This is much harder and more time-consuming than it sounds, especially when you go from a small(ish) number of features to more general knowledge. Worse, it doesn't scale to arbitrary domains - you'll always need a human there to give meaning to the models and effectively train them.
Reinforcement learning is designed to get around this by letting an agent "learn" meaning on its own by interacting with the world and getting feedback from its current state and actions. https://en.wikipedia.org/wiki/Reinforcement_learning
I tend to agree. Image classifiers and such appear much more powerful than they really are. People assume 'Oh, this computer knows what a Banana is!'. No it doesn't, It has no conception whatsoever of physical objects, texture, taste, smell or even the 3D shape of Bananas or anything else. It only works with 2D pixel data, and subtle manipulation of the image in ways barely detectable to humans can make even the best Banana classifier ever misidentify a car crash as a Banana. It's skipping directly from image to identification, without the intervening stages a human has in the real world of identifying shape, orientation, texture, relationship to other 3D objects and the ability to test a possible identification by utilizing other senses and interactions. Even for 2D picture we build mental models of the scene in ways image classifiers flat out don't and can't.
Reinforcement learning is really interesting though for several reasons. Algorithms like this and genetic algorithms can 'grow' sophistication far faster than we can program it. The agent takes actual actions that it learns from directly so there's richer feedback. Furthermore by analyzing them we can learn more about how systems learn across multiple problem types to achieve goals requiring multiple layers of sense, analysis, hypothesis, action and feedback.
No one approach is going to get this done. The brain consists of many layers and cortical columns, with many structures specialized for very different functions. I believe any strong AI will need to have such an architecture using various different approaches and techniques in concert. We have an advantage here because evolution only had neurons to work with so in the brain everything is a neuron but we can engineer whatever hardware or software implementation is most efficient for a specific function. It's till going to take probably another few generations though at least.
> Even for 2D picture we build mental models of the scene in ways image classifiers flat out don't and can't.
That's not true. Conditional Random Fields and other statistical models are being used to model spatial object relations. Example: https://arxiv.org/abs/1512.06790v2
I'm similar - I have an eclectic taste in music that is often not served well by static, long-term algorithms that place me into a "well" of the same sounding music. It's a problem I think about a lot.
Matrix factorization can deal with this a bit by using the high dimensional space to place your tastes into an area that reflects many different styles at the same time. In part 2 of the blog post we're going to talk about how we're modeling the acoustic qualities of music, which can find common patterns from completely different genres (for example, you may like soothing music with female vocals in both jazz and indie rock). In part 3 we'll talk a bit about how we can combine recent signals (like thumbs) to take into account your current mood, which I find helps pinpoint interesting music to surface right now.
> we're going to talk about how we're modeling the acoustic qualities of music, which can find common patterns from completely different genres
That's neat! I was curious of this is / was being looked into. It seems like I often get music that's matched based on a demographic (if that makes sense), rather than music matched on the characteristic features of the current song / band.
What we call genre is sometimes circumscribed by stylistic elements, sometimes by subculture (demographics as seen from the listeners' own perspective). But quite often since the rise of radio, it's circumscribed by target demographic as seen from advertisers' perspective.
The worst case of that is probably "new age", a label rejected by virtually all the artists so labeled (and most of the listeners), and having no common traits to speak of, but lumped together as whatever sold better in bookstores than in record stores.
>how we can combine recent signals (like thumbs) to take into account your current mood
This is the most interesting part of the problem to me. I always worry that the signals I send my favorite music radio service permanently alter the course of the channel. In some cases, I want that. In others I don't.
Agreed, music map is a great interactive way to discover music. Last year one of our internal hack week teams built something like that using our vector space model. What was cool is you could switch between the different types from the UI, e.g. Nirvana -> Red Hot Chili Peppers -> DC101 Alternative Rock. You could also start stations from the UI - it was really fun discovering music with it!
This is much harder and more time-consuming than it sounds, especially when you go from a small(ish) number of features to more general knowledge. Worse, it doesn't scale to arbitrary domains - you'll always need a human there to give meaning to the models and effectively train them.
Reinforcement learning is designed to get around this by letting an agent "learn" meaning on its own by interacting with the world and getting feedback from its current state and actions. https://en.wikipedia.org/wiki/Reinforcement_learning