Research on continual learning has led to a variety of approaches to mitigating catastrophic forgetting in feed-forward classification networks. Until now surprisingly little attention has been focused on continual learning of recurrent models applied to problems like image captioning. In this paper we take a systematic look at continual learning of LSTM-based models for image captioning. We propose an attention-based approach that explicitly accommodates the transient nature of vocabularies in continual image captioning tasks -- i.e. that task vocabularies are not disjoint. We call our method Recurrent Attention to Transient Tasks (RATT), and also show how to adapt continual learning approaches based on weight regularization and knowledge distillation to recurrent continual learning problems. We apply our approaches to incremental image captioning problem on two new continual learning benchmarks we define using the MS-COCO and Flickr30 datasets. Our results demonstrate that RATT is able to sequentially learn five captioning tasks while incurring no forgetting of previously learned ones.

We exploit categorical image annotations available in many captioning datasets. The influence of the people category is clearly visible. Figure 2: RA TT ablation on the MS-COCO validation set using different attention masks. Evaluation is the same as MS-COCO (figure 4). In figures 6 and 7, we give a comparison of performance for all considered approaches on the MS-COCO validation set. These learning curves and heatmaps allow us to appreciate the ability of RA TT to remember old tasks.

This is caused by the recurrent connections which amplify each small change in the weights.

Strengths: The paper is one of the first to study continual learning in recurrent settings and shows promising performance on the image captioning task. It proposes RATT, a novel approach for recurrent continual learning based on attentional masking, inspired by the previous HAT method. In its proposed method, three masks (a_x, a_h, and a_s) to embedding, hidden state, and vocabulary are introduced, and in its ablation study, the paper shows that all these three components are helpful to the final continual learning performance. In addition to the proposed novel approach, the paper also explores adapting weight regularization and knowledge distillation-based approaches to the recurrent continual learning problem. In its experiments, the paper shows strong results, largely outperforming simple baselines (such as fine-tuning) and previous regularization or distillation-based approaches (EWC and LwF).

The paper received two accept reviews and one borderline reject [R1]. The main concern of R1 is the paper relies on simple/not the most recent approaches for both captioning and continual learning. The other reviewers and I agree to that but believe that for one of the first papers in continual learning for captioning that this is reasonable, even if it is not optimal. R1 did not respond after the rebuttal. The reviewers appreciate the the paper's contributions, including 1) First paper in continual learning in image captioning.

Research on continual learning has led to a variety of approaches to mitigating catastrophic forgetting in feed-forward classification networks. Until now surprisingly little attention has been focused on continual learning of recurrent models applied to problems like image captioning. In this paper we take a systematic look at continual learning of LSTM-based models for image captioning. We propose an attention-based approach that explicitly accommodates the transient nature of vocabularies in continual image captioning tasks -- i.e. that task vocabularies are not disjoint. We call our method Recurrent Attention to Transient Tasks (RATT), and also show how to adapt continual learning approaches based on weight regularization and knowledge distillation to recurrent continual learning problems.

Research on continual learning has led to a variety of approaches to mitigating catastrophic forgetting in feed-forward classification networks. Until now surprisingly little attention has been focused on continual learning of recurrent models applied to problems like image captioning. In this paper we take a systematic look at continual learning of LSTM-based models for image captioning. We propose an attention-based approach that explicitly accommodates the transient nature of vocabularies in continual image captioning tasks -- i.e. that task vocabularies are not disjoint. We call our method Recurrent Attention to Transient Tasks (RATT), and also show how to adapt continual learning approaches based on weight regularization and knowledge distillation to recurrent continual learning problems.

Transformers have shown dominant performance across a range of domains including language and vision. However, their computational cost grows quadratically with the sequence length, making their usage prohibitive for resource-constrained applications. To counter this, our approach is to divide the whole sequence into segments and apply attention to the individual segments. We propose a segmented recurrent transformer (SRformer) that combines segmented (local) attention with recurrent attention. The loss caused by reducing the attention window length is compensated by aggregating information across segments with recurrent attention. SRformer leverages Recurrent Accumulate-and-Fire (RAF) neurons' inherent memory to update the cumulative product of keys and values. The segmented attention and lightweight RAF neurons ensure the efficiency of the proposed transformer. Such an approach leads to models with sequential processing capability at a lower computation/memory cost. We apply the proposed method to T5 and BART transformers. The modified models are tested on summarization datasets including CNN-dailymail, XSUM, ArXiv, and MediaSUM. Notably, using segmented inputs of varied sizes, the proposed model achieves $6-22\%$ higher ROUGE1 scores than a segmented transformer and outperforms other recurrent transformer approaches. Furthermore, compared to full attention, the proposed model reduces the computational complexity of cross attention by around $40\%$.