Andrew Lavin Email & Phone Number

California, United States

Published five years of original research in the paper "On the Efficiency of Convolutional Neural Networks." Major contributions include a theory of neural network efficiency that unifies model efficiency, computational efficiency, and latency; the Waterline performance model for sequences of parallel kernels, which corrects errors with the widely used Roofline analysis; memory-efficient CUDA kernels for ConvFirst and MBConv+SE blocks that achieve up to 14x and 5x speedup over PyTorch Inductor, respectively; the ConvFirst model that achieves ~4x speedup over ConvNeXt with equal accuracy when using our custom kernels. The paper shows that co-optimization of the model and program yields superior performance.Developed the Spio kernel library for PyTorch. Spio is a CUDA kernel framework with named tensors, run time compilation, kernel performance models, and torch.compile integration. The first Spio kernels perform grouped convolution several times faster than the builtin PyTorch kernels.

Mountain View, California, Us

Developed memory-efficient CUDA kernels for MBConv+SE blocks and ConvFirst blocks, implementing entire blocks with fused kernels. Led the development of a graph rewriting compiler for neural networks. Improved the performance and reliability of core system-level code for a production ADAS system.

Mountain View, California, Us

Developed an efficient inference engine for the NVIDIA Xavier SoC using block-fusion kernels written in CUDA and PTX. Achieved ~4x speedup on whole network object detection, semantic segmentation, and other tasks versus NVIDIA TensorRT. Developed a client-server framework for testing and benchmarking, enabling the compile and test framework to run in a Python desktop environment while the inference server runs on the embedded system. Administered the company's deep learning cluster.

Created the ConvFirst building block for convolutional neural networks, predating the similar ConvNeXt block by three years. Performed research on high-performance pre-processing pipelines for neural networks. Also devised multi-scale convnet models. Began contract work on the Phantom AI Inference Engine.

Santa Clara, California, Us

As a contract employee, I researched machine learning algorithms and their deployment on graphics processors.

Austin, Texas, Us

Evaluated the computer vision software stack for the autopilot team. Suggested architectural and algorithmic changes.

Santa Clara, California, Us

Improved efficiency of neural network software on Intel Gen graphics processors using OpenCL. Achieved approximately 4x speedup over the existing codebase.

Austin, Texas, Us

Developed high-performance NVIDIA GPU kernels for the Autopilot system.

Los Angeles, Ca, Us

Optimized a computer vision pipeline for desktop and embedded platforms over the course of a short term contract. Achieved speedup of 2.5X on embedded platform while reducing memory use by nearly half.

Introduced Winograd's fast convolution algorithms into convolutional neural networks while working as an independent researcher. Devised the mathematical formulation of fast tensor convolution as matrix products of coefficients in transform space. Presented the results at CVPR 2016 in the research paper "Fast Algorithms for Convolutional Neural Networks."cuDNN and other deep learning software libraries and ASIC and FPGA designs have all implemented the algorithm. The paper has more than 1,100 citations.Created the winCNN python module for the automatic generation of modified Cook-Toom (i.e., minimal Winograd) convolution algorithms.

San Jose, Ca, Us

Developed the world's first high efficiency convolution neural network kernel for NVIDIA GPUs, reaching 95% computational efficiency for popular deep learning network layers. Used the Maxas assembler and modified the SGEMM sample to perform direct convolution.Profiled the cuda-convnet2 deep learning framework and identified the image preprocessing bottleneck for multigpu systems. Removed bottleneck by performing scan line mean subtraction and color noise addition pipelined with scan line JPEG decoding.

Boston, Ma, Us

Designed and implemented a cross platform C++ library for parsing and rendering Apple Keynote files. Implemented code for iPad screen sharing. Evaluated codecs for next generation screen sharing. Implemented OpenGL based renderer for screen sharing and video conferencing.

Boston, Ma, Us

Designed and implemented next generation video conference user interface. Cross platform code written in C++, with original rendering implementation written for OS X. Advised engineers who ported to Linux and Windows. Extended the Chameleon framework (for porting iOS apps to OS X) with support for drag and drop and tooltips, using Objective-C on OS X. Improved OS X screen capture. Contributed many bug fixes and UI enhancements on OS X and iOS using Objective-C.

Palo Alto, Ca, Us

Created C++ middleware for system programming on Linux, MS Windows, and Mac/iOS. Profiled and optimized existing visual matching algorithm. Designed and implemented API for approximate nearest neighbor search and benchmarked different algorithms.

Developed face detection and automatic image segmentation software. Maintained MS Windows APIs for legacy products. Profiled and optimized image segmentation algorithm. Developed commandline tool for automatic image alignment. Translated several computer vision algorithms from Matlab to C++.

Created a Cocoa server framework for home automation on OS X.Created a flexible, efficient framework for distributed objects on iPhone and OS X. Improved the reliability and speed of legacy home automation system software. Eliminated GUI blocking and communication failures by creating an asynchronous, multithreaded distributed object system using C++ and TCP/IP on Linux.Created a framework for the rapid implementation of RS232 device drivers. Added the facility to control HVAC by implementing a BACnet client protocol stack.

Implemented Trax Systems’ Traceware package recognition system. Created the image analysis algorithm in just 2 weeks. Achieved perfect recognition accuracy for all well formed images regardless of perspective. Implemented the recognition module and web service interface on Linux using C++ and PHP.Wrote parsers for large, complex undocumented data files for the Thirty Meter Telescope Project. Wrote scripts to import daily data files into a relational database and a web site for data visualization. Used Perl, PostgreSQL, and PHP on OS X.Created the Picturepusher photo sharing service. Implemented a fast, cross platform photo management application, featuring a thumbnail viewer, camera import, and photo upload. Implemented image resizing algorithms that achieved high quality while remaining much faster than many publicly available libraries. Implemented a web site with user selectable access controls, transfer and storage quotas, and a scalable architecture. Used C++, PHP, Python, and PostgreSQL on LInux, MS Windows, and Mac.Implemented a sales reporting tool for a department store chain. Used PHP, PostgreSQL, and Oracle to pull daily sales data from the corporate database. Wrote parsers for the customer traffic monitoring system. Generated HTML reports comparing sales and profits per customer across any combination of stores.Designed and implemented open source libraries for cross platform sockets, threads and system logging.Contributed to the wxWidgets for Mac preemptive threads implementation.Created the Threadhandler classes for multi-threaded event handling in wxWidgets.

7th employee in a Silicon Valley startup in digital imaging technologies innovation.Improved the speed of the immersive imaging viewer by an order of magnitude using hardware graphics acceleration. This was accomplished during the first 2 weeks of employment and was key in winning my employer a contract for the development of an immersive video renderer and viewer.Co-inventor of 4 US patents for immersive video.Created an immersive video viewer.Created the image processing and visualization modules for the well reviewed Powerstitch large scale photo panorama software.

San Francisco, Ca, Us

10th employee of a small research group in one of the world’s largest software companies.Improved the speed of Picture This Home’s modeling algorithm by a factor of 2 in just 2 days.Created C++ libraries for linear algebra, nonlinear optimization, and geometric operations.Created the solid modeling user interface for the Origami image based modeling prototype.Provided programming support for researchers from leading universities such as Caltech, MIT, and Stanford.