<ul><li>Increasingly complex neural network architectures have achieved phenomenal performance but require massive computational resources and consume substantial amounts of electricity, raising environmental concerns.</li><li>Research on large pre-trained models shows redundancies exist. Previous focus was on model compression for performance rather than electricity consumption impact.</li><li>Study quantifies energy usage of uncompressed and compressed large pre-trained models to reduce electricity consumption.</li><li>Compression techniques like steganographic capacity reduction show significant benefits in reducing energy usage, while pruning and low-rank factorization do not offer significant improvements.</li></ul>

Energy Considerations for Large Pretrained Neural Networks

Discover more