3 ms·

Thanks for sharing. I guess those Data science optimization requires some understanding of the inner workings? Embedded system programming is also fun. I did i

• PaulHoule 5y ago
  In that case I mentioned part of it was a math problem (being able to prove that if I cut this corner it causes less than ε error) and the other part understanding how to do the computation with mechanical sympathy. When it comes to ML/AI things often the work just starts when a "data scientist" develops a model, often a lot of engineering goes into making the model fast and reliable. I worked on an neural network powered search engine for patents that came out just before GPUs were popular for neural networks. Somebody hard-coded the math to train and do inference on the network in SIMD instructions. If we wanted to change anything about the network, error function, training procedure, etc. we would have had to do assembly programming. It's terrible tedious work, and, worse than that, if you get the math slightly wrong you get a system that seems to train a network, it just trains the network slowly and poorly. Thank God I never had to change the network but if I did I probably would have lied to my boss and wrote an automatic differentiator and compiler to SIMD instructions because it probably wouldn't have taken much longer than a few rounds of doing it by hand and developing exhaustive test suites to know I did it right.