LegoNet: Memory Footprint Reduction Through Block Weight Clustering

A team of researchers has introduced LegoNet, a groundbreaking compression method that dramatically reduces the memory footprint of neural networks without sacrificing performance. Developed by Joseph Bingham, Noah Green, and Saman Zonouz, the technique works by constructing blocks of weights from across the entire model, regardless of layer type, and then clustering these blocks. In their tests on the standard ResNet-50 model trained on ImageNet, using just 32 of these 4x4 blocks achieved a compression ratio of 64x. This means the model requires 64 times less memory, all while maintaining the original model's accuracy. The most significant advantage is that this compression is achieved post-training, requiring no retraining, fine-tuning, or access to the original dataset.

This approach stands in contrast to other compression methods like pruning or quantization, which often involve trade-offs with accuracy or require extensive computational resources to retrain. LegoNet's block-wise clustering preserves the model's functional architecture, allowing it to be deployed immediately after compression. The researchers demonstrated that even more aggressive compression is possible, finding an arrangement of 16 blocks that yields a 128x reduction in memory footprint with less than a 3% drop in accuracy. Published at the IEEE DASC 2022 conference, this work directly addresses the critical barrier of model size, which currently prevents powerful AI from running on ubiquitous embedded systems with limited cache and RAM, such as mobile phones, IoT sensors, and edge computing devices.