ConVAEr: Convolutional Variational AutoEncodeRs for incremental similarity learningDownload PDF

29 Sept 2021 (modified: 13 Feb 2023)ICLR 2022 Conference Withdrawn SubmissionReaders: Everyone
Keywords: catastrophic forgetting, incremental similarity learning
Abstract: Due to catastrophic forgetting, incremental similarity learning in neural networks remains an open challenge. Previous work has shown that keeping image exemplars during incremental similarity learning is effective for preserving base knowledge (past learnt features and embeddings). It is also generally accepted that the output layers learn more task-specific feature embeddings during the later training stages compared to the input layers’ general features earlier on. Building on these insights, we start by freezing the input layers of a neural network. We then investigate the viability of generating “embedding” exemplars from a VAE that can protect base knowledge in the intermediate to output layers of the neural networks. These generated exemplars replace the necessity for retaining images from previously learned classes. We experimented with three metric learning loss functions on the CUB-200 and CARS-196 in an incremental similarity learning setup. We train different VAEs to generate exemplars from the intermediate convolution layers and linear output layers. We use these generated exemplars to rep-resent base knowledge. We compared our work to a previous technique that stores image exemplars. The comparison is done for base knowledge, new knowledge and average knowledge preservation as metrics. The results show that generating exemplars from the linear and convolutional layers retained the highest ratio of base knowledge. We note that using embeddings from the linear layers leads to better performance on new knowledge than convolutional embeddings. Overall our methods yield better average knowledge performance across all experiments. These results support the view that for incremental similarity learning to overcome catastrophic forgetting, emphasis can be placed on learning embedding exemplars for intermediate to output layers. Further, we note that most incremental similarity learning for new classes depends on the linear layers rather than the convolutions. Further investigation into the relationship between transfer learning and similarity learning and the protection of intermediate layer embedding space for catastrophic forgetting is required.
One-sentence Summary: Investigating into incremental similarity learning.
5 Replies

Loading