Robust-first Computing Wiki

Implement a convolutional neural network (CNN) that will be able to classify an input image. It will use the layers based around a very basic CNN including: convolution layer, pooling layer, rectifier layer, and a fully-connected layer. This will have to start off static in the MFM, as passing data would be very difficult otherwise. This will require five elements that I will go into detail further down this page.

• Neuron:
  • Will contain the weights from the pre-trained network that it will be using to classify the image
  • Deal with all of the computations for the convolution, pooling, normalization, and fully-connection
  • Will update the neighboring pixels based on the output from the computation
• Pixel
  • Holds the gray-scale value for 7 pixels, then has space for a couple more bits for flags in the future
• Router
  • Will hold routing information based on which way for a packet to go based on where it is going
• Path
  • Will help with guiding the packets to the specified Neuron cluster
• Packet
  • Will hold information about 7 pixels, and will follow the path and routers to its destination to deliver the information

• Main Goal: Run a simple MNIST example with smaller training set of 0's, of image size 18×18(?)
  • First Step: Get a layout that will work for the CNN (X)
  • Second Step: Get 7 packets sent from one cluster of pixels to another cluster of pixels ( )
  • Third Step: Get pixel packets from all other clusters of pixels to one Neuron cluster, then run that specific convolution layer over all pixels from the image ( )
  • Fourth Step: Get the above goal to work with all clusters of Neurons ( )
  • Fifth Step: Implement the pooling step ( )
  • Sixth Step: Set-up the rectifier layer ( )
  • Seventh Step: Create the fully-connected outcome ( )
  • Eight Step: Classify the image ( )

Week 7 Update:

• Now understand the concept of CNN better through messing with tiny_dnn
• Found “final” design for blocks of CNN
• Will now work on communication between the blocks in the CNN
• Also will work on fixing a size for the CNN based around a smallish image

Week 8 Update:

• Updated project page to more clearly state what my projects is and my goals
• Messed around with trying to get my own images as inputs for tiny_dnn
• Researched how small my input images could be for a CNN

Week 9 Update:

• Created Demo video for progress on project
• New Title: Object Classification in the Movable Feast Machine
• Scaled back project a lot to 3×3 hand made images without padding. So they represent object as the pixel values can only be 0 or 1.
  • These 3×3 images can contain a dot (1) or a 2-length line (0).
  • I get a 80% successful classification rate in tiny_dnn with the new CNN structure and these input images
• Researched into CNN structures and tested them in tiny_dnn with hand-made images of 9×9, 7×7, 5×5, 3×3, with padding and without
• New CNN structure is a 12 layer Convolution layer, that goes into a Fully_connected layer that takes in 12 input and outputs 2 results.
• Implemented working convolution layer in MFM, which takes pixels from a image element and will calculate the output from this layer
• Created shell for Fully-Connected layer for MFM, contains weights and biases, but nothing else yet.
• Scratched communication concept, as I can place all layer in a sort of column fashion, so there is no need for this

Week 10 Update:

• Implemented the fully connected layer
• Compared values with weights and results from each layer with the tiny-dnn output, they compared well
  • The results weren't exact, but I found a bug where the values were changing barely based on certain Neurons exceeding their stages and skipping certain steps
• Found some major problems in the weights I was using after testing every single possible input pattern
  • Will find better weights in week 11
• Forgot to put my week 10 update on week 10….

Week 11 Update:

• Found better patterns and shapes to use together, am using a 2×2 box and a horizontal 2 length line
  • Will look into adding more objects that can still be successfully classified
  • Implemented these weights and biases into the project, and it classifies 8/10 possible patterns (6 from the horizontal 2 length line, and 4 from the 2×2 box)
• Does not pull the image (pixels) from a element anymore, now can be hand drawn by the user
• Tested hitting the network with “radiation”, and it failed horribly every time
• Working towards getting a reset to work
  • The neurons in the convolution already had their weights stored separately from their output values, whil the fully connected layer did not
    • I have successfully separated the weights and the output values for the fully-connected layer, and it works perfectly
• Added a reset element, as when the Neurons see it, they will reset the network
• Created a presentation for this project

Week 12 Update:

• Fixed bug with fully_connected layer when the simulation was going too fast for it to correctly pass the weights
• Created scripts to get data from modifying parameters for training in tiny_dnn and then will compare the results to what I get from my pre-trained model
• Added different objects, so It can classify up to 5 now, with a much worse percentage…
• Wrote the abstract for this project