PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning

CVPR 2018  ·  Arun Mallya, Svetlana Lazebnik ·

This paper presents a method for adding multiple tasks to a single deep neural network while avoiding catastrophic forgetting. Inspired by network pruning techniques, we exploit redundancies in large deep networks to free up parameters that can then be employed to learn new tasks. By performing iterative pruning and network re-training, we are able to sequentially "pack" multiple tasks into a single network while ensuring minimal drop in performance and minimal storage overhead. Unlike prior work that uses proxy losses to maintain accuracy on older tasks, we always optimize for the task at hand. We perform extensive experiments on a variety of network architectures and large-scale datasets, and observe much better robustness against catastrophic forgetting than prior work. In particular, we are able to add three fine-grained classification tasks to a single ImageNet-trained VGG-16 network and achieve accuracies close to those of separately trained networks for each task. Code available at https://github.com/arunmallya/packnet

PDF Abstract CVPR 2018 PDF CVPR 2018 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Continual Learning Cifar100 (20 tasks) PackNet Average Accuracy 67.5 # 8
Continual Learning CUBS (Fine-grained 6 Tasks) PackNet Accuracy 80.41 # 5
Continual Learning Flowers (Fine-grained 6 Tasks) PackNet Accuracy 93.04 # 6
Continual Learning ImageNet (Fine-grained 6 Tasks) PackNet Accuracy 75.71 # 5
Continual Learning Sketch (Fine-grained 6 Tasks) PackNet Accuracy 76.17 # 6
Continual Learning Stanford Cars (Fine-grained 6 Tasks) PackNet Accuracy 86.11 # 6
Continual Learning Wikiart (Fine-grained 6 Tasks) PackNet Accuracy 69.40 # 6

Methods