What I find fascinating about RFDiffusion is that it puts together two very powerful yet distinct deep learning architectures: Diffusion models and Graph Neural Networks. I wrote about this here: https://www.assemblyai.com/blog/ai-trends-graph-neural-netwo...
Hi, I mentioned "distinct" metrics. Just like the ones you mention. In order to apply AlphaTensor to sparse matrices, one would only need to change the dataset of randomly generated synthetic data (to contain only examples of sparse matrices). Then one could optimize for any metric of choice.
AlphaTensor (from DeepMind) discovers mathematical algorithms with Reinforcement Learning. While much of the attention has been on AlphaTensor's results, its successes more truly lie in the novel approach it uses rather than the results themselves.
In DeepMind's AlphaTensor Explained, I outline the details of the model and its key ideas. An objective assessment of the results obtained by AlphaTensor for matrix multiplication algorithms is also given at the end.
Interesting article. Classical ideas from physics and dynamical systems have been inspiring a number of successful new models in AI lately. I wonder, what is coming next
Thanks! Yeah, I've felt for a while that the greatest strides in deep learning will come from novel approaches that pull from higher level math/physics (or from novel approaches to automating the data cleaning process)
