home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!rosie!next.com
- From: paulking@next.com (Paul King)
- Newsgroups: comp.ai.neural-nets
- Subject: neural nets and generalization (was Why not trees?)
- Message-ID: <4458@rosie.NeXT.COM>
- Date: 24 Jul 92 02:45:18 GMT
- References: <arms.711643374@spedden>
- Sender: news@NeXT.COM
- Lines: 43
-
- arms@cs.UAlberta.CA (Bill Armstrong) writes:
- > The question is: What is the advantage of making connections in a
- > non-tree fashion?
-
- sef@sef1.slisp.cs.cmu.edu (Scott) writes:
- > 2. Why try to minimize the number of independent free connections in
- > the net?
- >
- > Training is indeed faster with more connections, but generalization
- > is better with fewer connections. The extreme case is seen in various
- > image processing tasks (e.g. char recognition) in which it is almost
- > mandatory to use shared connections or some sort of bottleneck -- else
- > you would need many more training examples than pixels in the image.
-
- Here is an additional observation on the issue of generalization.
-
- The collection of training patterns for a neural net could be viewed as:
-
- possible input patterns --> [Black Box] --> correct output patterns
-
- Neural networks seem to operate on the assumption that what goes on
- in the black box is rational and deterministic -- that a causal chain
- of events transform an input pattern into an output pattern. The
- "goal" of the neural net is not only to memorize the input-to-output
- mappings, but also to reverse engineer the causal chain of intermediate
- states that presumably exists inside the black box.
-
- Hidden units are interesting because they occasionally converge
- on an intermediate representation that highlights regularities
- in the behavior of the black box. These intermediate states can
- then be used as inferred information about internal states of the
- black box, which can be useful as inputs to further networks.
-
- As an example, consider handwriting recognition. Mapping input
- drawings to target characters might result in hidden-units that
- correlate with subfeatures in the input drawings. This ability
- to map input drawings to subfeatures might then benefit a higher-level
- system that maps subfeatures to words, bypassing the character
- representation altogether. One now has a network that can recognize
- words even if certain letters are incompletely written.
-
- Paul King
- paul_king@next.com
-