Below, read about two recently wrapped passion projects and three ongoing ones by Metis Sr. Data Scientists.
COMPLETED PROJECTS (Q2 2018):
Legible Machine Learning - Machine Learning from Scratch with Python by Zach Miller
Zach developed a repository of machine learning programs from scratch. His library has over 50 unique machine learning algorithms, many of which we teach in the bootcamp. Each algorithm is presented as a fully documented notebook containing all the code needed to run the algorithm and a demonstration of the algorithm in action. He set out to write a number of algorithms from his own perspective to ensure that, when reading his code, you can understand exactly how the algorithms are working.
The Math Underlying Leading Neural Net Architectures by Seth Weidman
For his project, Seth (a lover of neural networks) developed convolutional and recurrent neural networks from scratch to understand how they work. The former is typically used to classify images and the latter to recognize patterns in text, voice, and images. Seth implemented a framework to build multi-layer RNNs with long short-term memory cells and showed how this framework can learn to predict the next character when given Shakespeare’s collected works.
CURRENT PROJECTS (Q2 2018):
High-Dimensional Time Series Forecasting with Neural Networks by Joe Eddy
Joe will tackle the problem of making forecasts with large-scale (thousands or hundreds of thousands) related time series. He plans to use sequence-to-sequence neural networks to create models that are simultaneously complex and generalizable and develop a user-friendly, open-source package for building such models. For a sneak peek into his thinking, check out his recent related blog post on Forecasting with Neural Networks - An Introduction to Sequence-to-Sequence Modeling Of Time Series.
Applications of Deep Learning Architectures for Feature Engineering in Image Processing by Roberto Reif
Roberto will apply convolutional neural networks to tackle image classification problems in two applications. One is to create a shoe recommendation system where a user can select the image of a shoe and be shown images of similar types of shoes. The other consists of classifying images from the retina of healthy and diabetic retinopathy patients, using this classification for medical diagnosis. As a start, check out his two recent blog posts: Limitations for Applying Dimensionality Reduction using PCA and Image-Based Product Recommendation System.
Automated Narrative/Conversation Generator by Vinny Senguttuvan
Vinny will use deep neural networks to create a narrative system. As a proof of principle, he plans to create a detective game set in a mansion. Each character will be built as an agent (using recurrent neural networks) who can freely and independently have a dialog with the player.
For more on our Sr. Data Scientists, click here and scroll down to reach about each member and get links to more of their work.