Hacker News Reader

Pretty much this. In the real world you have to deal with

- Too few labeled / garbage labeled data (70000 digits? How about only 1000 complex class objects?)

- Obscure bugs in custom implementation (yeah my custom layer works and gradient is correct... or wait why it diverges after 10k iteration? hmm).

- Timing/RAM constraints (it should segment an image under 10ms on Jetson TX1, well, good luck with GoogLeNet)

View on HN

Nailed it. Neural net execution speed is so critical for may production systems and it's very difficult to hit the sweet spot on trade offs but I never hear about these issues in wild.

> Nailed it. Neural net execution speed is so critical for may production systems and it's very difficult to hit the sweet spot on trade offs but I never hear about these issues in wild.

That's probably because you're not listening. There's plenty of literature on scaling down neural nets to smaller devices because everyone knows it's an issue and can trivially get a smaller device, see techniques such as Quantization , Distillation, Pruning or architectures specifically designed for the task such as YOLO.

Reply on news.ycombinator.com