todolist.txt

TODO List:
-> Enable running for gray and color images
-> Check; It is going to work for new learning algorithms
-> Download Features Learned by steven's algorithm and test the accuracy
_________________ Approach #1
-> Implement Uniform DeSTIN by replacing nodes of layer N by a typical node
    -> the "Typical-Node, TN" will have to do learning over every possible input from the N-1th layer
    -> if N=1, the typical layer#1 node will run over every possible overlapping 4by4 image patch, then ...
    -> the centroids learned by the TN will be shared by every node of the layer (Uniform DeSTIN Achieved)
    -> (if the node by chance suffered from improper initialization or so, a problem might happen)
_________________ Approach #2
-> Iterate over every possible shifted versions of the image and store the beliefs and average them
-> Requires a sound way of taking uniform centroids for the layer
-> Trains the whole network over a single image
_________________ Approach #3
For Every Image example
    -> Train Layer1 TN as in approach#1
    -> assign layer#1 nodes centroids learned from TN
    -> calculate beliefs of the layer#1 nodes
    -> then, do the same thing for layer#2 and so on



-> Revise the correctness of the implementation, readability, modularity
->