Hi Neal,

I think I understand your frustration.  I hope my perspective will help this
conversation a bit and help you and others decide whether NuPIC and the CLA
are worth your time.

 

The NuPIC open source project and the HTM/CLA technology are unique.  You
should care about NuPIC if the following two criteria apply to you.

 

1) You are interested in Machine Intelligence

2) You believe the best way to achieve machine intelligence is by employing
the principles used by the neocortex

 

Our goal is to make intelligent machines, machines that learn, that interact
with their world, and have goals. Machine intelligence and machine learning
are not the same thing.  There are many problems that can be solved well
with existing machine learning techniques and new ones are being created all
the time.  But most machine learning techniques are not on a path to machine
intelligence.  They are derived to solve particular problems and any
technique that solves the problem or does best on a benchmark is good.
Machine intelligence is about building universal learning machines that can
solve almost any problem via learning and behavior.  I measure our progress
in machine intelligence not just by what the CLA can do today but by how
many of the learning/behavior principles in the neocortex we understand and
employ.

 

It is obvious to me that we must understand how the neocortex works to build
intelligent machines.  However this is not a majority view.  I constantly
run into researchers who dismiss the need to understand brains or say it is
impossible.  In fact I believe the biggest obstacle to achieving machine
intelligence is getting more people to embrace brain theory as the path
forward.  I wrote On Intelligence to make this case; that a) we can
understand how the neocortex works and b) this is the path to intelligent
machines.  The good news is this view is gaining momentum.  DARPA, iARPA,
IBM, and others are funding projects based largely on the arguments I have
put forward and the progress we have made so far with the CLA.  It is too
early to say we have succeeded but we are definitely making progress getting
people to embrace brain theory as a necessary component of a program to
build intelligent machines.  The CLA is the best, and arguably the only,
example of a learning methodology based on cortical principles that fits
within a framework for machine intelligence.

 

Of course I too would like to have more success stories today.  When we
started Numenta I thought we could be where we are today in four years.  It
has been close to nine.  Still, I am thrilled with our progress.  The
technology and theory are progressing and from my vantage point I see
acceleration.  This past month I made good theoretical progress on temporal
pooling and vision invariance (I will write this up  soon).  Progress in
theory doesn’t translate quickly into products but I know it will eventually
so it is exciting.

 

Still, it is important to demonstrate progress on real world problems too.
Today, Grok is the primary example of the CLA enabling a cool new product.
We hope to make a big business based on Grok but that will take time so we
can’t yet point to it as a success. Another exciting application I see for
the CLA is based on CEPT’s work in word SDRs.  I was blown away by Subutai
Ahmad’s hackathon hack feeding three word sentences into the CLA.  (There
were other cool demos but this one struck me as the one with the most
immediate business potential.)  Subutai’s hack
<http://numenta.org/blog/2013/11/06/2013-fall-hackathon-outcome.html#fox>
looked simple but what was happening under the covers was impressive.  There
are several great businesses lurking there.  But as always, it takes time,
money, and foresight to seize those business opportunities.  I don’t know if
anyone in the NuPIC community has what it takes to do that, but CEPT’s work
and the CLA has enabled fundamentally new possibilities in language
understanding.  There are no doubt other applications NuPICers are working
on that we don’t know about.  Again, achieving business success with the CLA
is important but it has to play a supporting role in the overall mission. 

 

Your point about the code not exactly matching the theory is understandable.
Implementing the theory in SW is challenging.  We constantly struggle with
how to make the code work in real time within the constraints of our
computing environments and how to test it to know that it works, etc.  All I
can say is doing significant things is hard.  The first computers were far
from Turing's "universal machine".  

 

In summary, NuPIC  isn’t for everyone.  The CLA isn’t the most amazing
machine learning method ever invented and we have many theoretical problems
yet to solve.  But the CLA is the best learning model that fits within a
framework for biological and machine intelligence and even today it has many
unique capabilities.  If you subscribe to the two principles above I don’t
know of a better place to be.

 

Jeff

 

 

From: nupic [mailto:[email protected]] On Behalf Of Pedro
Tabacof
Sent: Friday, January 10, 2014 4:24 AM
To: NuPIC general mailing list.
Subject: Re: [nupic-discuss] Any results, anywhere?

 

Hello Neal,

 

I understand your frustration. With so much buzz around deep learning, big
data, etc, on the mainstream machine learning community, it's hard to work
with such a different approach. I see this a real scientific research, and
there is always big risks with projects as such, but if we succeed it will
be a real AI progress. Note also that there are many interesting brain
features that are yet to be incorporated into the CLA (such as motor action,
hierarchy, feedback). 

 

I don't know if you read this thread, but I applied the CLA to a real-world
dataset that was used on a serious competition, achieving an error that
would put me around the third place (note that it's a somewhat unfair
comparison since I had the test data at my disposal):

 

http://comments.gmane.org/gmane.comp.ai.nupic/1047

 

Considering it was my first time using the new NuPIC and I didn't spend a
lot of time on it, I consider it to be a good, even surprising, result. I
believe we will see much more interesting applications later on, we either
have to be patient or actually work on them. This NuPIC version is really
new, I even had difficulties getting it running, so we're not talking about
a mature technology (even though the algorithm is a few years old), but
rather a very young one.

 

Pedro.

 

 

On Fri, Jan 10, 2014 at 4:52 AM, Ian Danforth <[email protected]>
wrote:

Neal,

 

 Which application domain are you most interested in? Remember that NuPIC
has been used almost exclusively internally by Grok/Numenta until very
recently, and so on a limited set of tasks around prediction and anomaly
detection the technology are much more mature. In most other areas there
haven't been more than tech demos implemented. The kind of extensive use,
methodology and rigor it sounds like you're looking for probably won't
happen until the code base is a bit cleaner and easier to use. It would be
useful to know what you expected NuPIC to be 'doing' so the community can
provide entry points and demonstrations that serve the audience you
represent.

 

Ian

 

On Thu, Jan 9, 2014 at 7:50 PM, Neal Donnelly <[email protected]> wrote:

Hey everyone - 

 

I got involved with NuPIC with starry eyes after reading the CLA whitepaper.
The abstractions that it describes are exciting and the level of work that
was obvious in the codebase were very compelling. I was mildly concerned
that there were no results presented, but I was told that was because there
are no established metrics for measuring its performance since it's so
novel.

 

Now, after mucking around in the network engine codebase for a while, I've
realized that the abstractions presented in the CLA whitepaper seem to have
little bearing on the implementation. Spatial and temporal pooling are
accomplished as separate types of regions, which are composed not of cells
but of Nodes, which are themselves comprised of elements...?

 

This confusion has left me wondering why I believe in this project if
neither theory nor results back up the implementation. This line of thinking
has left me frustrated that I can't find a single result of NuPIC actually
doing anything. When none of the existing benchmarks are fitting,
researchers invent a new one. I realize I've seen no learning curves, no
applications to real data, no demonstrations of performance of any kind.

 

My hope is that this will provoke a flood of links and papers that I missed.
My fear is that I've been terribly naive to assume that NuPIC would work
when there aren't results out front and center.

 

Thanks. 

Neal Donnelly

 

_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

 


_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org





 

-- 
Pedro Tabacof,
Unicamp - Eng. de Computação 08.

_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to