An implementation of a image recognition neural network for handwritten digits using the ML.NET framework.
- Dataset from publicly available MNIST database
- Trained regression model using Stochastic Dual Coordinate Ascent (non-calibrated)
- Using Softmax and ReLU activation functions
- Model accuracy was about 90% with a deviation of 6%
- Occasional prediction inconsistencies, likely due to bitmap preprocessing bugs
- UI and formatting improvements to the Windows Form app
- Improve bitmap preprocessing for better model accuracy
- Add additional network layers for better model accuracy
- Change to a nonlinear training method (kNN, SVM, or convolutional) to achieve 97%+ accuracy
Developed by Jie "Jason" Liu (@jiejasonliu)
Contact: jasonningliu@gmail.com