task consolidation

[Christian Omlin]

Lisa Meeden writes:

>To investigate the efficacy of protoplans, a transfer of learning
>experiment was done.  The protoplans learned in one controller network
>were saved in an associative memory and used to guide a second
>controller network as it learned the same task from scratch.  The
>associative memory mapped the precursor sensor states of a protoplan
>to the protoplan itself.  Controllers trained with protoplans instead
>of goals as input converged more quickly on good solutions than the
>original controllers trained with goals.  Protoplans were able to
>guide the robot's behavior by marking the important moments in the
>interaction with the environment when a switch in behavior should
>occur.  This kind of timing information was indirect--no specific
>action was indicated--but knowing when to change from a particular
>strategy to a new one can be very important information.

This is similar to work done on training (recurrent) networks with
prior knowledge. We have investigated algorithms for the extraction
and insertion of symbolic knowledge in recurrent networks trained
on temporal learning tasks. For a testbed, we learned regular grammars.
We have shown how partial prior knowledge about a regular grammar
can be encoded in a fully-recurrent neural network with second-order
weights. The improvement of convergence time is `proportional' to the
amount of prior knowledge. A description of the learned grammar can
also be extracted from networks in the form of deterministic finite-state
automata (DFAs). We have shown that the extracted DFAs outperform 
the trained networks, i.e. the DFA correctly classifies more strings
than the trained network itself.

The details can be found in the following book which has recently
been published:

@INCOLLECTION{omlin94b,
    AUTHOR = "C.W. Omlin and C.L. Giles",
    TITLE = "Extraction and insertion of symbolic information in
             recurrent neural networks",
    EDITOR = "V. Honavar and L. Uhr",
    BOOKTITLE = "Artificial Intelligence and Neural Networks:
                 Steps toward Principled Integration",
    YEAR = "1994",
    PUBLISHER = "Academic Press",
    ADDRESS = "San Diego, CA",
    PAGES = "271-299"}