💡 Colab Link

<aside> ❓ Goal: Build foundational knowledge in PyTorch using fully-connected neural nets and convolutional neural nets. Also MIT OLL Course Code was busted so had to go on my own 🙃

</aside>

Results

Successfully developed ML infrastructure to test different architectures on the MNIST database - achieved 98.98% test accuracy using a CNN.

Little further experimentation was done to push the SOTA FCNN and the CNN architecture further, but strong results were achieved and I was satisfied with my learnings.

Fully-Connected Feed-Forward Neural Networks

• My own architecture from experimenting: 96.22%
  • 784-128-128-10; LR = 3e-4; 5 epochs
• Copying SOTA neural net architecture for MNIST: 97.14%
  • 784-800-10; LR = 3e-4; 5 epochs

Convolutional Neural Networks

• Achieved 98.98% test accuracy with a Convolutional Neural Net
  • LR = 3e-4; 10 epochs