Email: de@deigen.net
Link to Google Scholar page
Designed and led development of model packaging and encapsulation, including input/output formats, dependency installs and runtime resource estimates. Models can be shipped between clusters, including air-gapped environments, without being tied to container base images.
Led efforts to integrate third-party open source LLMs into our training and inference pipelines, including larger GPU deployments, profiling and train adapter training.
Developed system to automatically scale model deployments in a cloud environment based on inference request load. Can load and scale up neural network models from zero on demand to handle both unexpected bursts and slowly adapting traffic, and efficiently share GPUs between models.
Led efforts to integrate our training and experimentation system (see below) with our data platform to automatically create image-based object detection and classification models in multiple cluster environments. Handles data validation, model training, evaluation, and inference deployment. Defined benchmarks to measure accuracy and speed for different types of models, and found best price/performance points.
Developed prototype method to estimate camera field of view, pitch and roll from single image. Located and assembled initial datasets and models, simplified and improved models adding normalized angle regression and NLL loss. Combined FOV with horizon estimates to find pitch with its estimated certainty.
Wrote core execution engine for video object detection and tracking: frame buffering, parallel inference, serialized tracking and completion, and interface with custom video streaming protocols used by customer.
Led ML work to create object detector for near-realtime detection on aerial videos in a government contract. Developed further improvements to detection methods, data cleaning, and measurement, resulting in significant performance gains. Top-performing system in comparison to other contract competitors.
Mentored interns on projects in few-shot learning. Published works at CVPR 2019 and NeurIPS Meta-Learning Workshop 2018.
Wrote object detection code framework for use with in-house neural network library and tensorflow. Created object detection models performing at state-of-the-art accuracy and ~1.5x faster compared to concurrently developed opensource object detectors.
Built job scheduler and experiment tracking system targeted for ML model building, comparison, reproducibility and change tracking. Allowed team members to independently run modified codebases in our cluster, encapsulating most dependencies and automatically tracking changes for reproducibility.
Developed fast classifier training system used by customer end users and internal company teams to quickly build thousands of classifiers in diverse applications. Includes work on frozen classifier embeddings, quantization, data balancing, and neural network architecture and settings. Developed benchmarks based on first customers' uses in beta deployment to rework key components for general release.
Built a system to train detection models to recognize logos in images based on synthetic data; filed patent application.
Created industry-competitive face detection and recognition system, based on data collected from publicly available sources. Initial labeling for the detector based on combinations of open-source detectors with data filtering, and refined with hard example mining.
Created industry-leading classifier for image-based content moderation and filtering, using target labels created automatically from a word-based classifier applied to weak labels and user-supplied text.
Extended our depth-prediction network (below) also to surface normals and semantic segmentation, as well as higher resolution outputs. ICCV 2015.
Predicts depth from a single image using a multi-scale convolutional network that first predicts the depth coarsely looking at the full image, then refines using local information. State-of-the-art single-image depth estimation. NIPS 2014.
Part of team to develop then-state-of-the-art ImageNet object localization and detection system. My contributions concentrated on developing the location regressor and localization task (winning entry 2013), and helping integrate this with the object detection system. ICLR 2013.
Removed structured corruption from images, such as those in pictures taken through a layer of dirt or water droplets using a convolutional network. ICCV 2013.
Trained a network model to learn image features from unlabelled or partially labeled data, using a feed-forward convolutional sparse coding approximation. Inspired by Deconvolutional networks, our model produces similar features while requiring 1-2 orders of magnitude fewer coding iterations. Unpublished work.
In this project, we explore two ways of customizing nearest-neighbor results for individual queries in the context of a kNN image classifier. First, we learn per-descriptor weights that minimize classification error, using backprop through the NN lookups and calculations. Second, we adapt the training set used for each query based on image context; in particular, we condition on common classes (which are relatively easy to classify) to improve performance on rare ones. The first technique helps to remove extraneous descriptors that result from the imperfect distance metrics/representations of the data. The second contribution re-balances the class frequencies, away from the highly-skewed distribution found in real-world scenes.
Created a system to classify IP addresses as likely spam or ham senders for email based on recent trap rates, using as input live streams of spam trap hits and overall mail volume estimates.
Created a system to automatically classify web page content into 30 categories, based on Naive Bayes classification methods. The system categorized over 10 million sites with an estimated misclassification rate of under 10% at 50% recall.
Developed a system that rates HTTP requests with a score indicating the chance the request might fetch malicious content. The system combines multiple sources of data consisting of URL portions or IP ranges; each source may contribute positively or negatively. Reputation is used on web proxy devices to block potentially malicious requests, as well as divert from further scanning traffic that is highly likely to be clean.
Developed systems to process traffic samples sent back from web proxy appliances. IronPort has several thousand appliances deployed at customer sites throughout the world, forming a sensor network of web-related data. This data is automatically fed back into Web Reputation and other systems. We also use it to evaluate efficacy and measure new techniques.
Extracted static client code and dependencies into a dynamically updateable package, taking into account potential differences in system libraries and hardware architecture between client platform releases. Distinct from system upgrades, the engine is automatically updated live on the appliance to the current provisioned version, without any input or interaction by the administrator.
The NVLog is an intent journal used in WAFL (Write-Anywhere File Layout) to ensure data integrity in the event of a system crash or abrupt shutdown. Since all writes are logged to the journal, this part of the system was a serialization point and bottleneck. I rewrote the way journal writes are done to nearly eliminate lock contention, leading to an overall system performance gain of over 10% by throughput.
Outlined a design for dynamic microcores, a reporting and debugging feature. Microcores are partial coredump files, only a few megabytes. The project aimed to let engineers write descriptive recipes to identify memory regions, and trigger microcore generation upon hitting system events. For example, if a system message warns about possible corruption in a block, one could identify interesting memory regions relative to the in-core structures for the block and inode in the message.
Investigated the question of whether the use of different interaction techniques might impact the memory of a user. I designed a set of two experiments to address this: The first, a preliminary study confirming a well-known difference in performance between positional- and velocity-based controls, helped to verify my experimental methodology. The second, a comparison between three interaction modes in an immersive environment, was statistically inconclusive. Anecdotal evidence, however, suggested that for many subjects, memory performance improved with a full-body walking interaction.
Wrote and reviewed proposals for a 6-week project in a mock grant proposal process during a class on scientific visualization. Carried out research on my project on electrode parameter settings in deep brain stimulation for obsessive compulsive disorder, collaborating with Benjamin Greenberg, a psychiatrist at Butler Hospital. Presented a poster of this project in SIGGRAPH 2004.
Created a software package for creating interactive differential geometry visualizations, and produced class labs and demonstrations using this software. Staff and students continue to use this package in new applications and to explore mathematical concepts in several courses at Brown, including differential geometry, combinatorial topology, calculus, geometry, and linear algebra. It has also been used by Prof. Banchoff in classes at UCLA, Notre Dame, University of Georgia, and Stanford.