| Honesty, repeatability, numerical analysis. Canonicalization. Honesty: how many times was the exact same data processed? Was the result cherry picked and the best one published? For the sake of integrity how is it possible to scientifically improve on this result? (example, your AI outputs some life altering decision?) Repeatability: In science, if a result can be independently verified, it gives validity to the "conclusion" or result. Most AI results cannot be independently verified. Not indepentently verifiable really ought to give the "science" the same status as an 1800's "Dr. Bloobar's miracule AI cure". Numerical Analysis: performing billions / trillions of computations on bit-restricted numerical values will introduce a lot of forward propagated errors (noise). What does that do? Commentary: Video cards don't care if a few bits of your 15 million pixel display are off by a few LSB bits, they do that 60 or 120 frames a second and you don't notice. It is an integral part of their design. The issue is, how does this impact AI models? This affects repeatability -> honesty. If error of a quantized size is a necessarily required property to achieve "AI learning that converges", there is still an opportunity for canonicalization -- a way to map "different" converged models to explain why they are effectively the "same". This does not seem to be a "thing", why not? In my opinion, in 2020, the AI emperor still has no clothes. |
When we talk about controlling machines, our control algorithms have mathematically proven strict error bounds, such that if we provide an input with a particular maximal error (e.g. from a sensor that has some error tolerance), we can calculate what's the maximum error possible in the response that our model would produce, and then use that to evaluate whether this is even an input that should be handled by the current algorithm or not.
These control algorithms all take some inputs, and use them to "predict" what will happen, and using that prediction, compute some response to correct it. These predictions need to happen much faster than real time, since you often need to perform an optimization step to compute an "optimal" response.
These predictions are usually computed using a reduced-order model, e.g., if you had to solve a PDE over 10^9 unknowns to compute the actual prediction, you can instead reduce that to a system with 10 unknowns, by doing some pre-computation a priori. Most tools to do these kinds of reductions developed in the last 60 years come with tight error bounds, that tell you, depending on your inputs, the training data, etc. what's the largest error than the prediction can have, so you can just plug these in into your control pipeline.
People have been plugin in neural-networks to control robots, cars, and pretty much anything you can imagine into these pipelines for 10 years, yet nobody knows what the upper bound on the errors that these neural-networks give for a particular input, training set, etc.
Until that changes, machine learning just makes your whole pipeline unreliable, and e.g. a car manufacturer must tell you that in "autonomous driving" mode you are liable for everything your car does, and not them, so you have to keep your hands on the driving wheels and pay attention at all times, which... kind of defeats the point of autonomous driving.
---
Prediction: we won't have any tight error bounds for real-world neural networks in the 2020-2030 time frame. These are all non-linear by design (that's why they are good), error bounds for simple non-linear interpolants are pretty much non-existent, people have tried for 20-30 years, and real-world NNs are anything but simple.