Learning to Predict Gradients for Semi-Supervised Continual Learning

TL;DR

The paper tackles learning new visual concepts without forgetting by introducing semi-supervised continual learning (SSCL) that leverages unlabeled data through a gradient learner. The core idea is to train a gradient learner on labeled data to predict pseudo-gradients for unlabeled samples, enabling backpropagation without ground-truth labels and without assuming known unlabeled-class mappings. Empirical results across CL, adversarial CL, and SSL tasks show improved average accuracy and backward transfer, with evidence that unlabeled data can enhance generalization when used judiciously via gradient prediction. The approach is lightweight, generalizable, and extends to SSL, offering a practical pathway to incorporate unlabeled data into continual learning systems.

Abstract

A key challenge for machine intelligence is to learn new visual concepts without forgetting the previously acquired knowledge. Continual learning is aimed towards addressing this challenge. However, there is a gap between existing supervised continual learning and human-like intelligence, where human is able to learn from both labeled and unlabeled data. How unlabeled data affects learning and catastrophic forgetting in the continual learning process remains unknown. To explore these issues, we formulate a new semi-supervised continual learning method, which can be generically applied to existing continual learning models. Specifically, a novel gradient learner learns from labeled data to predict gradients on unlabeled data. Hence, the unlabeled data could fit into the supervised continual learning method. Different from conventional semi-supervised settings, we do not hypothesize that the underlying classes, which are associated to the unlabeled data, are known to the learning process. In other words, the unlabeled data could be very distinct from the labeled data. We evaluate the proposed method on mainstream continual learning, adversarial continual learning, and semi-supervised learning tasks. The proposed method achieves state-of-the-art performance on classification accuracy and backward transfer in the continual learning setting while achieving desired performance on classification accuracy in the semi-supervised learning setting. This implies that the unlabeled images can enhance the generalizability of continual learning models on the predictive ability on unseen data and significantly alleviate catastrophic forgetting. The code is available at \url{https://github.com/luoyan407/grad_prediction.git}.

Figures (12)

• Figure 1: Conceptual comparison between the challenge in the novel semi-supervised continual learning (SSCL) problem and the one in the semi-supervised learning (SSL) problem. The key difference is that the underlying classes w.r.t. the unlabeled data could be unknown in the SSCL problem, while the ones in SSL are assumed to be from the known classes. To suitably adapt to the continual learning paradigm, we do not impose such a constraint in the novel SSCL problem. Instead, the underlying labels of unlabeled data can be from either known classes or unknown classes. The faded-out samples in the task 2 indicate that the samples in the task 1 are not available in the task 2 according to the protocol.
• Figure 2: Problem of semi-supervised continual learning. Conventional supervised continual learning requires labels to compute gradients for model update (see blue flows). In contrast, this work proposes to predict gradients so that the unlabeled images can be incorporated in the continual learning paradigm for better generalizability (see red flows).
• Figure 3: Overview of the proposed gradient learning and gradient prediction process with the gradient learner $h(\cdot;\omega)$. The backbone network is shared between the two processes.
• Figure 4: Comparison between the predicted gradient $\bar{g}$ and the gradients $\frac{\partial \ell(z,\hat{y})}{\partial \theta}$ generated with pseudo labels $\hat{y}$. Assume the proposed gradient learner is trained with the samples in categories cat and dog, given an unlabeled image $\tilde{x}$, the proposed gradient learner would take all learned class-specific knowledge (i.e. w.r.t categories cat and dog) into account, instead of taking one category (i.e. cat or dog) into account in pseudo labeling methods.
• Figure 5: Geometric interpretation of supervised learning (top) and semi-supervised learning (bottom). Through leveraging the semantics of unlabeled images, the generalizability of models is expected to be improved. Experimental results in Table \ref{['tbl:conn_rota']}--\ref{['tbl:acl_miniImageNet']} validate this finding.