3 ms·

I firmly believe that general AI will not be developed without agency for the AI. The "info only" helper AI (called Tool AI) means that information will have to

by daveguy 10y ago

I firmly believe that general AI will not be developed without agency for the AI. The "info only" helper AI (called Tool AI) means that information will have to be added manually by some intelligent agent (human or otherwise). No exploration of how actions and interactions affect results can be explored. Tool AIs will never "want" anything because the meaning of want will be completely foreign.

visarga 10y ago
You're not alone. Yann LeCun's famous cake analogy associates the cherry on the top with reinforcement learning (agency, being a part of an environment from which it can learn and explore). I especially liked this passage. It connects many ideas onto one theme: > CNNs with adaptive computations will be computationally faster for a given accuracy rate than fixed-iteration CNNs, CNNs with attention classify better than CNNs without attention, CNNs with focus over their entire dataset will learn better than CNNs which only get fed random images, CNNs which can ask for specific kinds of images do better than those querying their dataset, CNNs which can trawl through Google Images and locate the most informative one will do better still, CNNs which access rewards from their user about whether the result was useful will deliver more relevant results, CNNs whose hyperparameters are automatically optimized by an RL algorithm will perform better than CNNs with handwritten hyperparameters, CNNs whose architecture as well as standard hyperparameters are designed by RL agents will perform better than handwritten CNNs… and so on. (It’s actions all the way down.) This is the general trend. Going meta, through RL. Optimize the optimizer, learn gradient descent by gradient descent.
- gwern 10y ago
  > learn gradient descent by gradient descent. As far as 'learning gradient descent by gradient descent' goes, I think it's still an open question if a tiny RNN actually can improve meaningfully over ADAM etc :) I don't recall the paper showing any wallclock times, which suggested to me that the RNN was way slower even if it trained faster. The individual parameter adjustments are so low in the hierarchy of actions that the value may be minimal compared to higher up like architecture design.
  - AlexCoventry 10y ago
    What paper are you referencing?
    - detaro 10y ago
      I assume this https://arxiv.org/abs/1606.04474 https://arxiv.org/abs/1606.04474 (since the title is basically the quoted text)
- skybrian 10y ago
  I'm not sure that trawling Google Images counts as active. It doesn't seem like it can be any more powerful than being given the entire Google Images database in advance.

by daveguy 10y ago

visarga 10y ago

You're not alone. Yann LeCun's famous cake analogy associates the cherry on the top with reinforcement learning (agency, being a part of an environment from which it can learn and explore). I especially liked this passage. It connects many ideas onto one theme: > CNNs with adaptive computations will be computationally faster for a given accuracy rate than fixed-iteration CNNs, CNNs with attention classify better than CNNs without attention, CNNs with focus over their entire dataset will learn better than CNNs which only get fed random images, CNNs which can ask for specific kinds of images do better than those querying their dataset, CNNs which can trawl through Google Images and locate the most informative one will do better still, CNNs which access rewards from their user about whether the result was useful will deliver more relevant results, CNNs whose hyperparameters are automatically optimized by an RL algorithm will perform better than CNNs with handwritten hyperparameters, CNNs whose architecture as well as standard hyperparameters are designed by RL agents will perform better than handwritten CNNs… and so on. (It’s actions all the way down.) This is the general trend. Going meta, through RL. Optimize the optimizer, learn gradient descent by gradient descent.

gwern 10y ago
> learn gradient descent by gradient descent. As far as 'learning gradient descent by gradient descent' goes, I think it's still an open question if a tiny RNN actually can improve meaningfully over ADAM etc :) I don't recall the paper showing any wallclock times, which suggested to me that the RNN was way slower even if it trained faster. The individual parameter adjustments are so low in the hierarchy of actions that the value may be minimal compared to higher up like architecture design.
- AlexCoventry 10y ago
  What paper are you referencing?
  - detaro 10y ago
    I assume this https://arxiv.org/abs/1606.04474 https://arxiv.org/abs/1606.04474 (since the title is basically the quoted text)
skybrian 10y ago
I'm not sure that trawling Google Images counts as active. It doesn't seem like it can be any more powerful than being given the entire Google Images database in advance.