Artificial intelligence (AI) models like those for medical images or voice recognition use complex data structures called tensors. Tensors are like matrices but with more than two dimensions, making them hard to work with. These models need a lot of computation, using a lot of energy.
MIT researchers have devised a system to make AI models work better by using two kinds of data redundancy: sparsity and symmetry. Sparsity means many values in the data are zero, so you only need to work with the non-zero parts. Symmetry means parts of the data mirror each other, so you only need to compute half.
Before, using these redundancies was hard and you could only use one type at a time. But the MIT system, called SySTeC, lets you use both at once. This can make computations 30 times faster in some cases.
Automatic generation of complex code
SySTeC is easy to use, so even people new to deep learning can improve their AI algorithms. The system improves the generation of complex code.
SySTeC starts by figuring out how to use symmetry. If a tensor is symmetric, SySTeC only works on half of it. If steps in between are symmetric, it skips extra work. Then, it deals with sparsity by only keeping non-zero data. In the end, SySTeC generates ready-to-use code.
“In this way, we get the benefits of both optimizations," says researcher Willow Ahrens in an MIT press release. "And the interesting thing about symmetry is, as your tensor has more dimensions, you can get even more savings on computation.”
The MIT researchers have described SySTeC in a preprint published in arXiv and will present it at the International Symposium on Code Generation and Optimization.
The MIT researchers argue that SySTeC could make AI models not just faster with less memory and bandwidth demands.