[Houshalter]

There is no mathematical distinction between different "tasks". Playing Go requires thousands or even millions of different heuristics and pieces of knowledge. Each of those is a "task" in and of itself. The same technology which beat Go could also, in principle, be used to play a real time video game that requires hunting and stalking things.

Insects are indeed tiny. But they are highly specialized. They can't learn to do things differently than they are programmed. GPUs are extremely general purpose. Not only can neural nets running on them learn many tasks, but they can execute many algorithms which aren't neural networks at all.

[Cybiote]

> The same technology which beat Go could also, in principle, be used to play a real time video game that requires hunting and stalking things.

If you generalize things enough to mean application of linear algebra, the chain rule and map, then yes that's true. However, the fact is that Neural networks still take a huge amount of skill and experimentation to design new architectures. The more powerful deeplearning becomes, the less skill will be required to do complicated things.

The AlphaGo bot was very data inefficient; at ~1 hr 18 min in http://techtalks.tv/talks/deep-reinforcement-learning/62360/, David Silver points out how they broke correlations by brute forcing millions of games while only keeping one move per game. We want something with better generalization ability than that. As amazing as AlphaGo was, it's more a very early waystation than our final destination. Recognizing the weaknesses of our models instead of glossing over them is how we get to AI!

[Houshalter]

>If you generalize things enough to mean application of linear algebra, the chain rule and map, then yes that's true. However, the fact is that Neural networks still take a huge amount of skill and experimentation to design new architectures. The more powerful deeplearning becomes, the less skill will be required to do complicated things.

Sort of. Hyperparameter tuning is indeed difficult, but in principle it can be automated and there has been great work on this in the past few years. But I never claimed that it was just plug and play. The point is it's the same underlying technology in speech recognition or self driving cars or face recognition. Even if the number of layers and convolution sizes and learning rates are totally different, the principles are the same.

>The AlphaGo bot was very data inefficient... David Silver points out how they broke correlations by brute forcing millions of games while only keeping one move per game.

This problem was solved in the past with temporal difference learning. Maybe there are technical reasons they didn't do that, I don't know. At the very least they could pick a random move from each game per batch, instead of discarding the rest of the moves. I don't really see how that helps solve anything.

>Recognizing the weaknesses of our models instead of glossing over them is how we get to AI!

There is a huge difference between recognizing weaknesses, and the AI Effect. Where AI solves problem X, and people disregard it "because it can't do Y".

[Cybiote]

> Sort of. Hyperparameter tuning is indeed difficult, but in principle it can be automated and there has been great work on this in the past few years.

There's far more to modification than just hyperparameter tuning; this is why people can keep turning out papers on the latest breakthrough. The gates of LSTMs are a non-trivial addition to RNNs. Attaching a differentiable stack or figuring out how to best do soft attention requires far more thought than just bashing at hyperparameters. CTC, dilated convolutions, residuals, extending auto-encoders to capture the data distribution were not trivial ideas. VGGNet and other conv nets are a very manual heavy design, differing significantly from vanilla MLPs and not something that could be automatically arrived at (at least not yet). Things like FractalNet or recursive tree based architectures differ greatly from vanilla networks.

> Even if the number of layers and convolution sizes and learning rates are totally different, the principles are the same.

Again, the underlying captured algorithms are very different. If you could look at the underlying source code these neural net programs represent, they'd vary at least as much from each other as the code for say a merge sort, insertion sort, binary search or an AVL tree would.

> This problem was solved in the past with temporal difference learning. Maybe there are technical reasons they didn't do that, I don't know.

The problem pointed out in the lecture, part of why it took so long to find a good Go policy, was that immediate moves differed so little from each other that it was too difficult to get any kind of signal without the brute force approach to decorrelating they took. There are probably better ways but that was the shortest delta available to progress.

[YeGoblynQueenne]

>> There is no mathematical distinction between different "tasks".

Sure there is. For instance, the minimax algorithm is a mathematical model of optimal behaviour in a specific kind of game (two-player, zero-sum, complete information games). Performing minimax to dominate such a game is a very clearly delimited task and that's exactly why a machine can do it in the first place.

More broadly, any "task" can be described mathematically by means of a program for a given machine. Saying there's no mathematical distinction between different tasks is like saying you can't write a program that does a specific thing.

>> Insects are indeed tiny. But they are highly specialized.

Agreed- see my comment above.

However, that on its own doesn't say much. Sure, insects have highly specialised firmware-style programming. How did they come to have them and how can we reproduce them? We can't hope to create such programs "by hand" so we're currently trying to create algorithms that learn them. We're still far from achieving it though.

That neural networks and other learning algorithms can learn all sorts of different things is true, however you have to train a new instance of an algorithm for every different thing you want it to learn (what I call a "task", so a new program). For each instance you need new data (and a lot of it) another huge chunk of processing power and a long time to train. Then you have to find a way to combine all the trained models into one ... thing that can operate as a coherent whole.

We haven't even started on the last bit.

Just having the technology to create pigments is far from being the same as creating a Mona Lisa.