Machine Learning

Stream: general

[Simon Hudon]

@Andrew Ashworth Let's switch threads

[Simon Hudon]

You should consider changing names :stuck_out_tongue_closed_eyes:

[Andrew Ashworth]

but yeah, ml isn't going to break into real time radar imaging anytime soon, mostly because you need a ton of compute horsepower

[Simon Hudon]

if it did, what would it improve?

[Andrew Ashworth]

depends on what your end application is

[Andrew Ashworth]

for radar, a big problem is distinguishing objects from noise

[Andrew Ashworth]

we use tons of linear approximations everywhere that aren't totally applicable

[Andrew Ashworth]

but who cares because everything is a gaussian if you actually want to compute things

[Andrew Ashworth]

unfortunately noise is rarely linear, rarely uncorrelated, and better ways of doing it are too complicated to do in real time

[Andrew Ashworth]

but maybe a well trained ml model would be efficient enough? I don't know, but i think it might be true in the future

[Andrew Ashworth]

gotta stay ahead of the curve and do some learning :)

[Simon Hudon]

If it was computationally feasible, would it bring in safety risks?

[Andrew Ashworth]

hmm, possibly, but we already accept some risk of error in the process

[Andrew Ashworth]

since it's statistical, we design the system to some probability pf of a false alarm

[Andrew Ashworth]

(of detecting an object, in this example)

[Andrew Ashworth]

can you talk about that false alarm probability in a deep learning model? kinda... sorta... I don't know well enough to say that rigorously

[Simon Hudon]

Already having an estimate for the probability of failure of a ML model seems pretty difficult, isn't it?

[Andrew Ashworth]

yup. but even a handwavy one would be nice though

[Andrew Ashworth]

i mean, current estimates are based on assuming gaussian noise, which doesn't reflect reality

[Andrew Ashworth]

it's a complicated subject

[Simon Hudon]

I'm wondering if the best that can be done with an ML system is counting errors and retiring / replacing models whose number of errors exceeds the tolerance. I'm not sure if it's any good though

[Simon Hudon]

Sorry, I might not be focusing on what you care about, I think a lot about guarantees. In your case, the hope seems to be an increase in average accuracy, does that make sense?

[Andrew Ashworth]

yeah. it doesn't make sense to speak of guarantees in object detection, just minimizing the ratio of detections to false alarms

[Simon Hudon]

You could think of a guarantee as an upper bound on the probability of failure

[Simon Hudon]

Whenever I think of ML lately, I think of a software tool that was on the news recently where judges would use ML to help them make their judgement faster. It was found that the process exaggerated an existing racial bias. I see that as a conflict between efficiency and correctness. If you can't make any statement about the correctness of the software, what good is it?

[Andrew Ashworth]

haha, that's a dangerous statement to make

[Andrew Ashworth]

cowboy c coders brought us the computer revolution :)

[Simon Hudon]

My skin has grown impervious to the rocks I'm being thrown :stuck_out_tongue_winking_eye:

[Simon Hudon]

Which revolution do you mean? The age of personal computing or putting white supremacists in the White House?

[Andrew Ashworth]

both, because as you know, it's a fact that automation has put many out of work, but that's getting a little off-topic

[Simon Hudon]

That's true (both things)

[Andrew Ashworth]

I'm wondering if the best that can be done with an ML system is counting errors and retiring / replacing models whose number of errors exceeds the tolerance. I'm not sure if it's any good though

back in the day, when humans interpreted radar returns, firing a bad operator was exactly what was done

[Simon Hudon]

I'm slowly getting to respect the development of cowboy developers and I realize that everything doesn't need to be bullet proof. But there is an extent where rolling out a technology should be after proper study