Hi:
Hi, Eric. 95% and in 5 mins, That's pretty good if just with a SP +
KNN. Did you turn the SP learning on? or just turn off and use it to
get the SDR. Recently, i changed the training order - the order of
training sequences. before that, I just followed the order of MNIST
dataset. Now the training sequences are from 0 to 9. I got 97% on a
small dataset (6000/1000). I am still wondering if there are some
errors in my code which I just added, because all of the errors happen
to input 0 and 1. So, now i am trying to train the full size dataset
with SP learning on and classifier learning on to see what will
happen. With the current parameters, training is really slow. I am
still waiting for the result.
Actually, I'm thinking about using other non-python engine. I am using
the nupic python currently.
The world record 0.23 is using deep CNN as the basic unit. Then using
them as column structures...Anyway, the result is applaudable.
What about the parameters which you used for the SP?
Thank you.
An Qi
Tokyo University of Agriculture and Technology - Nakagawa Laboratory
2-24-16 Naka-cho, Koganei-shi, Tokyo 184-8588
[email protected]
On Sun, 18 Jan 2015 12:05:42 -0500
Eric Laukien <[email protected]> wrote:
Hmm, I got 95% after five minutes. Perhaps you can try CHTM on this
task.
Especially with the latest spatial pooler, which is far more
efficient than
what I had before, it should work very well.
Also make sure to use multiple layers, and make the NN pass through
all of
them.
The world record is held by Yann Lecun using "sparse codes" which
are SDRs.
So the approach definitely works.
Regards,
Eric Laukien
On Sun, Jan 18, 2015 at 5:06 AM, Fergal Byrne
<[email protected]>
wrote:
Hi An Qi,
That's very interesting, thanks for sharing. Could you place your
code and
setup on Github so we can take a look at exactly what you did? It
seems at
first glance that you're just using SP, which is perhaps the least
powerful
part of HTM. I think a saccading system which also does Temporal
Pooling
(which we haven't quite got yet) would be able to do a much better
job on
this kind of task, but 89.6% is still a very good start for a plain
SP-based approach.
Very well done! Hopefully we can help you do even better.
Regards,
Fergal Byrne
On Sun, Jan 18, 2015 at 7:00 AM, <[email protected]> wrote:
Hello.
Sorry for the last email. Thx to the rich formatting :( ... I have
to
type again.
Recently, I got the result of the test. I followed the source code
and
built the Spatial Pooler + KNN classifier. Then I extracted images
from
MNIST dataset(Train/test : 60000/10000) and parsed them to the
model. I
tried to test with different parameters (using small dataset:
Train/Test -
6000/1000 ), the best recognition result is about 87.6%. After that,
i
tried the full size MNIST dataset, the result is 89.6%. Currently,
this is
the best result I got.
Here is the statistics. It shows the error counts for each digits.
the
Row presents the input digit. the column presents the recognition
result.
Most of the "7" are recognized as "9". It seems the SDR from SP is
still
not good enough for the classifier.
I found some interesting things. When I let the "inputDimensions"
and
"columnDimensions" be "784" and "1024", the result will be around
68%. If i
use "(28,28)","(32,32)" and keep others the same, the result will be
around
82%. That 's a lot of difference. It seems the array shape will
effect SP a
lot.
Did any one get a better result? Does any one have some suggestion
about
the parameters or others?
Thank you.
An Qi
Tokyo University of Agriculture and Technology - Nakagawa Laboratory
2-24-16 Naka-cho, Koganei-shi, Tokyo 184-8588
[email protected]
--
Fergal Byrne, Brenter IT
http://inbits.com - Better Living through Thoughtful Technology
http://ie.linkedin.com/in/fergbyrne/ -
https://github.com/fergalbyrne
Founder of Clortex: HTM in Clojure -
https://github.com/nupic-community/clortex
Author, Real Machine Intelligence with Clortex and NuPIC
Read for free or buy the book at
https://leanpub.com/realsmartmachines
Speaking on Clortex and HTM/CLA at euroClojure Krakow, June 2014:
http://euroclojure.com/2014/
and at LambdaJam Chicago, July 2014: http://www.lambdajam.com
e:[email protected] t:+353 83 4214179
Join the quest for Machine Intelligence at http://numenta.org
Formerly of Adnet [email protected] http://www.adnet.ie
