically, we outperform the state-of-the-art method by an absolute margin of 19.27% for the CUB dataset.

Multimodal Parameter-Efficient Few-Shot Class Incremental Learning

“…Prototype modeling was also used to assign prototypes in the embedding space to reserve it for future incoming classes [62], or to use the average of new class embedding representations as a class prototype to replace classifiers [63]. Different methods addressed the problem by synthesizing features into a mixture of sub-spaces for incremental classes by using a VAE [7], or by adapting general deep learning architectures to enable a few parameters to be updated for every new set of novel incoming classes [30]. More recent approaches tried to combine features emerging from supervised and self-supervised models for boosting classifiers [1], or to calibrate distributions to avoid forgetting by retrieving distributions for old classes while estimating distributions for new classes [26].…”

“…We also introduce prompt regularization to improve performance and prevent forgetting. Our experimental results demonstrate that CPE-CLIP significantly improves FSCIL performance compared to state-of-the-art proposals while also drastically reducing the number of learnable parameters and training costs.Recent research has focused on solving these problems through various approaches, such as meta-learning [57, 34], regularization techniques [30], or knowledge distillation [38,6,62]. These methods have shown promising results in achieving incremental learning over time with a limited amount of data available.…”

“…Recent research has focused on solving these problems through various approaches, such as meta-learning [57, 34], regularization techniques [30], or knowledge distillation [38,6,62]. These methods have shown promising results in achieving incremental learning over time with a limited amount of data available.…”

Federated Learning via Attentive Margin of Semantic Feature Representations

“…Recently, OCL for embedded devices [4], [8], [23], [24] further restricts the scope to parameterefficient updates with unitary batch size due to low-resource devices. In parallel, few-shot CL (FS-CL) [25], [26], [27], [28], [29] and FS-OCL [10] train a model to recognize new classes based on few labelled data only, via fast and efficient model updates with minimal human effort. Along all these definitions, three main scenarios have been considered [30]: 1) class-incremental (CI) where new classes are introduced to models over time; 2) domain-incremental (DI) where the same problem is learned in different contexts (e.g., with a fixed set of classes); and 3) task-incremental (TI), where new distinct tasks are presented to models.…”

S3C: Self-Supervised Stochastic Classifiers for Few-Shot Class-Incremental Learning

“…It can also arrest the catastrophic forgetting of the previously learnt classes to a certain extent. As is common in most FSCIL approaches [44,25], we propose to freeze the feature extractor and learn only the SC at each incremental step. In order to compute features from the base classes which generalize to unseen classes, inspired by recent works [22,47], we use self-supervision along with SC giving our final S3C framework.…”