Recurse Center - Batch 2 - Cycle 20241008-202410010 - Convolution

Convolution

Another cycle that was motly focused on ARENA plus working through the Neural Networks from Scratch book.

Day 1

Today was Impossible Day at RC, so I started by finishing Chapter 5 in NNFS and then worked through translating this tutorial on building an ASR system with a Transformer model from Keras to PyTorch. I got it mostly done, but wasn't able to get it running on Heap due to some issues with Heap's Python environment. Something to work through on another day :)

• LJ Speech Dataset

• LJ Speech Tools

• OpenSLR

A fellow Recurser also presented on FFTs and showed off their guitar tuner written in C!

Day 2

Today was the start of working on Convolutional Neural Networks

• A Comprehensive Guide to Convolutional Neural Networks

• Fundamental Algorithm of Convolution in Neural Networks

• But what is a convolution?

• Visualizing and Understanding Convolution Neural Networks

• torch.nn.Conv2d

• torch.nn.functional.conv2d

And then I wrapped up finishing Chapter 6 in NNFS

Day 3:

This morning a fellow Recurser introduced me to Piper, a text-to-speech program that can run on a Raspberry Pi. Something to look into and play with!

In our ARENA check-in call, we talked about the following projects:

• einops in Rust

• polars

• candle

• rye

Today I continued working on CNNs and encounted the followint topics:

• Batch Normalization in Convolutional Neural Networks

• Deep Residual Learning for Image Recognition

• torch.nn.MaxPool2d

• torch.nn.Sequential

• torch.var

• torch.nn.BatchNorm2D

Things for next cycle

• Finish Chapter 0 in ARENA

• Finish Chapters 7, 8, and 9 in NNFS

• Update my ml_ai_self_study repo on Github

• Work on my ASR system on Heap with transformers :)

• Think about creating a synthetic speech dataset for numbers 0-9...

  • https://github.com/Jakobovski/free-spoken-digit-dataset

  • https://adhishthite.github.io/sound-mnist/