1. 9 "Modeling the Neo-Cortex"

2. no

3. no

4. 3

5. 5

6. 0

This project report was very poorly written. With constant run-on
sentences, disorganized structure, and many grammatical errors per
sentence, it was a struggle to read. While there was an overall
structure, it seemed as though bits and pieces were out of place. For
example, there was no evaluation section, but instead all of the
evaluation was crammed into the results section, and no clear
distinction was made between phase one and phase two results.

The overall idea of the project, modeling a neural network after the
neo-cortex, fits well into the framework of developmental robotics.
Unfortunately, the implementation does not. First of all, the fact
that features for objects were hand labeled by a person rather than
derived from actual data violates the grounding principle of
developmental robotics: the data was not grounded in the robot's own
sensory perception, but that of an external observer. Also, hand
labeling the features defeats the idea of the project. The purpose of
the neo-cortex is to take in raw sensory data and output the features
from the data, but since this was done by humans, the algorithms
don't have a purpose.

The best thing about this project is its idea. Modeling an algorithm
off the way the neo-cortex works is very interesting. Hawkin's model
of the cortex provides some useful insight into how the brain works,
but as of yet, no one has gotten a very good working model of it. If
this project could deliver on that, it would be a very useful
contribution to science.

This project didn't even come close to delivering on that. The worst
thing about this project is the implementation. The methods used to
implement this project lack real foundation in developmental robotics.
The phase one neural network, since it was all just linear
combinations of activations, didn't provide useful results.

The methods in this project line up with what was proposed in the
initial proposal. The results and contributions, as I predicted in
my initial review, are useless, just like the initial propsal 
indicated they would be.

There are quite a few details left out. First of all, HOW DO YOU TAKE
THE DFT OF THE CENTROID OF AN OBJECT?! I would love to know how...
Also, details of the evaluation of phase two are almost non-existent,
as well the accuracy for phase two isn't even reported. How was the
evaluation performed? What was the training and test set? There is no
related work section, which would have been very helpful to have.

The results in this paper do not demonstrate success; they are just
noise. The results from phase one indicate, as expected, that when
the update rules for weights for each neuron don't take into account
the firing rates of the other neurons, all the neurons eventaully
converge to the same output. When they do take that into account,
they start to have different outputs, but the outputs are meaningless.
There is no relation between objects with similar outputs. Also, only
38 datapoints were used, as well as in one test only 10 datapoints,
that were randomly generated, were used. This is no where near enough
datapoints to train a neural network or provide meaningful results.
Finally, randomly generating data points is not a valid method for
acquiring data. These networks are supposed to discover structure in
the data, but by definition, there is no structure in random data!

The results for phase two are equally as meaningless. Of what little
is reported, the ROC is reported as being >0.6, which the authors
claim indicates effective learning. It, in fact, does not. An ROC of
0.5 is chance, so being greater than 0.6 is not very informative.
Looking at the provided confusion matrix, it would seem as though
the results were random chance. In fact, calculating the accuracy
from the confusion matrix yields an accuracy of 0%. According the the
matrix, it got none of the classifications correct. This is not
unexpected, as the authors are attempting to get a multilayer
perceptron to classify a small set of objects using hand labeled
binary features. 

There is some hope though. The initial idea was very promising. If
the authors wish to improve this project, the first thing that needs
to be done is a more thorough literature search. Modeling the neo-
cortex with neural networks has been attempted several times, some
with success. Hawkins even attempted to implement an algorithm based
off the ideas in his book. As well, the authors might consider
looking into Deep Belief Networks, they have shown much promise and
are very similar to what the authors are trying to do.

This paper, in its current form, is not publishable. If it were
majorly overhauled, then it might be able to be published, but
as it currently stands, it isn't.