Description
The growing energy requirements of modern machine learning are limiting the ability to run models on large compute clusters owned by a handful of companies and preventing interesting end-user and edge applications. Therefore, it is becoming increasingly important to optimize not only the task performance, but also the energy efficiency of these models. Recurrent neural networks are well suited for energy efficient edge computing. A recently developed model [1] has demonstrated the potential of activity sparsity in recurrent networks. Furthermore, it has been shown that recurrent networks can operate in sparse parameter space if the sparsity is adapted online [2]. In this project, we will combine activity and online parameter sparsity, investigate these two types of sparsity, and identify optimal trade-offs between performance and energy efficiency.
Keywords: Efficient Machine Learning, Recurrent Neural Networks, GRU, Sparsity
Prerequisites
- Experience with the Python programming language, ideally previous experience with machine learning in Python, e.g. PyTorch, Jax, etc.
- Completed courses, internships, or projects related to artificial intelligence or machine learning.
References
[1] Anand Subramoney, Khaleelulla Khan Nazeer, Mark Schöne, Christian Mayr, David Kappel. Efficient recurrent architectures through activity sparsity and sparse back-propagation through time. ICLR 2023. https://arxiv.org/pdf/2206.06178.pdf
[2] Guillaume Bellec, David Kappel, Wolfgang Maass, Robert Legenstein. Deep rewiring: Training very sparse deep networks. ICLR 2017. https://arxiv.org/pdf/1711.05136