3 ms·
It is reductionistic, because so much comes from how you arrange those switches, the current state of the switches, and the sequence in which you tell those swi
by ericd 6y ago
It is reductionistic, because so much comes from how you arrange those switches, the current state of the switches, and the sequence in which you tell those switches to switch. And I think most in the 1940s would be very surprised at what computers can do now. It only takes little imagination because we’ve already seen it in action.
Similarly, you can say that deep learning is just a series of matrix multiplications with nonlinearities applied, which is true, but it certainly wasn’t obvious to most that when you scale that way up, that that would lead to computers being able to interpret images, sound, and text. For example, my AI course professor described neural nets as something of interest mostly for historical reasons, something that was tried, but was an evolutionary dead end. And he was no slouch.
The hype has gotten ahead of its current abilities, but I simultaneously think that it will be one of the most impactful inventions we’ve ever created.
- mmcdermott 6y ago> The hype has gotten ahead of its current abilities, but I simultaneously think that it will be one of the most impactful inventions we’ve ever created. This statement is one of the things I find most fascinating about the current state of AI. Do I think deep learning is incredibly useful? Absolutely and I look forward to seeing more awesome stuff done with it in the future. Do I think the singularity or "true" AI will be based on deep learning? Not really. I strongly suspect anything capable of reaching scifi levels of intelligence (which a lot of hype says or strongly implies is imminent) will have to be a difference of kind, not of degree. I could be wrong, of course, but it will be interesting to wait and see.
- ericd 6y agoYeah, I tend to be more in the difference in degree camp, with the caveat that it's pretty clear we're at least a few discoveries of structural mechanics away from mimicking human-style learning. We need to make advances in semi-supervised learning, as well as increasing sample efficiency, and reducing catastrophic forgetting. But my intuition says that the solutions will be pretty simple, and seem obvious in hindsight. But I'm also guessing we'll come up with an endless number of complex tweaks to get the current stuff closer before someone figures out the simple mechanics and sweeps most of that away.