Inspiration
We had a lot of experience working with high level libraries like pytorch and fastai that made the job easy for us. But could we replicate the same model with a decent accuracy using just python numpy?
What it does
Classifies handwritten digits using the MNIST dataset with an accuracy of 83% on validation set with just python nupy.
How we built it
Using Python numpy.
Challenges we ran into
Understanding difficult concepts like backpropagation and trying to implement it in python. We had to compromise while choosing the activation functions as it had to be something that we could replicate in python in a weeks time.
Accomplishments that we're proud of
Achieving 83% accuracy in a vanilla python model.
What we learned
Backpropagation and different activation functions
What's next for Computer vision using MNIST dataset
Try to implement Negative log likelihood loss and add biases. Normalise the dataset before working on it.
Log in or sign up for Devpost to join the conversation.