Machine learning is a kind of artificial intelligence? Lately I've been meeting more and more such projects. Is this likely to mark the beginning of a new era?
It can be quite confusing. Machine learning, artificial intelligence, deep learning - those are actually different things.
First of all, for those unfamilliar with these things or just heard it somewhere, I really recommend to watch Nvidia GTC 2017 on their official Youtube channel, what Jensen Huang tells about AI, and some of the examples. Even for those who not technical, it's very fun to watch. And as specialist in deep learning i can say that what they say is quite accurate and state-of-the art in AI field. They put a great effort to make GPUs for deep learning, it's kinda hard to say what they focused more on these days, gaming or deep learning...
Machine learning (or data science, data mining) usually means such things as cluster analysis, staticstical methods, and neural networks. Although when you see "machine learning" it's usually means some simple neural networks or statistical methods without using neural networks. Those things are working fine on CPUs.
Deep learning, on the other hand, is the edge of artificial intelligence right now, and this is the task Volta GPUs with tensor cores were developed for. These neural networks are deep in terms of layers. If in perceptron there's 3 layers (or actually 1, not counting input and output), in deep net it's not surprise to see more than 100 high-dimensional layers (tensors). It's very computational intensive. People waiting for weeks to train one such net on several Titan GPUs, and it's just for one experiment, with parameters that can be optimized for speed, not for accuracy, because it's already takes too long. Exaples of such networks are ResNet, Goog
LeNet, VGGNet(if interested, google vggnet first). Application includes autonomus driving, photorealistic scene and face generation, understanding and reasonong about scene, and a lot of other things. I'm personally working VQA, something that can lead to universal AI, applicable to any field, like our own intelligence.
When i've tried to create tensor in Torch with algo from a paper that was interesting to me, it said that i don't have enough ram. It asked for something like 50 gigs of ram. So i've been forced to change the code and reduce the size of tensor just to be able to launch the script. Therefore, if i'll have hypootecicaly unlimited computational power i would try different configuration of neural network architecture, datasets, and other hypermarameters. Without need to wait for month to see the result of my experiment, i'll launch multiple experiment in parallel, tweaking my code, see what's works. Computational power is the main constraint in this field, so the demand is huge, if not say infinite.
The term Artificial Intelligence is more broader, it describes the entire field of intellectual systems and often serves in marketing purposes, because many scientists hate that term, because it's too sci-fi and bring back memories of high-demand-low-deliver. There's a field of AI called Knowlege Engineering, which usualy means creating the Expert System that will give the same conclusion as an Expert in the corresponding field, using knowlege base created by developers. For example, system asks you for some symptoms and give you recomendations on treatment. If it's unsure, it'll ask additional questions. You can ask it to describe why it gave such recommendation. This type of AI doesn't require machine learning or significant computational power. Only a lot of human work is required, which may be the one reason it didn't became very popular.