[h4b4n3r0]

Sure, but that’s DARPA, it’s by definition for stuff that won’t be practical for the next 20 years, and it’s orthogonal to what they are pursuing in the near future. In the near future the goal is pretty simple: do the same things as they do now, but with more drones and a lot fewer people. A lot of off the shelf stuff is ready to go now: classification, object detection, segmentation, etc. And it’s not even that hard to get off the ground if you have a large, high quality dataset.

[tsomctl]

Ok. I've spent the last several years working on neural networks for audio and music, and trying to reuse off the shelf classification and segmentation networks. It is ridiculously hard to get to work. I'm going to agree with BucketSort; I get the feeling that a lot of AI cheerleaders here have never actually trained a neural network.

[Spooky23]

I think his point isn’t that it is trivial, just possible. Off the shelf tools exist and practitioners (of which there are thousands) can build solutions now to identify tanks or people on a kill list, or whatever the military wants.

DARPA looks at what is plausible and spends money to get the brainpower (a number in the hundreds or less) who can create the tools and create the process thinking about things that the military wants or needs.

[h4b4n3r0]

What are you talking about? It’s ridiculously easy, to the point where my 14 year old can train a classifier. For classification in particular you don’t even need to do anything: just clone, point it to your dataset in one of the established formats (eg imagenet) and let it train for a few days. Object detection is quite a bit harder (and takes way longer to train even from a pretrained net), but again, totally doable half a dozen different ways using existing code you can get from GitHub.

If you’ve spent years doing this stuff and training a classifier is an insurmountable obstacle, you should consider changing your field of work.

[tsomctl]

Some of us are working on more complicated data than just handwritten digits. Part of the problem is that existing networks are tuned for the dataset that the original authors were working on. If you want to use it on completely different problems, you have to change the sizes of the layers, convolution size, max pooling, etc. The other problem is figuring out how to preprocess your data to make it as easy as possible for the network to digest. Then, to make it harder, changing the preprocessing means you have to change the network architecture, and vice versa. Fun times!

[ericd]

It can certainly be tricky. That said, if you've never used it, I highly recommend trying out adaptive max pooling.

[newen]

Try training a classifier that can detect a person holding weapon aiming to kill someone vs a person not. Make it a 99.9999% accuracy classfier under different weather conditions. Now make one for night vision images. Even seemingly straighforward classification problems can be hard. Not to mention the vast array of other problems out there.

[h4b4n3r0]

Such a thing doesn’t exist and it doesn’t need to, because humans have much lower accuracy than that. You don’t need to run faster than the bear, you need to run faster than the other guy, and that’s not hard to do if you pick the right task.

[star-techate]

^ everybody remember this line for when it's given in response to an angry Congressman.

I imagine another line will be "it's marginally better than statistical human shooting, according to three studies!"

And then in the 'big house' people will refer to you as "Butterstats" or something. Because you were screaming "marginally better" the first night they left you in your cell.

Yo, butter stats, how's your appeal going?