3 ms·

That's not how ML works. You need to acquire labeled data to train on.

m_mueller 7y ago
Again, I'm talking about possible future technology. What if you can infer labels some other way? The fundamental element you need is not labels but a cost function - that could e.g. be curiosity and fear (of disappointing the end user). The most fundamental part that ML is missing to be AI, is to have a model of the world, with which such a cost function could evolve organically. Once you have that, AI can go and explore on its own.
- walterbell 7y ago
  Until humans stop providing free input to their machine (AI) adversary, e.g. because of new business models that pay some humans for high quality feedback. That will leave parasitic AIs competing for low quality signals.
- bunderbunder 7y ago
  If you can accurately, inexpensively, and automatically figure out what the labels on the classifier's training data should be, then you have no need for the classifier.
- codesushi42 7y ago
  That makes no sense. If you can already infer the labels beforehand, then you already have a working model, right? The job of the model is to infer those labels. The model needs to be trained to accurately predict those labels by using examples that were labeled beforehand.
  - m_mueller 7y ago
    By model of the world I mean something much more than what's called an ML model today. I mean the cognitive ability to predict the past and future of some part of the real world based on observation. Most simple example: Seeing two video frames of a massive object moving in front of a reference background and predicting the curve it's going to do, and where it came from. Only having such an understanding of the world (physical laws, human interactions with it etc.) allows to program in much more general cost functions that can get the system to explore on its own, not just within a given rule set, but the rule set itself.
    - codesushi42 7y ago
      Again, what you're proposing has no basis in today's reality of how ML is implemented. Especially not for classifying image data. What you're proposing is some kind of machine intelligence that can only exist right now in a William Gibson novel. Reinforcement learning is something similar, along with unsupervised learning. But you still need to define some kind of objective function and rules to play by. The AI cannot come up with this on its own. And the real world is far too complex to model all of these physical rules you've mentioned. Reinforcement learning can work for Go because the rules of the game are relatively simple. It is the number of possible moves/outcomes that make it enormously complex.

Again, I'm talking about possible future technology. What if you can infer labels some other way? The fundamental element you need is not labels but a cost function - that could e.g. be curiosity and fear (of disappointing the end user). The most fundamental part that ML is missing to be AI, is to have a model of the world, with which such a cost function could evolve organically. Once you have that, AI can go and explore on its own.

walterbell 7y ago
Until humans stop providing free input to their machine (AI) adversary, e.g. because of new business models that pay some humans for high quality feedback. That will leave parasitic AIs competing for low quality signals.
bunderbunder 7y ago
If you can accurately, inexpensively, and automatically figure out what the labels on the classifier's training data should be, then you have no need for the classifier.
codesushi42 7y ago
That makes no sense. If you can already infer the labels beforehand, then you already have a working model, right? The job of the model is to infer those labels. The model needs to be trained to accurately predict those labels by using examples that were labeled beforehand.
- m_mueller 7y ago
  By model of the world I mean something much more than what's called an ML model today. I mean the cognitive ability to predict the past and future of some part of the real world based on observation. Most simple example: Seeing two video frames of a massive object moving in front of a reference background and predicting the curve it's going to do, and where it came from. Only having such an understanding of the world (physical laws, human interactions with it etc.) allows to program in much more general cost functions that can get the system to explore on its own, not just within a given rule set, but the rule set itself.
  - codesushi42 7y ago
    Again, what you're proposing has no basis in today's reality of how ML is implemented. Especially not for classifying image data. What you're proposing is some kind of machine intelligence that can only exist right now in a William Gibson novel. Reinforcement learning is something similar, along with unsupervised learning. But you still need to define some kind of objective function and rules to play by. The AI cannot come up with this on its own. And the real world is far too complex to model all of these physical rules you've mentioned. Reinforcement learning can work for Go because the rules of the game are relatively simple. It is the number of possible moves/outcomes that make it enormously complex.