Hi Quinn,

You might find useful a couple of papers that were written about doing
handwritten digit recognition with the previous (pre-CLA) version of HTM:

http://www.ict.griffith.edu.au/~johnt/publications/AI2008.pdf
http://bias.csr.unibo.it/maltoni/HTM_TR_v1.0.pdf

My thought about your question #3 is that what really potentially sets CLA
apart from other solutions for visual recognition problems is the temporal
pooler, which allows it to hierarchically learn invariant representations
based on the statistics of how features occur sequentially in time. This
would require training the network with sequences of moving images (ie.,
animations of the digits zig-zagging across the input field), as was done
in the above papers. There is growing evidence that the brain's visual
system works a lot like this.

Without the temporal pooler, CLA is similar to a Kohonen network or vector
quantization, standard techniques that would probably learn to discriminate
the digits to some extent, but because they wouldn't have any position or
scale invariance they would basically have to learn every possible look
(and position within the input field) that each digit could have, which
would require a very long training period and many columns.

However, while digit recognition was successfully modeled with the original
version of HTM, that doesn't seem to be the case with CLA yet. I asked some
questions about using CLA for vision last month in a thread that starts
here:

http://lists.numenta.org/pipermail/nupic_lists.numenta.org/2013-June/000199.html

Jeff said that Numenta had gotten a single layer vision experiment working
well, but were just starting to go on to multi-layer experiments when they
stopped working with vision. CLA works on the same basic principles as the
original version of HTM, so it should be possible. But in my own
experiments I've been finding it tricky to get the invariant
representations working successfully. I've had to put that experiment aside
for the summer to work on other things but I'd be happy to let you know if
I get it working, and would love to hear how your experiment turns out.

-Mike

_____________
Michael Ferrier
Department of Cognitive, Linguistic and Psychological Sciences, Brown
University
[email protected]


On Sat, Jul 13, 2013 at 10:35 PM, Ian Danforth <[email protected]>wrote:

> Can't answer most of your questions but I did want to say the diagram is
> great :)
>
> As for question 3 the answers are Yes and Yes.
>
> I can also tell you that # of columns, % active, and perm inc and dec,
> were always being tweaked. There was no one 'right answer.' That lead to an
> automated outer loop to explore the space of possibile parameters. I
> believe Numenta intends to make this outer-loop available as part of the
> open source NuPIC code eventually, but you'll probably implement one
> yourself before then. :)
>
> I'd love to hear more about how you progress, keep us up to date!
>
> Ian
>
>
> On Sat, Jul 13, 2013 at 4:27 PM, Quinn Liu <[email protected]> wrote:
>
>> Hi everyone,
>>
>>     My name is Quinn and I am a CS student at Virginia Tech. While trying
>> to understand the CLA for the past year I built an object-oriented model of
>> it view-able here:
>> https://github.com/quinnliu/WalnutiQ/tree/master/WalnutiQ
>>
>> I am now trying to train the model on the popular MNIST handwritten digit
>> data set to see how the CLA compares to other algorithms. However, before I
>> do so I was hoping I can get advice on how to structure the model to best
>> learn and classify the handwritten digits.
>>
>> In hopes of making my questions easier to answer I drew up a diagram and
>> the questions can be viewed here:
>> http://www.walnutiq.com/#!questions-about-cla/cmr
>>
>> Thank you very much for your time and congrats with all the success with
>> Grok!
>>
>> Best Regards,
>>
>> Quinn Liu
>>
>> www.walnutiq.com
>>
>> [email protected]
>>
>> _______________________________________________
>> 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
>
>
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to