3 ms·
Hype harms the brain. >> In the last couple of years, magic started happening in AI. Techniques started working, or started working much better, and new techni
by YeGoblynQueenne 10y ago
Hype harms the brain.
>> In the last couple of years, magic started happening in AI. Techniques started working, or started working much better, and new techniques have appeared, especially around machine learning ('ML'), and when those were applied to some long-standing and important use cases we started getting dramatically better results. For example, the error rates for image recognition, speech recognition and natural language processing have collapsed to close to human rates, at least on some measurements.
Let's see.
a) Magic? Really?
b) "The last couple of years" goes back to 2014. Machine learning's big break into the mainstream happened at least in 2012, when a conv net won ILSVCR [1]. Did the op not look further back before reporting on all this "magic"?
c) Unfortunately no "new techniques" have appeared. Convolutional neural networks date from 1988 [2]. LSTM Recurrent Neural Networks were first described in 1997 [3]. Machine learning in general is pretty much as old as GOFAI, with the earliest connectionist ideas detailed in the 1950s [4].
All this is really well known and understood and not at all controversial- Geoff Hinton himself is on record explaining that the recent boom is due to more data and more processing power (I can dig up a link if there is any doubt).
d) "the error rates for image recognition, speech recognition and natural language processing have collapsed to close to human rates, at least on some measurements"
In image recognition, the big success story is the ImageNet results detailed above. For speech processing I'm told there's been a big jump also, someone else could give an example.
For natural language processing: not in your wildest dreams. We're nowhere near "close to human rates" in any way, shape or form, unless you count very restricted results like how good this or that natural language parser does on a one-million word corpus like the Brown corpus. Unfortunately, that counts for nothing.
Now, most of that is stuff you can find easily on wikipedia with a bit of reading. There's absolutely no reason to waste peoples' times with disinformation and pointless overhyping.
[1] https://en.wikipedia.org/wiki/ImageNet https://en.wikipedia.org/wiki/ImageNet
[2] https://en.wikipedia.org/wiki/Convolutional_neural_network#History https://en.wikipedia.org/wiki/Convolutional_neural_network#H...
[3] https://en.wikipedia.org/wiki/Recurrent_neural_network#Long_short_term_memory https://en.wikipedia.org/wiki/Recurrent_neural_network#Long_...
[4] https://en.wikipedia.org/wiki/Perceptron#History https://en.wikipedia.org/wiki/Perceptron#History