Electrical Engineering
Daniël Reichman, Ph.D.
CEO atAi-RGUS
I am a data scientist with extensive experience in designing and implementing software solutions to automate complex tasks. I gained my expertise by obtaining a doctorate in artificial intelligence and through several consulting projects I have been engaged in. If you have data that needs to be analyzed or need advise on how to architect software systems to support your (data-centric) business I'd be most enthusiastic to hear from you.
What can I do for you ?
Belgium, I enjoy weight lifting and mountain hiking.
The task is to identify the presence of landmines in previously unseen locations, automatically, quickly, and accurately. The data is collected using an electromagnetic modality that measures the composition of the subsurface. This data can be visualized as imagery and therefore, my research focuses on developing software that uses techniques from computer vision, deep learning, image processing, machine learning, and statistical learning theory. I have worked on this problem at Duke University for my Ph.D. and am currently a Research Scientist working on this project.
Example of the automated detection system running on data from the subsurface. The subsurface data can be visualized as imagery in which the buried explosive threats typically appear with a hyperbolic shape. The detection algorithm receives the data column by column and automatically determines whether the data it has collected corresponds to what it has learned to recognize as a buried threat. The red line represents the algorithm in action and after it is done processing, it will provide a prediction of buried threat presence at that location. The dark regions correspond to low probability and the light regions correspond to high probability of buried threat presence.
Several projects of mine have involved color imagery for which I have developed new methods and applied existing methods to solve the problem. My experience in image processing spans the use of hand-crafted feature design, feature learning (e.g., dictionary learning) as well as using convolutional neural networks and deep learning. These methods were applied to natural imagery and to the problem of rooftop detection in satellite imagery (color and height maps).
The top row shows examples of satellite images captured over different US cities. The bottom row shows the output of a rooftop detector using deep learning algorithms. The white spots indicate locations where the algorithm estimates rooftop presence.
Much of my work is to write software to run scientific experiments on the dataset at hand. To ensure that experiments are repeatable and to minimize errors in experiment implementation, I have designed large frameworks of code to support that goal. These frameworks have increased productivity throughout the lab by systematizing the workflow for other members so they may (1) share code more easily, (2) minimize the rewriting of code, and (3) reduce the number of bugs in the code. I have written frameworks for four different projects that are used by multiple people and have been deployed in different computing environments (cloud, different operating systems, varying hardware). They have had to meet various degrees of data sensitivity and for use in analyzing big data as well fine-grained detailed analysis.
Illustration of a part of the class hierarchy in the framework that underpins our work on the landmine detection projects
Software to analyze large archives of textual documents is challenging, partly, because such documents are not necessarily structured in any particular format. I was approached by a researcher of charitable foundations who planned to write a book on the 60,000 documents spanning from the years 1986-2008 of that foundation. This required developing software to automatically generate a taxonomy of the documents. This taxonomy was then used to identify word clouds and common topics across the documents and across the years. I also made a GUI to be able to browse through the topics in the corpus in a convenient manner.
The project was split into three parts where the first part is completed with a book published on the topic. The publication for the second part is currently in preparation.
Screenshot from the Graphical User Interface (GUI) that was developed for the purpose of visualizing and analyzing the set of 60,000 documents.
At my summer internship at New York Life Insurance Company, I applied machine learning to systematically and drastically reduce the run-time of models that ensure the company's solvency (i.e. does not go bankrupt). An insurance company has many liabilities (the policies it issues) that it must match with assets (stocks, bonds, etc) to ensure its ability to pay off the policies. The model determines whether the given set of assets, under different interest rate scenarios over time, will balance out against the liabilities. I worked on a clustering algorithm that can summarize the assets while maintaining model accuracy. This reduced the run-time of the model to a few hours down from a few days.
The flowchart corresponding to the Asset Liability Modeling (ALM) to compute the company's solvency under different economic scenarios (e.g., low interest rates).
