Reproduce “BottleFit: Learning Compressed Representations in Deep Neural Networks for Effective and Efficient Split Computing” on OpenResearch

About this paper

Yoshitomo Matsubara, Davide Callegaro, Sameer Singh, Marco Levorato, Francesco Restuccia

BottleFit is a framework that learns highly compressed intermediate representations within deep neural networks for efficient split computing, achieving up to 77.1% data compression with only up to 0.6% accuracy loss on ImageNet. The method reduces end-to-end inference latency by up to 89% and mobile device power consumption by up to 49% compared to local computing, enabling robust AI deployment on resource-constrained edge devices.

Key insights

• Modifying existing DNN layers to embed bottleneck functionality, rather than adding external autoencoders, drastically reduces the head model's size and computational demands on edge devices.
• A multi-stage training strategy, pretraining the bottleneck with Generalized Head Network Distillation, is essential for maintaining high accuracy even with strong compression rates on large-scale datasets.
• Compressing intermediate DNN representations effectively can reduce data transfer by over 77% (or 93% with quantization) while outperforming JPEG-compressed input images in terms of data size and accuracy.

Summary by alphaXiv.