Enabling Continual Learning in Neural Networks
This is an article about a month ago from DeepMind. It does seem that we are getting a bit closer to being able to use artificial neural networks (ANN) to replicate the learning that takes place in the human brain.
And I think that is where the problem lies.
We are trying to replicate the human brain, but is the human brain the best model for learning?
An example. My eight-year-old son and I both watch the same Star Wars movie together. Yet he can recall a lot more details (the scenes, the people, even the music) about the movie compared to what I can recall.
If what DeepMind is doing is a hint, it may mean that our brains are also trying to optimize the limited synapses (although the number of synapses in the brain is way beyond what any ANN has now) in the brain. Which means that once we have used our brains to learn a certain amount of things, any additional learning is going to mean readjusting the connections (weakening our memories in some areas) to learn those new things.
Going down this road, it suggests that making machines think and learn like humans may not be optimal, as the way our brains learn has a serious flaw--it needs to weaken our ability to recall what we have already learnt, in order for us to learn new things.
If only there is some other way to learn things without having to forget what we have learnt. Well, for our mortal brains, that may not be an option, but it is something that we can try in our attempts at giving intelligence to machines. And I think that is the Holy Grail of artificial intelligence--to arrive at a model for machine learning that surpasses the human brain.
#DeepMind #AI
This is an article about a month ago from DeepMind. It does seem that we are getting a bit closer to being able to use artificial neural networks (ANN) to replicate the learning that takes place in the human brain.
And I think that is where the problem lies.
We are trying to replicate the human brain, but is the human brain the best model for learning?
An example. My eight-year-old son and I both watch the same Star Wars movie together. Yet he can recall a lot more details (the scenes, the people, even the music) about the movie compared to what I can recall.
If what DeepMind is doing is a hint, it may mean that our brains are also trying to optimize the limited synapses (although the number of synapses in the brain is way beyond what any ANN has now) in the brain. Which means that once we have used our brains to learn a certain amount of things, any additional learning is going to mean readjusting the connections (weakening our memories in some areas) to learn those new things.
Going down this road, it suggests that making machines think and learn like humans may not be optimal, as the way our brains learn has a serious flaw--it needs to weaken our ability to recall what we have already learnt, in order for us to learn new things.
If only there is some other way to learn things without having to forget what we have learnt. Well, for our mortal brains, that may not be an option, but it is something that we can try in our attempts at giving intelligence to machines. And I think that is the Holy Grail of artificial intelligence--to arrive at a model for machine learning that surpasses the human brain.
#DeepMind #AI
1 comment:
Right after I posted this article, I came upon this:
AI Learns Gender and Racial Biases from Language
http://spectrum.ieee.org/tech-talk/robotics/artificial-intelligence/ai-learns-gender-and-racial-biases-from-language
I think at the end of the day, if we based our AI on humans, it will just be good replicas of human beings, but nothing more.
Post a Comment