- Visualizing neuralnets as crumpled paper.ipynb An education visualization I made during my time teaching at MIT Lincoln Labs! Shows how neural networks form these surfaces in high dimensional spaces to fit to the data. Different architectures have different influences on the 'shapes' and 'textures' of the surfaces; I found it helpful for being an intuition beyond a mathematical understanding.
- NLP political sentiment analysis tool.pdf Did you know the word "billionaire" when used in reference to Elon Musk on Twitter is most commonly connotated with the words "narcssist", "fascist", and "nazi"? I built a tool allowing me to procedurally scrap tweets containning a certain keyword to build a corpus, then training a word2vec model to learn word embeddings within this corpus. By doing this, I can then analyze which words share similar embeddings and find which words are similarly connotated Which allows me to make unsupervised, real-time sentiment analysis. Here's the link to the repo: https://github.com/macmacmacmac/Twitter-sentiment-analysis/
- Identifying depression from brain region sizes.ipynb Did you know that long-term depression manifests physically as atrophy and hypertrophy in certain brain regions? I analyzed a dataset of depression vs. control patients and their differences in brain region sizes. In total, there were 253 different brain regions which had a statistically significant (proven via hypothesis tests using t-test for difference of means with alpha value = 0.05) size difference between healthy and depressed subjects. Surprisingly, in certain regions depressed subjects had a higher average region size than control subjects!
- Visualizing how normalizing affects gradients.ipynb A visualization showing the impact of different normalization methods on model convergence! I found it helpful for being an intuition on why and how normalization impacted gradients.
- Implementing pytorch sequential from scratc.ipynb Implementing neural networks from scratch without any reference tutorials using just numpy + I wanted it to be a pytorch-sequential-api-like package as well. Very proud of this one because I did not look up any reference code aside from numpy docs.
