Hi, I’m Andrew Ganse. I’m an data scientist who builds and leads ML systems end-to-end — from data pipelines to deployment. In the commercial world and in the past in the academic science world, the focus of my work has been in machine learning, inverse problems, optimization, signal processing, and data analysis. (My PhD as well as much of my past work as a physicist concerned inverse problems in geophysics). Here on research.ganse.org are some of my publicly shareable research results and toy problems and tools, both in data science topics and in applied physics topics. Note the categorized menus at top, and below is a randomized selection of some of the articles on this site. You can also use the search box at top to find content by keyword. You can contact me at andrew@ganse.org.
A classic textbook for predictive (tracking) filters is Applied Optimal Estimation, edited by Gelb (1974). In section 6.1 of that book are two simple radar tracking examples (6.1-2 and 6.1-3) which demonstrate several nonlinear filters. Both involve a body falling through the atmosphere while a ground radar measures its range, so nonlinearity enters through both the drag dynamics and the measurement geometry. The book uses them to compare the extended Kalman filter, the iterated extended Kalman filter, and a second-order filter, showing how each copes with linearization error. These examples are instructive as the differences are easy to see in simple plots. I've programmed up those examples into a Matlab script and added a few additional filters to compare and contrast them in both linear and nonlinear cases.
If you've seen any of my other DS/ML content you know I'm a big fan of MLflow. The bottom line is that making a new model is typically the quick and easy part. It's figuring out how to make it consistent and how to improve it that are the hard part - and the key is being fastidious and objective about what you (and your team) know about the model and data that you're working with. This is equally true with GenAI language models and RAG as it is for classic machine learning and deep learning. MLflow 3 has an expansive feature set for GenAI development and quantifiable evaluation that provides a lot of the toolset straight out of the box (or "container" 🙂?).
The nature of an icy satellite's interior relates fundamentally to its composition, thermal structure, formation and evolution history, and prospects for supporting life. Gravity measurements via radio Doppler information during spacecraft flybys are an important tool used to infer gross interior structure of these moons. Liquid water and ice layers have previously been inferred in Jupiter's icy satellites Europa, Ganymede, and Callisto via magnetic field measurements by the Galileo probe. We apply geophysical inverse theory tools to assess the icy moon's interior density anomaly distribution that could be estimated from radio Doppler measurements, to support the search for mass anomalies in the ice shell (meteorites or diapiric upwellings) or near the H2O/rock interface (seamounts).
For some types of semantic analysis we don't always need to always pull big dataframes of data out of the database to do the analysis externally if we already have embedding vectors stored in the database. PostgreSQL's pgvector extension provides powerful functionality for semantic search and aggregation right there in the SQL queries. Let's experiment with some of this by grabbing an arbitrary dataset (how about Kaggle's Seattle AirBnB listings/reviews), popping that into a database, running an embedding model on that to create embedding vectors, and exploring SQL queries using this functionality.
Normalizing flow models are invertible neural networks, a type of generative model that in this image context offers a nice two-for-one benefit: simultaneously enabling unsupervised learning for image anomaly detection (by mapping unlabeled images to a distribution where statistical anomaly detection techniques can apply) and also enabling image simulation (by mapping randomly generated samples from a probability distribution into the image space). Now it may not be my ultimate use case, but it turns out there are, you know, a ton of cat images and datasets on the internet, so let's experiment with INNs and cats!
I love MLflow a lot, and I use it all the time for both work and personal projects. There are projects in which I go all-out and auto-log extensive graphical artifacts to better enable modeling improvements - like with example images for the different error cases, and ROC curves, and so on. And for those it's really nice to use its website as a one-stop UI to where those artifacts are logged and reviewed by others. But there are also a lot of projects where I really don't log much more than the modeling parameters and performance metrics, just numbers, and yet I'm still leaving my terminal-based workflow in Vim and looking up and comparing those numbers on the (in this case needlessly point-and-click) MLflow website. So I wrote this Vim plugin…
Let's implement a random forest classifier from Scikit-Learn to see how well we can predict whether a bank client will have good loan behavior (meaning they won't default or become delinquent) if they are given a new loan. We'll use a public bank transactions/loans dataset from the PKDD99 Challenge conference for the modeling. In the process we'll fit and explore the assumptions made for this model, and learn about some limitations of Scikit-Learn's tree-based models.
Let's explore a supervised learning problem in medical imaging based on a public dataset and MLFlow's "Projects" functionality. A self-contained modeling module is trained, has its performance logged in MLFlow, and is able to be checked out as a deployable model image. I created a configurable implementation of this in a repo. Let's try the malaria detection dataset from the Tensorflow datasets, which contains a balanced, labeled dataset of about 27,000 thin blood smear slide images of cells, and let's see how well we can detect malaria parasite presence in the images. This dataset is used to train/test different variations of image classification models, including VGG-16 and various sizes of more basic convolutional networks.
Normalizing flow models are invertible neural networks, a type of generative model that give us a.) a one-to-one mapping to points with probabilities in the latent space corresponding to input points, b.) a way to compute the likelihood of the data exactly, and c.) computationally efficient training and inference, because the model’s transformations have a clever construction allowing the Jacobian determinants to be computed very cheaply. Those details allow for doing things in an understood Gaussian latent space that couldn't be done (or not easily) in the original data space; let's look at that a bit visually in some 2D examples here.
This wonderful kids' book series is fun not only for the stories themselves, but also because each of the first several books involves a cipher puzzle with "fairy hieroglyphics" - I love code puzzles! In the electronic form of the books I discovered the hieroglyphic sequence was moved to the back of the book, all perfectly lined up in matrices over a few pages at the end. And I thought, hey that seems like it'd be easy to parse and decrypt on a computer, just like the main character did. Using intensity profiles to split out the 2600+ glyphs and DBSCAN clustering to label them, I got close enough to read the fairies’ message!
Normalizing-flow models are invertible neural networks (INNs), a type of generative model that I've recently been fascinated with. They are a real Swiss-army-knife of the modeling world, as they allow not only generating new samples from the learned distribution (which GANs and VAEs do too) but also exact likelihood computation as well (which GANs and VAEs do not). By modeling probabilities directly, INNs allow for a range of applications. This series of articles will implement and run a set of examples from key research papers on INNs. It's all the same model for all these applications, just with a few variations in the partitioning of the inputs and outputs. Let's start here with an overview and introduction to INNs in general.