I write a blog on Custom ML Solutions
There are many unique ML problems, and many problems require custom solutions to address the business problem at hand. The thought process for designing these custom ML solutions, however, is very standard and reused. My blog (once filled out) will describe and give code for many custom ML systems for generating real business value.
​
These are projects I have built, but I have not built many blog posts yet. The blog will populate over time and these topics will link there.
Insight Gaining
Natural Language insights: What phrases are customers saying most positively/negatively about a category?
​
Marketing Attribution/optimization/forecasting: Not sure how much I should say about this yet.
Dimensionality reduction and clustering: Most datasets outside of academia do not group into pretty clusters when you run PCA or TSNE. When you need to find clusters and your off the shelf math algorithms don't work, how do you find clusters? Applications include visualizations for user-base understanding and preparing to solve cold start problems.
​
ML for Business Application
​
Introduction to ML: Regression vs Correlation for measuring ROI/Feature Impact
​
​
Recommendation Engines: Solving the cold start problem
​
Recommendation Engines - Special Case: Handling new parents and their predictable trajectories rather than static preferences.
​
Computer Vision: Object Localization and Tracking (Productionized to count pedestrians, cars, strollers by intersections)
​
NLP: Named entity detection and relationship detection. Which object in text is affecting what other object in text?
​
NLP: Building a chatbot that can adopt personas on command.
​
Low Level ML
​
+Deploying the simple Neural Network in a more realistic setting (and watching it initially fail)
​
Recommendation Engines: Solving the cold start problem
​
NLP: Modern Standard Language Modeling Techniques (Attention and knowledge bases)
​
Computer Vision: Recovering 3D reconstruction of object given one or more images (Traditional methods solve easily with many images, deep learning now solves with one image)
​
​