3 ms·

I wonder how do unit tests work with NN (or if they're even a relevant concept at all). You could replay some testing video frames and make sure the objects ar

by bsaul 8y ago

I wonder how do unit tests work with NN (or if they're even a relevant concept at all). You could replay some testing video frames and make sure the objects are correctly identified, but i suppose that's already what training is about... If an issue like that resurface, does it mean that the original frames leading to the 2018 accident aren't part of the training (or at least frames from someone driving in this kind of scenario) ?

inertiatic 8y ago
Well, testing in the strict sense would be measuring accuracy on data not in the training set, and even accuracy on the training set data (which isn't guaranteed to be 100% I believe)
bluGill 8y ago
The frames from the 2018 should not be part of the training. They should be part of the test set to prove whatever training they do works.
neilalexander 8y ago
Yes, the problem is indeed that there's no real way to "look into" a neural network and understand how it has been trained. All you can do is observe that the given inputs generate the desired outputs. Even if there was training based on the 2018 frames, that doesn't mean that you have verifiably fixed the problem. It's difficult to train a neural network selectively – every time you "train" the network with additional data, you are increasing the chance that you are also teaching it something you didn't intend which then can have a side-effect in some seemingly unrelated scenario. You can see this in real life with image recognition networks. Teach them too much and they gradually become less effective at identifying anything.
jononor 8y ago
Unit tests are not sufficient for neural networks. Say the network takes 2D images of typical 224x224 pixels with 3 channels (RGB) of 8 bit values, this input space has 256^(224x244x3) = 3.5×10^362507 possibilities. Billions of years to test them all. This is before we consider stereo vision, 3D images and state over time. How to know which subset of these inputs are necessary to give reasonable coverage? Right now I don't think we have very good answers to these things. Of course one can always add a regression test (with some K mutations) when someone crashes. It is better than nothing, but hardly good assurance that something like this will never happen again. The entire area of safety and quality assurance with neural networks is still in active research actively, from multiple angles. For some examples of how chaotic neural networks can be, look up 'adversarial examples'.
dv_dt 8y ago
My first thought was, well someone is in store for a revamp of their regression test suite.
ummonk 8y ago
Should be using integration tests, not unit tests. I.e. SITL simulations + having the car drive around test circuits / scenarios.
skwb 8y ago
You don't unit test a NN. You can unit test certain functions, but fundamentally this is an integration test. This is why serious scientific training is needed to understand these complex systems when health and safety are on the line.
- leggomylibro 8y ago
  "Autonomous vehicle integration test track" could make for a great setting in a spy thriller. The villain could own the megacorporation which makes the cars, and the heroes could find evidence of their evil plot among the sprawling acres of labs and potemkin streets. But then, in the distance, the sound of revving engines... Seriously though, I wonder if that sort of physical test track will become popular. You would load your build onto an idle car, queue it up, and make sure that it didn't hit any of the silhouettes which spring up, unusual traffic and weather conditions, etc. They must already do that in some capacity, right?
  - saalweachter 8y ago
    > "Autonomous vehicle integration test track" Just say "the 101", it's shorter.

by bsaul 8y ago

inertiatic 8y ago

Well, testing in the strict sense would be measuring accuracy on data not in the training set, and even accuracy on the training set data (which isn't guaranteed to be 100% I believe)

bluGill 8y ago

The frames from the 2018 should not be part of the training. They should be part of the test set to prove whatever training they do works.

neilalexander 8y ago

Yes, the problem is indeed that there's no real way to "look into" a neural network and understand how it has been trained. All you can do is observe that the given inputs generate the desired outputs. Even if there was training based on the 2018 frames, that doesn't mean that you have verifiably fixed the problem. It's difficult to train a neural network selectively – every time you "train" the network with additional data, you are increasing the chance that you are also teaching it something you didn't intend which then can have a side-effect in some seemingly unrelated scenario. You can see this in real life with image recognition networks. Teach them too much and they gradually become less effective at identifying anything.

jononor 8y ago

Unit tests are not sufficient for neural networks. Say the network takes 2D images of typical 224x224 pixels with 3 channels (RGB) of 8 bit values, this input space has 256^(224x244x3) = 3.5×10^362507 possibilities. Billions of years to test them all. This is before we consider stereo vision, 3D images and state over time. How to know which subset of these inputs are necessary to give reasonable coverage? Right now I don't think we have very good answers to these things. Of course one can always add a regression test (with some K mutations) when someone crashes. It is better than nothing, but hardly good assurance that something like this will never happen again. The entire area of safety and quality assurance with neural networks is still in active research actively, from multiple angles. For some examples of how chaotic neural networks can be, look up 'adversarial examples'.

dv_dt 8y ago

My first thought was, well someone is in store for a revamp of their regression test suite.

ummonk 8y ago

Should be using integration tests, not unit tests. I.e. SITL simulations + having the car drive around test circuits / scenarios.

skwb 8y ago

You don't unit test a NN. You can unit test certain functions, but fundamentally this is an integration test. This is why serious scientific training is needed to understand these complex systems when health and safety are on the line.

leggomylibro 8y ago
"Autonomous vehicle integration test track" could make for a great setting in a spy thriller. The villain could own the megacorporation which makes the cars, and the heroes could find evidence of their evil plot among the sprawling acres of labs and potemkin streets. But then, in the distance, the sound of revving engines... Seriously though, I wonder if that sort of physical test track will become popular. You would load your build onto an idle car, queue it up, and make sure that it didn't hit any of the silhouettes which spring up, unusual traffic and weather conditions, etc. They must already do that in some capacity, right?
- saalweachter 8y ago
  > "Autonomous vehicle integration test track" Just say "the 101", it's shorter.