Basic question: Why is this faster than running Intel Linux apps in an emulated Intel Linux VM? Because Rosetta is faster than QEMU, and you only need to emulate one application rather than the entire kernel?
In the same theme, there’s a recent documentary called “The Weight of Gold”, narrated by Michael Phelps and other Olympic champions. Really poignant stories.
1. I wonder how many times the test set can be used on "incremental changes in future versions of the model" before losing statistical validity.
2. This article describes their process, but not the FDA's process. Are there specific regulatory requirements for ML models beyond their four types of reports?
This article is about optimization (finding good parameters), not the approximation power of neural networks (which is well-known through the universal approximation theorem).
I was thinking MPC might work for set intersection. For oblivious query, can't you do it by sending the query directly into the remote trusted execution environment, encrypted with the TEE's public key?
The article claims that oblivious query, set intersection, and machine learning on private data are not possible without FHE. However, aren't they all possible either with secure MPC or hardware based enclaves e.g. AMD SEV?
A bunch of the papers I found from a quick Google search use secure MPC rather than FHE for interbank fraud. Which one is more practical these days wrt latency?
There are two ways to implement autograd, reverse-mode and forward-mode. Reverse-mode is what minigrad uses, and what most ML libraries these days use by default, since it computes gradients of all inputs (wrt one output) in a single pass. It's exactly what you describe in the 2nd paragraph.
Forward-mode autograd is the technique that can use dual numbers. It computes all gradients of one input (wrt all outputs) in a single pass. Dual numbers is a pretty neat mathematical trick, but I'm not aware of anyone that actually uses them to compute gradients.
The most approachable explanation of dual numbers I've seen is in Aurelien Geron's book Hands-On Machine Learning (Appendix D). There are articles online but I found them more technical.