3 ms·
The following opinion proved quite unpopular on Hacker news: "If I was going to try machine learning I would code up from scratch rather than relying on black
by Safety1stClyde 10y ago
The following opinion proved quite unpopular on Hacker news:
"If I was going to try machine learning I would code up from scratch rather than relying on black boxes."
Another one, a preference for Python (or Go, or Rust, or whatever it is next week) is due to its fashionability, rather than any intrinsic merit of the language.
See also this discussion:
https://news.ycombinator.com/item?id=14057953 https://news.ycombinator.com/item?id=14057953
- PaulHoule 10y agoI would disagree with that one. I worked on a neural network product that was coded in C++. It worked, and it came out early, before deep learning got bigger than beatlemania. It used intrinsics to incorporate SIMD instructions, thus it was tied to a particular processor version. I seemed to be the only person who noticed or cared if the derivatives were calculated correctly, or to implement constraints with Lagrange multipliers rather than wishful thinking. Once we got it working we would be crazy to change anything. Thus tensorflow, keras, and similar tools can save you a lot of grief.
- Safety1stClyde 10y ago> I seemed to be the only person who noticed or cared if the derivatives were calculated correctly, or to implement constraints with Lagrange multipliers rather than wishful thinking. In other words, it was a black box.
- PaulHoule 10y agoIt may have been to the people who made it. I was the one who had to understand it in order to make it ready to go in front of the customer.