Some novel DL workloads in 2021
Table of Contents
Most Deep Learning Systems/Compilers in 2021 are still trying to optimize workloads like ResNet & VGG, though these models are really useful and we can publish papers on decent conference by proposing new compiler techniques for them (e.g. TASO/PET). I'm not so enthusiastic about competing on this track, Graph Neural Networks brings structured data into Deep Learning, which is good, but unfortunately the application of large scale GNNs are mostly limited to recommendation systems. I always expect my work to have broader positive social impact and apparently recommendation system can't.
I started surveying the evolving Deep Learning workloads this semester and try to find the "future" of Machine Learning Systems. Currently I feel like AlphaFold, Diffusion Model and Neural Radiance Field as three models that worth investigating into.
AlphaFold
TODO Multiple Sequence Alignment
TODO AlphaFold
TODO SE(3)-Transformer
Tensor Field Networks
\[ \textbf{f}(\textbf{x}) = \sum_{j=1}^{N} \textbf{f}_j \delta(\textbf{x} - \textbf{x}_j)\]
NVIDIA's optimization of SE(3)-Transformer
NVIDIA released an acceleration of SE(3)-Transformer.
RoseTTAFold
Diffusion Model
Lillian Weng's blog post is a good start point to know about diffusion models.