Aadhav Prabu

Handwritten Digit Detection Neural Network

I developed this neural network using the Tensorflow and Keras libraries. My model, which I trained on a small data set of handwritten samples (1000 samples of 20px by 20px images), can take an image of a digit and categorize it as that particular digit, based on how it conforms to the other training samples in that category. I initially built this to categorize only binary digits (0 and 1) as depicted above. This neural net uses three Tensorflow dense layers with sigmoid activations. I then expanded the functionality from binary to multiclass digit recognition. I achieved this by expanding the dataset from 1000 to 5000 samples and switching from sigmoid to ReLU and softmax activations (still using three TF dense layers). My final model can take in a 20px by 20px image of a digit and classify it as 0-9 with 99.7% accuracy (15 misclassifications in a test set of 5000 samples).

Skills Used

Neural Networks (Tensorflow/Keras)




ML Algorithms (Regression/Classification)


Data Visualization (Matplotlib)