As someone who works on ML, in particular 'deep neural networks', I am always surprised when someone seriously mentions using them to solve complex (anything with humans in the loop fits solidly within this category) real-world (i.e. real-world inputs and outputs) problems. I am very surprised when the problem definition additionally involves potential injury and death.
And yet it seems that Karpathy firmly believes that Neural-network-based self-driving systems are not only reasonable, but within reach using something close to today's algorithm.
This always leaves me wondering whether (i) I am somehow using an entirely different class of algorithms from what these guys are using (ii) the AI/ML bubble has even the top minds in our field drinking a little too much kool-aid (iii) My colleagues and I are just not good enough at ML to unlock it's full magical potential.
(To be more exact, I am not debating the truly impressive power of today's ML/deepNN algorithms when it comes to statistical inference. Yet, does anyone here believe that you can get to Level 5 on today's roads without something close to general-purpose intelligence? How do you react to written roadsigns explaining temporary route alterations? How do you know to take special precautions in situation where context indicates the likelihood of unmodeled irrational behavior?)
I am genuinely curious, who here believes/doesn't believe today's AI can really generalize that far when (i) adversarial attacks are still a real issue [1] (ii) random seeds can sometimes have a greater influence than model architectures [2] (iii) we are not even sure how well these models approximate learning which takes place in biological brains [3]
[2] "We do not find any evidence
that the considered models can be used to reliably learn
disentangled representations in an unsupervised manner
as random seeds and hyperparameters seem to matter
more than the model choice." - https://arxiv.org/pdf/1811.12359.pdf
If somebody is paid extremely good money to work on doing FSD using ML snake oil, and is completely insulated from any failures of the system, what else would you expect them to say?
Same thing when you have G/FB/T people believing that they just need to adjust the algorithm a bit, and that will miraculously fix all the negative social effects of their product.
When you are a hammer, the whole world looks like a nail.
I also work in ML/DL. I'd use a different word than "surprised". "Appalled" is more appropriate. If the video proves anything, it's that everything they do is a hack. There is no principled way to do anything. There is something seriously wrong when they have to resort to oversampling to deal with catastrophic forgetting, aka, learning instability. Yet I'm supposed to trust my life with this thing? Today's DNN are still incredibly crude. There are fundamental problems that need to be solved before we can deploy them in anything other than toys.
3 comments
[ 3.5 ms ] story [ 35.6 ms ] threadAnd yet it seems that Karpathy firmly believes that Neural-network-based self-driving systems are not only reasonable, but within reach using something close to today's algorithm.
This always leaves me wondering whether (i) I am somehow using an entirely different class of algorithms from what these guys are using (ii) the AI/ML bubble has even the top minds in our field drinking a little too much kool-aid (iii) My colleagues and I are just not good enough at ML to unlock it's full magical potential.
(To be more exact, I am not debating the truly impressive power of today's ML/deepNN algorithms when it comes to statistical inference. Yet, does anyone here believe that you can get to Level 5 on today's roads without something close to general-purpose intelligence? How do you react to written roadsigns explaining temporary route alterations? How do you know to take special precautions in situation where context indicates the likelihood of unmodeled irrational behavior?)
I am genuinely curious, who here believes/doesn't believe today's AI can really generalize that far when (i) adversarial attacks are still a real issue [1] (ii) random seeds can sometimes have a greater influence than model architectures [2] (iii) we are not even sure how well these models approximate learning which takes place in biological brains [3]
[1] https://sites.google.com/view/lidar-adv
[2] "We do not find any evidence that the considered models can be used to reliably learn disentangled representations in an unsupervised manner as random seeds and hyperparameters seem to matter more than the model choice." - https://arxiv.org/pdf/1811.12359.pdf
[3] https://www.cell.com/trends/cognitive-sciences/fulltext/S136...
Same thing when you have G/FB/T people believing that they just need to adjust the algorithm a bit, and that will miraculously fix all the negative social effects of their product.
When you are a hammer, the whole world looks like a nail.