Comparing various architectures (fully-connected and multiple convolutional variants) on the USPS dataset for the optical recognition task. Currently, four different weight initialization presets are available for training the models: effective, too slow, too fast, and default. In addition to weight initialization schemes, various learning rates have been tested and categorized in a similar manner (effective, too slow, too fast). Please refer to the materials folder for a more detailed write-up.
python3 study.py --net 1 # fully-connected net
python3 study.py --net 2 # locally-connected CNN
python3 study.py --net 3 # fully-connected CNN
python3 study.py --net 1 --init 1 # effective learning
python3 study.py --net 1 --init 2 # fast learning
python3 study.py --net 1 --init 3 # slow learning
- add visualization functionality for kernel filters and activations for each layer
- add options for dropout, momentum, and bagging