Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

2 times 3 can sometimes equal 7 with Android's Neural Network API

alexanderganderson.github.io

Post reply on HN
Print view View original

31–40 of 92 posts

Re: 2 times 3 can sometimes equal 7 with Android's Neural Network API

#31

Post by Judgmentality »

Hasnep wrote: post #29
Judgmentality wrote: post #21

Earlier quoted context omitted.

> I'd trust a machine to drive more accurately than most people I see on the highway. Humans aren't special, in fact more often than not we're sloppy, subject to fatigue, and a whole bunch of other negative things. Humans are much, much, much more capable than the absolute state-of-the-art robots when it comes to doing things in an uncontrolled environment. https://www.youtube.com/watch?v=g0TaYhjpOfo

One of the advantages of an autonomous driver is that its superhuman reflexes, never driving while tired, never getting road rage, etc., will make it less likely to get into an uncontrolled environment. Would you prefer your pilots to fly your plane with no AI assistance?

> One of the advantages of an autonomous driver is that its superhuman reflexes, never driving while tired, never getting road rage, etc.

First of all, when you actually understand a self-driving car stack, you'll realize those super-human reflexes are more human than you think. The stack is complicated and not only are there delays to be expected, some hardware syncing requirements guarantee certain delays in the perception pipeline. It's still better than a person, but it's nothing close to approaching instantaneous. Likewise, sensors can get dirty, and blah blah blah there are other weaknesses robots have that humans don't. My point is simple: robots aren't perfect. In fact, they are almost always much worse than most people realize.

> will make it less likely to get into an uncontrolled environment

You're misunderstanding me. I'm not saying less likely to get into an accident. I'm saying the world, where cars drive, is an uncontrolled environment - and the current state of robotics is such that humans are better for doing things in the real world. There is no "less likely to get into an uncontrolled environment" because by definition you are always putting it into that situation.

> Would you prefer your pilots to fly your plane with no AI assistance?

AI assistance is fine. AI replacement is not.

Re: 2 times 3 can sometimes equal 7 with Android's Neural Network API

#32

Post by danfang »

Grimm1 wrote: post #9
Avalaxy wrote: post #7

Using a neural network for things that have clear cut rules is wrong. When you know the exact rules, implement them as such, instead of bruteforcing a guesstimation. This is also why I'm sceptical of the usr of GPT-3 for all sorts of purposes where accuracy is important. Think of the code generation case. Bugs may be very subtle and may go unnoticed.

Code generation only needs to generate code with n bugs where n is less than the number of bugs a human developer generates for it to have usefulness, and maybe some other factor of severity where they are generally less severe than human developers. I think it'll make neat autopilot functionality for developers but not replace the need to have someone look over and understand the code.

This is naive. The point is that code is a well defined system with clear rules that can be expressed through logic and mathematics. GPT is suited to approximate systems where the rules are not well defined. Until AI can actually learn the principles of logic, it may not be useful for code generation on a meaningful scale, other than things just like simple auto-completions.

Not only that, AI would also have to learn the principles of system design, performance, security, readability, maintainability. That's what makes "good" software. It's a far stretch to say that AI could achieve anything of the sort based on current abilities.

Re: 2 times 3 can sometimes equal 7 with Android's Neural Network API

#33

Post by justicezyx »

vmception wrote: post #5
justicezyx wrote: post #4

Earlier quoted context omitted.

Or one can claim that it's entirely obvious when relating that with human beings making mistakes, where not only 2*3 can be 7, millions can die of some obscure disctators whim, without much conacusoly realized the insanity...

Did someone just train a GAN on HN comments?

“The real question is not whether machines think but whether men do. The mystery which surrounds a thinking machine already surrounds a thinking man.”

-- B F Skinner

Re: 2 times 3 can sometimes equal 7 with Android's Neural Network API

#35

Post by skybrian »

Avalaxy wrote: post #7

Using a neural network for things that have clear cut rules is wrong. When you know the exact rules, implement them as such, instead of bruteforcing a guesstimation. This is also why I'm sceptical of the usr of GPT-3 for all sorts of purposes where accuracy is important. Think of the code generation case. Bugs may be very subtle and may go unnoticed.

This is good reason to stick with use cases where a human can verify the results, like autocomplete or something like a search engine interface.

Re: 2 times 3 can sometimes equal 7 with Android's Neural Network API

#36

Post by shoyer »

As someone who builds neural networks routinely, this sort of non-reproducibility sounds troubling to me. We expect small differences for floating point arithmetic between platforms, but integer math is typically exact.

This is all the more concerning for 8-bit quantized arithmetic, where off-by-one means a relative error of about half a percent. If a individual layers in a quantized neural net have off-by-one errors with a consistent bias, I can imagine these errors accumulating into significant losses in model quality in deep networks. There isn't a huge margin for error in quantized neural nets.

One concern about the article: it uses the word "non-deterministic" in a slightly misleading way. I assume any specific hardware is still expected to produce consistent results when run twice on the same input. So it's more non-reproducible than non-deterministic. Compensating for inconsistent arithmetic on different devices sounds much more feasible than compensating for stochastic arithmetic.

Re: 2 times 3 can sometimes equal 7 with Android's Neural Network API

#39

Post by Grimm1 »

Judgmentality wrote: post #21
Grimm1 wrote: post #16

Earlier quoted context omitted.

I don't want to die when I crash my own car, and I already debug my own apps at 12am. If your argument is that things need to be perfect than my god you must never leave your home! I'd trust a machine to drive more accurately than most people I see on the highway. Humans aren't special, in fact more often than not we're sloppy, subject to fatigue, and a whole bunch of other negative things. That considered, I had a p…

> I'd trust a machine to drive more accurately than most people I see on the highway. Humans aren't special, in fact more often than not we're sloppy, subject to fatigue, and a whole bunch of other negative things. Humans are much, much, much more capable than the absolute state-of-the-art robots when it comes to doing things in an uncontrolled environment. https://www.youtube.com/watch?v=g0TaYhjpOfo

"That considered, I had a pretty strict qualifier in my above post which means the machine must perform better than the average human in the respective task and therefore I'd be more likely to die driving my own car than a machine meeting my prerequisites."

Have to read the whole comment before replying. You can't just grab individual statements out of an entire argument and choose to go after those. I mean you can, but you can't expect someone to actually engage you then.

Re: 2 times 3 can sometimes equal 7 with Android's Neural Network API

#40

Post by amelius »

shoyer wrote: post #36

As someone who builds neural networks routinely, this sort of non-reproducibility sounds troubling to me. We expect small differences for floating point arithmetic between platforms, but integer math is typically exact. This is all the more concerning for 8-bit quantized arithmetic, where off-by-one means a relative error of about half a percent. If a individual layers in a quantized neural net have off-by-one errors…

> We expect small differences for floating point arithmetic between platforms, but integer math is typically exact.

Then perhaps think of the integers as fixed point numbers.

Post reply on HN