Are Deep Neural Networks Dramatically Overfitted? (from this blog)

• I would like to discuss a couple of papers on generalizability and complexity measurement of deep learning models in the post. Hopefully, it could shed light on your thinking path towards the understanding of why DNN can generalize.

Work for the next week:

• I need to build a set of panalyzers that are room/channel specific. Done!
  • It looks like it’s just creating a key in the channel loop that also has the split information:

    max = len(split_list)-1
    for i in range(0, max):
        for db in db_list:
            name_list = self.run_query('select username as new, id as old from user_view', db_name=db)
            for group in channel_list:
                key = "splt-chn_({}of{})-{}".format(i+1, max, group)

• Comparing the counts of place and space terms will (hopefully!) show the movement of the channel population between splits (splits should be similar across channels), and the actions (spaces) that are explored in each space.
  • Table that shows place/space term counts by channel and split
  • I think I can get this by just using the top-n common terms to identify the rooms, and the top-n (excluding room terms) for each splt-chn panalyzer. These room terms are the identifier (row label), and then each column is a different group, which gives us a set of 4 room nodes from which link to action nodes. Take that and output to graph?
  • Refining the table and graph can be done by finding embedding neighbors.
• Statistical difference in the populations of place terms within each split and within each channel.
• What is the math that we use to show spaces? Or do we just show that they are different from the place populations?
• Build a map from the table

Well that’s a nice way to start the day:

----------------------
split_1_of_4:
subject BOW
tymora1: 3948
tymora2: 732
tymora3: 506
Group: 63
1: 63
post BOW
goblin: 287
room: 239
orc: 227
stairs: 214
vines: 197
hit: 186
fire: 171
arrow: 170
gate: 134
behind: 133
----------------------
split_2_of_4:
subject BOW
tymora1: 3972
tymora2: 732
tymora3: 376
Group: 83
1: 83
post BOW
rope: 415
gate: 331
room: 324
orb: 309
statues: 251
across: 207
side: 196
around: 162
pit: 150
feet: 148
----------------------
split_3_of_4:
subject BOW
tymora1: 3004
tymora2: 816
tymora3: 1016
Group: 57
1: 57
post BOW
troll: 618
grogg: 596
chest: 274
hand: 252
room: 244
open: 209
gate: 199
club: 191
want: 169
key: 143
----------------------
split_4_of_4:
subject BOW
tymora1: 596
tymora2: 162
tymora3: 438
Group: 67
1: 67
post BOW
dragon: 176
coins: 174
room: 148
barrier: 108
light: 102
something: 102
platform: 100
eyes: 82
woman: 82
treasure: 62

Need to fix the “Group” “1” error and add an underscore to subjects in panalyzers – done

Note. This means that you left a paren on your callback: File "C:/Development/AntibubblesDungeon/src/analytics/SequenceAnalyzerGUI.py", line 438, in show_bow
self.main_text.delete('1.0', END)
AttributeError: 'NoneType' object has no attribute 'delete'

Added the beginnings of the table