catastrophic interference

[Sheldon Tetewsky]

Neil Burgess recently wrote that
>studies of catastrophic interference in BP networks are interesting when
>considering such a network as a model of some human (or animal) memory
>system.
However, he also questioned whether or not there was
>..any reason for doing that.

In the Tetewsky, Shultz, and Buckingham study ("Assessing interference and
savings in connectionist models of a sequential recognition memory task"),
referenced in Bob French's recent posting, we did some simulations and 
experiments indicating that neural networks can provide a good account of
human performance in a simple recognition memory task when memory is
assessed in terms of savings scores.  The paper is in progress and will
be announced soon. In the meantime, I'll just mention a few relevant
details.

Our work was motivated by Scott Fahlman's idea that his cascade-correlation
algorithm (CC) has certain inherent design features that should give it an
advantage over BP when it comes to dealing with the problem of catastrophic
interference.  We tested this idea by using an encoder version of CC to
model recognition memory.  Results indicated that in contrast to previous
findings, retroactive interference is not that serious a problem for BP
when memory is measured in terms of savings scores. However, CC also
produced a significant increase in savings, relative to BP.  Aside from
this difference in the magnitude of savings, we also have evidence that CC
was better than BP at accounting for the relative number of trials that
subjects used in the different phases of learning.

-- Sheldon Tetewsky