While deep learning seems to be the answer
While deep learning (or other types of machine learning) seems to be the answer to most of the hard problems in computer science like computer vision, prominent figures of the field admit that advancing in the area of general intelligence will require entirely different methods.
Geoffrey Hinton says that he is "deeply suspicious" about back-propagation, the technique used to train neural networks, and that "My view is throw it all away and start again" .
Yoshua Bengio, in a conference presentation called "Creating Human Level AI", says that we are "still far away".
"All industrial successes are based on pure supervised learning"
"Still learning supeficial clues that do not generalize well outside of training contexts and make it easy to fool trained networks: Current models cheat by picking on surface regularities, e.g. , background greenery implies that probably an animal is present "
"Still unable to do a good job of learning higher-level abstractions at multiple time scales, deal with very long-term dependencies"
In his talk at Cognitive Computational Neuroscience 2017 Yann LeCun agrees with Josh Tenenbaum that "All of these AI systems we see, none of them is 'real' AI" because
"The brain learns with an efficiency that none of our machine learning methods can match"
François Chollet, the creator of the Keras machine learning framework, says that"deep learning is not the road to artificial general intelligence". Instead it could be to create algorithms that:
"Learn subroutines reusable across diverse subproblems"
In this paper by Kevin Ellis, Lucas Morales, Mathias Sable Meyer, Armando Solar-Lezama, Joshua B. Tenenbaum an algorithm of learning is described which is inspired by the way humans are writing computer programs.
Geoffrey Hinton says that he is "deeply suspicious" about back-propagation, the technique used to train neural networks, and that "My view is throw it all away and start again" .
Yoshua Bengio, in a conference presentation called "Creating Human Level AI", says that we are "still far away".
"All industrial successes are based on pure supervised learning"
"Still learning supeficial clues that do not generalize well outside of training contexts and make it easy to fool trained networks: Current models cheat by picking on surface regularities, e.g. , background greenery implies that probably an animal is present "
"Still unable to do a good job of learning higher-level abstractions at multiple time scales, deal with very long-term dependencies"
In his talk at Cognitive Computational Neuroscience 2017 Yann LeCun agrees with Josh Tenenbaum that "All of these AI systems we see, none of them is 'real' AI" because
"The brain learns with an efficiency that none of our machine learning methods can match"
François Chollet, the creator of the Keras machine learning framework, says that"deep learning is not the road to artificial general intelligence". Instead it could be to create algorithms that:
"Learn subroutines reusable across diverse subproblems"
In this paper by Kevin Ellis, Lucas Morales, Mathias Sable Meyer, Armando Solar-Lezama, Joshua B. Tenenbaum an algorithm of learning is described which is inspired by the way humans are writing computer programs.
Comentarii
Trimiteți un comentariu