Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Neural net outperforms humans in pattern recognition

Author: Ratko V Tomic

Date: 15:12:47 10/01/99

Go up one level in this thread


That's interesting. Several years ago I was trying to use neural
nets for speech recognition (as a consulting job) and had abandoned
it due to poor performance (regular statistical matching recognized
better and much faster, although still poorly compared to human).
Similarly disapponting was neural network use for character
recognition, especially the for the handwriting. I think the key
underplayed point in this article was:

   In benchmark testing using just a few spoken words, USC's Berger-Liaw
   Neural Network Speaker Independent Speech Recognition System not
   only bested all existing computer speech recognition systems but
   outperformed the keenest human ears.

i.e. the phrase "few spoken words." That sheds the light how it was
possible. The net had a handful of choices to decide among the words
amid noise. While the researchers may have instructed humans to also
watch for the same handful of words, thus appearing to even the
conditions, the bulk of human recognition and perception occurs
within deeply subconscious (automatic) feedback loops in the lower
level sensory networks/filters. And these lower layers of sensory
tuning are not tunable at will (consciously) but are tuned by the
overall phoneme patterns a persons had learned. Thus even though
a human may be instructed for the test that only 5 words may occur,
human sensory perception will still be tuned for hundreds of
thousands of patterns they have learned in their entire life.

So this system seems to be one of those one trick dogs, outperforming
humans in a very narrow task (after all, calculator can outperform
everyone in arithmetic). Once the researchers move up to a system
with comparable vocabulary to the human one, they'll run into the same
problem everyone else did trying to teach neural nets to perform at the
human level -- the performance doesn't scale up, it does well on a small
problem, but as you scale the problem up, the network and the computation
grow exponentially.

In chess one could compare it to a computer being able to solve mates in
few moves better than any human.  But the search space grows exponentially
with the length of the sequence and at some size it is beyond the simple search
and requires more flexible strategies which are as yet not programmable.




This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.