3 ms·

> then we could name features This is much harder and more time-consuming than it sounds, especially when you go from a small(ish) number of features to more g

> then we could name features 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 https://en.wikipedia.org/wiki/Reinforcement_learning

simonh 10y ago
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.
- visarga 10y ago
  > 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 https://arxiv.org/abs/1512.06790v2

simonh 10y ago

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.

visarga 10y ago
> 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 https://arxiv.org/abs/1512.06790v2