3 ms·
Now, about our wondrous machines and algorithms. This needs a different comment. >> Selman, a professor at Cornell University, said: “Computers are basically s
by YeGoblynQueenne 11y ago
Now, about our wondrous machines and algorithms. This needs a different comment.
>> Selman, a professor at Cornell University, said: “Computers are basically
starting to hear and see the way humans do,” thanks to advances in big data and
“deep learning”.
I'm a student of AI. I'm studying for an MSc in Intelligent Systems at the
University of Sussex in the UK. I come from a classical AI background, with
studies in discrete maths, logic and logic programming in particular. I've
studied machine learning and to a lesser extent neural networks. I'm familiar with deep learning.
This limited knowledge nevertheless gives me the stenght of conviction to say
that, no- the quoted comment is not a good representation of the state of the
art. AI systems are still far from perceiving the world in "the way humans do".
Machine learning is a powerful tool, but it has severe limitations. Deep
learning is certainly not magic and neural networks in general are "just"
optimisation (except we're optimising systems of functions). I really don't see
how any technique that we have at this point in time can be used to produce
machines with anything approaching human levels of competence in the complex
tasks that humans are good at.
Sure, we have machines that can drive cars or play Chess or Go competently.
However, those are very clearly delineated tasks. Chess and Go are games with
well-known rules in a restricted domain. Driving is also a clearly strictly
defined domain - for instance, you won't see Google car driving offroad; my
theory is that if there's no roads or signs to guide it it would not be able to
navigate on its own. I might be wrong on that.
The important thing is that although machine learning algorithms can be trained
on any conceivable task, and the same algorithm can often be trained on a large
variety of tasks, changing tasks requires new training. This training requires
resources that scale, very often exponentially so, with the complexity of the
task. Training a Chess-playing AI requires a lot of data and computational
power. Training a Go-playing AI requires a lot of different data and another
lenghty period of training. For a game where not enough data is available, it
would be difficult to obtain good results from machine learning. Machine
learning itself has not given the best results in competitive AI.
Humans on the other hand excel at generalising from very few examples and can
often pick up a new task with minimal training. For instance, from playing
football to playing basketball, or from playing Magic: the Gathering to playing
Poker. We excel in very different types of tasks: competitive tasks, collaborative tasks, games, survival, commerce, anything you can think of.
I don't see how machine learning will ever be able to accomplish such feats, no
matter how competent individual algorithms can become in specific tasks.