Re: WHITE PAPER: Numenta's Hierarchical Temporal Memory (Jeff Hawkins and Dileep George)

From: Patrick Crenshaw (
Date: Mon May 22 2006 - 16:34:29 MDT

I've read just about everything Hawkins has written on the subject,
and while he is not trying to create a general intelligence (at least
that is not his stated goal) I think that an approach similar to his
would be doomed to success. He is essentially (though he came at it
from a different angle) taking the ideas of Judea Pearl and
incorporating time into the belief propagation nodes. I'm not sure
explicitly including time into the individual nodes is the best way to
do it, but it does appear that the brain does something along those
lines, so that is the path that he is following.

If anything, I am glad that I found the work of Judea Pearl through
Hawkins. Everyone on this list should read this book:

As for built-in structure, you should read the part of that book about
constructing trees or polytrees for a given set of input data.

On 5/22/06, Charles D Hixson <> wrote:
> Joshua Fox wrote:
> > Interesting white paper. Any comments on whether this will work?
> >
> > My (non-expert) impression, like neural networks, this lacks
> > sufficient built-in structure to achieve full general intelligence
> > (not that they claim it would), but that it may form a useful module,
> > which, along with other recent AI developments, another, more central,
> > module could put to good use.
> >
> > Joshua
> To me this doesn't appear to be an attempt to create a general AI, but
> rather only components that could be adapted to several different
> purposes, including an AI.
> It also doesn't appear to include sufficient detail to judge whether or
> not it would work. It's like many such papers in being designed to
> eventually allow claim to original discovery, without revealing enough
> to allow others to get there ahead of you. (No shame here. This is
> common practice in academic research, where priority is highly valued.)
> I find what it does say to be interesting, but not convincing. Much of
> it meshes with what I know or believe, in other places he uses language
> that it, to me, full of terms that are either undefined or incompletely
> defined. OTOH, I haven't even devoted much thought at all to visual
> (i.e., time varying 2D data) pattern recognition, which is where he
> appears to be placing most of his efforts.


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT