-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathCOS_583_project_structure.txt
100 lines (57 loc) · 2.12 KB
/
COS_583_project_structure.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
COS 583 Project Structure
main application gui....
enables: choosing input data set
visualizes: data set in batches, input and output.
enables: changing/setting hyper parameters and network shape
visualizes: view of network
enables: training of network
visualizes: visualizing backdrop.
visualizes: choosing different functions
There modes then
Input network training and viz
class neural network
constructor takes layers list and function…
can learn and train to some error? or for some iterations….
can run iteration
can return immutable model that returns output for some input
can be fed training data
can return model that you can show bprop on...
might need to return errors per layer to show layer by layer backdrop
Interface function. sigmoid…..
Visual net? network particularly for backprop?
Update notes with slide info.
helper objects like layer??
helper for maybe testing????
mike nielsen website in bookmark very helpful
python code too.
Citations/Resources:
http://neuralnetworksanddeeplearning.com/chap1.html
Java swing.
java to json for visualization…
so add code maybe in dialog or whatever to build network…. change data information to training information, ability to learn input, percentage, viz from son fie
test for errors
add tooltips
push neural net project to git
check for mismatch between nn and input data
Presentation…..
Show that neural net can learn and help others learn b comparing iteration etc…
characteristics:…
underlying network package distinct from application,
another app can use the package for all this, decoupled.
many features can be tuned and can see effect on accuracy and user provided input
can incrementally train a network
example;
TestDigit params….. layers: 784, 30, 10
epochs = 25. eta = 3.0 mbsize 20
Half Adder layers: 2 4 4 2
epochs = 3000. eta = 10. mbsize 4
Implication layers: 2 4 4 2 / 1
epochs: 2500. eta = 10 mbsize 4s
Sine: relu vs sigmoid?
2 4 4 2
epochs: 2000. eta = 5 mbsize = 400
future work…..
more functions
multiple networks side by side
visualization, at least showing steps of backdrop….
challenges…