I recently read the LLAMA paper from an industry pracitioner’s point of view. The quoted sections are exerepts from the original paper. The following comments, questions, and observations are based on my margin notes. However…
Before neural netowrk frameworks were released, there was no effective way to automatically run gradient descent, defined model layer dimension, or initialize weights. All these features are baked into Tensorflow and Pytorch. In a…
Writing Good Code vs Training More Models In this project, we had relatively poor model performance. I would like to do a little introspection below. Original Project Report Introduction In many real world settings, having access…
Author’s Note I wanted to reflect on this past project. My team decided to join an existing machine learning competition where the source data and a base line model was provided to us. The following is an excerpt from an original…
Introduction The following project is a classic example on why complicated and novel methods aren’t always the best. OMsignal is a company specializing in analytics on electrocardiogram (ECG) signals. The project requires using…
Introduction In order to alleviate the hardships of people with vision loss when navigating through a city, the Humanware project aims to create an application to identify text and numbers in real time using the camera of a phone…
Abstract The task of classifying the accent of recorded speech has generally been approached with traditional SVM or UBM-GMM methods (Omar and Pelecanos, 2010; Ge, 2015). However, modern deep learning methods yield the potential…
In deep learning, we use normalization to solve gradient vanishing or explosion. It kind of inspired me to apply normalization to a recipe. Let’s dive in. This is a gluten free, diary free, processed sugar free cake recipe. 1/…