DeepComponents

Image Recognition by Connected Component-based Deep Neural Networks

Data

• MNIST
• For each image, for each row within the image matrix, a representative sparse vector is created. In-a-row non-zero values correspond to the connected component of the image within that row. If there are more than one in-a-row non-zero sequences, there are more than one connected components in that image matrix row.
• For each connected-component representation of each image, an additional abstraction vector is created.
• In this connected-component-image-vector, for each row of the representation, a row with no connected components is mapped to a 0, and a row with connected components has the lengths of those connected components mapped to an element of the connected-component-image vector with the same value.
• MNIST data is 784-dimensional.
• This process reduces dimensionality of MNIST to 51-D at most, and 21-D at least - after removing zeros from generated representation vector
• In the future, I will train the algorithm on the noisier augmented-MNIST dataset.

Morse Function Representation

• Digits: 1 (Morse Functions)

• Digits: 1 (Original Pictures)

• Digits: 3 (Morse Functions)

• Digits: 3 (Original Pictures)

Neural Network

• PyTorch
• Very simple, 2-hidden-layer model with SGD optimization and momentum, Cross-Entropy Loss, ReLU activations.

Results

• With the 51-D representation, max accuracy of 88%

Generation of Images

• Generating Images (with noise) from 51-D representation sequences
• Under Visualizations/generated