Click Here to Join via Zoom
Meeting # 942 7359 5552

Zoom Meeting # 942 7359 5552

Abstract:

Deep neural networks have drawn much attention due to their success in various vision tasks, including classification. Class Incremental Leaning is a paradigm where instances from new object classes are added sequentially. Traditional methods are faced with catastrophic forgetting, where the updated model forgets the old classes and focuses only on the new classes. In this work, we introduce a framework called incremental PPCA for class incremental learning. It uses a self-supervised pre-trained feature extractor to obtain meaningful features and trains Probabilistic PCA models on the extracted features for each class separately. The Mahalanobis distance is used to obtain the classification result, and an equivalent equation is derived to make the approach computationally affordable. Experiments on standard and large datasets show that the proposed approach outperforms existing state-of-the-art incremental learning methods by a large margin. The fact that the model is trained on each class separately makes it applicable to training on very large datasets such as the whole ImageNet with more than 10,000 classes. To better handle so many classes, we take inspiration from our understanding of the human hierarchical cognition models and propose a framework called Hierarchical PPCA for image classification. The framework uses probabilistic PCA models as basic classification units and groups the image classes into a smaller number of super-classes. During classification, Hierarchical PPCA assigns a sample to a small number of most likely super-classes, and restricts the image classification to the image classes corresponding to these super-classes. Experiments on ImageNet indicate the hierarchical classifier can achieve a 4-16 times speedup compared to a standard classifier without any loss in accuracy.

Speaker's Bio

Attachments:

File	Description	File size
Adi16.jpg	Headshot	55 kB