Classifying handwritten digits using KNIME, DL4J and a LeNet variant

The workflow downloads, uncompresses and preprocesses the original MNIST dataset. The two "Normalize Images" components use the KNIME Streaming functionality to convert the input files into KNIME image cells that can be used by the DL4J Learner and Predictor. The "LeNet" metanode (taken from the Node Repository) is a variant of the originally described LeNet convolutional neural network. The images and the DL4J model is then used by the Learner to train a model (saved using the DL4J Model Writer), which is then applied to the test set, which is finally scored.

External resources

• Learning Deep Learning. A tutorial on KNIME Deeplearning4J Integration
• The MNIST DATABASE of handwritten digits
• KNIME Image Processing - Deeplearning4J Integration (64bit only) extension