3 ms·

Curious; If you train a neural net to drive a car with X number of samples, and the car does something erratic, how do you determine which data item introduced

by aantix 8y ago

Curious; If you train a neural net to drive a car with X number of samples, and the car does something erratic, how do you determine which data item introduced the erratic behavior and how do you go about untraining that behavior?

Hydraulix989 8y ago
Gradient descent is done in batches and over the aggregate collection of the training data so it's pretty lossy. Usually the error is more of the fault of the learned model itself, and then the culprit is attributed to neuron "mis"-activations post-hoc. Often, problems take the form of generalization error instead. And yes, all of these problems are still quite unsolved, which is why we don't have end-to-end neural network-based self-driving cars, we just have rule based systems with a perception framework that just uses convnets for vision.
- ghaff 8y ago
  I think there's this popular perception that you just stream in lots of data from road vehicles and the driving just gets better over time--rather than continuing to discover individual edge cases more or less manually and updating rules based on those.
repsilat 8y ago
That is one reason they don't do that. (Another is that the net would be useless if you upgraded the cameras...) Whatever the sensing technology, everyone uses it to build a "state of the world" model which is then fed into algorithms for planning movement. Nobody (to my knowledge) just wires up input pixel neurons to output pedal neurons.
piccolbo 8y ago
It's simple, you retrain the model several times removing one sample at a time (leave-one-out cross-validation), until the erratic behavior disappears. You repeat until there are no more instances of erratic behavior. Sorry couldn't resist the attempt at humor, but also consider it as an argument by contradiction that what you are asking is not possible. It's like asking which lesson caused me to hit a forehand over the tramlines. The answer is the one I didn't have time for.

by aantix 8y ago

Hydraulix989 8y ago

Gradient descent is done in batches and over the aggregate collection of the training data so it's pretty lossy. Usually the error is more of the fault of the learned model itself, and then the culprit is attributed to neuron "mis"-activations post-hoc. Often, problems take the form of generalization error instead. And yes, all of these problems are still quite unsolved, which is why we don't have end-to-end neural network-based self-driving cars, we just have rule based systems with a perception framework that just uses convnets for vision.

ghaff 8y ago
I think there's this popular perception that you just stream in lots of data from road vehicles and the driving just gets better over time--rather than continuing to discover individual edge cases more or less manually and updating rules based on those.

repsilat 8y ago

That is one reason they don't do that. (Another is that the net would be useless if you upgraded the cameras...) Whatever the sensing technology, everyone uses it to build a "state of the world" model which is then fed into algorithms for planning movement. Nobody (to my knowledge) just wires up input pixel neurons to output pedal neurons.

piccolbo 8y ago

It's simple, you retrain the model several times removing one sample at a time (leave-one-out cross-validation), until the erratic behavior disappears. You repeat until there are no more instances of erratic behavior. Sorry couldn't resist the attempt at humor, but also consider it as an argument by contradiction that what you are asking is not possible. It's like asking which lesson caused me to hit a forehand over the tramlines. The answer is the one I didn't have time for.