ECAL071 EX4 Grades
Click on the authors to see their work.
Click on the grade to toggle my comments on and off.
ronav
http://www.cs.bgu.ac.il/~ronav/ecal/ex4/
Very good work, and nice code. Good description of the
Neural Network structure and backpropagation algorithm.
Please read my comment about graphs below.
bonus-5
bonus-2
exp-alpha
exp-lr
no-err-per-N-historgam
graphs
not-so-good-graph-epoch
105
alexta, zhok
http://www.cs.bgu.ac.il/~zhok/work4.htm
Good report, brief but informative.
exp-sets
good-graph
100
mazory, liadi
http://www.liad.biz/ecal/ex4/
Your work is very good, yet your report is too brief - especially
the work description. I couldn't understand whether you've tested your
network on the entire dataset, or just a subset; I had to deduct some points on this.
The "N vs. Accuracy" graph is not clear too - see my comments about graphs and
reports below.
On the upside, the tabbed format of the report is very
convenient and easy to browse.
graphs
reports
bonus-5
no-err-vs-epoch-graph
100
einavbit, jasminme
http://www.cs.bgu.ac.il/~einavbit/ECAL071/Exe4/Exe4
Good report, good work, good discussion.
good-graph
not-so-good-graph-epoch
bonus-2
bonus-5
107
bas, benichay
http://www.cs.bgu.ac.il/~benichay/Ex4/
Your work and report are very good. Your report is a little brief
but coverts all the necassary information. Nice network topology figure.
good-graph
bonus-2
bonus-5
107
lisse, suissad
http://www.cs.bgu.ac.il/~suissad/ECAL071/ex4/doc4
Your work and report are very good, and nicely presented.
The discussion and conclusion part of your report is impressing.
good-graph
exp-sets
bonus-2
not-so-good-graph-epoch
102
abirs, yaakovog
http://www.cs.bgu.ac.il/~abirs/evo/ass4/
Very good work!
exp-sets
good-graph
exp-lr
bonus-2
not-so-good-graph-epoch
102
ilankad, wolfsonk
http://www.cs.bgu.ac.il/~wolfsonk/GA/ex4
Excellent work! very good analysis of the experiment results
and good discussion & conclusions.
bonus-2
good-graph
good-graph-epoch
exp-sets
102
eliyahuu, kertis
http://www.cs.bgu.ac.il/~eliyahuu/ass4/
Very good report, and good descriptions of the NN algorithm.
I especially liked the plot of error vs. ratio of neurons in the hidden layers -
it becomes very clear that high ratio causes bad results.
bonus-5
bonus-2
no-single-layer
fitness
train-test-separated
105
amitbenb, merhavi
http://www.cs.bgu.ac.il/~amitbenb/EvoAlgo/EvoAlgo4.htm
Excellent report. It took me some time to figure that
K is a constant instead of 1,000 (think about a network with 3K units
in each layer!). Your 3D graphs are visually appealing, but mind that
sometimes they may be misleading; for example, when the y value is 1
for all the cases, there's no need to display fractional value marks...
good-graph-epoch
graphs
no-single-layer
bonus-5
103
beckerli, ohanaar
http://www.cs.bgu.ac.il/~ohanaar/Ass4/NN.htm
Very good work and report.
The figure of the network architecture is nice.
bonus-2
good-graph
not-so-good-graph-epoch
102
meltzerh, zeharias
http://www.cs.bgu.ac.il/~zeharias/Assignment4/ass4.htm
Your work is very good and the report describes it well.
I liked the initial part when you state which experiments will be done
(main experiment, varying ratio, entire dataset etc).
Some of the "num of errors" graphs use fractional axes marks - this
is confusing and unecesarry - read my comment below.
bonus-2
bonus-5
graphs
exp-sets
train-test-separated
107
nirgi
http://www.cs.bgu.ac.il/~nigri/evol4task.htm
The work and report are very good. Excellent discussion and conclusions!
exp-sets
good-graph
good-graph-epoch
100
aharob, blesser
http://www.cs.bgu.ac.il/~aharob/Ex4/
Good work. Nice illustration of the neural network architecture and good description of the backpropagation algorithm.
exp-sets
bonus-2
bonus-5
107
Bonus for using two (or more) inner layers
+5
Bonus for using the entire dataset
+2
Nice experiments with the network performance when
modifying the alpha parameter
Nice experiments with the network performance when
modifying the learning rate parameter
Nice experiments with the network performance when
modifying the train / test set size ratios
Missing: Histogram of error per N
-2
Missing: Graph of network error / performance vs. epoch number
-2
You haven't done experiments with a single hidden layer.
The main goal of this assignment is to perform experiments with
single hidden layer neural network and analyze it's performance.
Testing a multiple hidden layer network is a bonus task, and doesn't
exempt you from testing the single hidden layer version.
-2
You talk about the "fitness" of a Neural Network -- In general,
the term "fitness" is used in evolutionary discussions; there is
no evolution (only learning - mind the difference!) in this experiment.
A more suitable term would be "performance"...
A general comment about graphs: even if the title describes
the axes meanings, give each axis a title - it makes the graphs
much more readable. Also, when information within a given
range is displayed within a graph, its axis should usually display
only the relevant range in the relevant units.
For example, percentage axes should be in the range [0,100] only;
an integer axis should not display fractions on its value marks etc.
Your reports should enable the reader to understand the process
of your work. It should indicate which experminents have you done, and
the settings of each experiment. If you've found encountered some problems
during the process of work, list them and tell how did you solve them.
Your work may achieve very good results, but the report should describe the
entire process, not only results.
Very good graphs of error vs. hidden layer size, where the
error axis if for both test and train sets; this is exactly what
we expected to see in this experiment.
Very good graphs of error vs. epoch number, where the
error axis if for both test and train sets; this is exactly what
we expected to see in this experiment.
The error vs. epoch graph could provide much more information
when it shows the test set errors as well as the train set errors.
To do so you should simply evaluate your network on the test set
every K epochs, where K can be a fairly large int (10, 100, 500, ...)
so the performance of the learning program stays almost uneffected.
Your graphs could be more informative if you'd chose
to display both train and test set results as two series in the same graph.
This method enables you to easily note the changes in your network performance
on both sets as a function of N or epoch number (depends on the graph type),
and the interaction between these two sets.