July 15, 2022
DECORE: Deep Compression with Reinforcement Learning
As deep learning has become an increasingly popular and powerful methodology for modern pattern recognition systems, the models for the most complex tasks have grown to contain millions or billions of parameters, making them difficult to employ on mobile devices for real world applications due to constraints on memory size or latency requirements.
We present DECORE: Deep Compression with Reinforcement Learning, a reinforcement learning-based approach to entirely automate the network compression process. DECORE solves for the importance of each channel/neuron in the network for the classification output of the model, and removes the channels with low importance to achieve compression. DECORE can be also be used to search network architectures for production scenarios with a fixed memory or floating point operations per second (FLOPS) budget.
We welcome conversations with companies interested in leveraging this technology in their products to deliver more powerful on-device AI processing and user experiences.
We also welcome members of the research community to evaluate this research, and share feedback. We’re expanding our team and always thrilled to meet talented researchers and engineers interested in working with us on challenging AI problems, using state-of-the-art techniques such as transformers, multimodal learning, graph deep learning, and self-supervised learning among others. Just drop us an email at email@example.com. Thank you!
Watch Manoj Alwani’s presentation:
Access the full paper at https://arxiv.org/abs/2106.06091