PRANC: Pseudo RAndom Networks for Compacting deep models

Parsa Nooralinejad, Ali Abbasi, Soheil Kolouri, Hamed Pirsiavash

Keywords: Computer Vision, Compacting Deep Models

TL;DR: A method to compress networks for sharing knowledge and storage

Abstract: Compacting deep models has various applications where the communication and/or storage is expensive including multi-agent learning. We introduce a simple yet effective framework for compacting neural networks. In short, we train our network to be a linear combination of many pseudo-randomly generated frozen models. Then, one can reconstruct the model by communicating or storing the single `seed' scalar used to generate the pseudo-random `basis' networks along with the learned linear mixture coefficients. Our method, denoted as PRANC, learns almost $100\times$ fewer parameters than a deep model and still performs reasonably well on several datasets and architectures. PRANC enables 1) efficient communication of models between agents, 2) efficient model storage, and 3) memory-efficient inference by generating layer-wise weights on the fly. We test PRANC on CIFAR-10, CIFAR-100, tinyImageNet, and ImageNet-100 with various architectures like AlexNet, LeNet, ResNet18, ResNet20, and ResNet56 and demonstrate a massive reduction in the number of parameters while providing satisfactory performance on these benchmark datasets.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2206.08464/code)