Vladimir Feinberg

Gemini Pretraining Area Lead

Gemini Pretraining Area TL

Gemini Flash Pretraining TL

Oct 2021 to Jul 2022 Google Research Cerebra Quantization for Ads DNNs, Software/Hardware codesign. Developed a static, calibrated quantization library and pushed for ads integration & compiler support to achieve XX% latency for click-through rate prediction models. Jul 2022 to Apr 2023 Google Brain, Algorithmic Efficiency Developed new quantized training technques for XX% LLM training speedup. Second-order optimization algorithms research (see NeurIPS paper, Sketchy, below) Apr 2023 Google DeepMind Led finetuning for initial Bard launch, earning a spot bonus from Jeff Dean. Led pretraining & finetuning & RLHF for several PaLM2 sizes (https://ai.google/static/documents/palm2techreport.pdf), earning company-level recognition (Google Tech Impact Award) Led architecture & training for Codey v2 model (Nature acknowledgements: https://www.nature.com/articles/s41586-023-06924-6, launched in https://blog.google/technology/developers/google-colab-ai-coding-features/, https://blog.google/products/search/search-labs-ai-announcement-/, and https://cloud.google.com/blog/products/application-development/introducing-duet-ai-for-developers). Co-led training for distilled on-device generative AI models used in the launches described by Rick Osterloh (https://youtube.com/clip/Ugkx3wQZoi0fudU06amUzEbH8UhTxGOaXTku?si=KhIkZgYbls3yFCFI) Currently Gemini, focussed on efficient large multimodal model training & serving.

I have contributed to the Sisu Diagnostics Platform as the first engineer and led our machine learning effort as the Head of Machine Learning. I have worked on scalable inference methods to explain drivers behind customer key performance indicators by building a new automated inference engine with false discovery control for high-dimensional sparse settings via a custom-built optimizer for a hard-fought 10x improvement over Vowpal Wabbit. My management contributions have been leading ML product roadmap, creating an ML team from scratch by hiring IMO medalists, ICML authors, math professors with articles in Transactions of the AMS, and training via biweekly reading group: https://sisudata.com/blog/learning-unsupervised/

Spark MLlib team. Implemented Spark Catalyst primitives to make online machine learning possible in a streaming environment. Created a proof-of-concept with online adaptive gradient descent with feature hashing for binary logistic classification. Also created keyed models for distributed training in spark-sklearn.

Cloud Dataflow team. I created and instrumented a Spanner instance for profiling workflows for optimization and auto-scaling, which speeds up execution by enabling better scheduling (up to 15% speedups). I also instrumented a critical path metric for workflows (DAGs of execution steps and dependencies), and patched deadlock-y shutdown sequence orchestration.

Weekly problem sessions for students in the MAT 216 and MAT 218 sequence (accelerated honors real analysis, in one and multiple dimensions, respectively).

Kudu team. I increased scan speed with query codegen (up to 2x speedup) and concurrent B+-tree prefetches (up to 1.5x speedup), used knapsack for compaction optimization (20x speedup), and worked on the client-side API as well. Over winter break, I reduced the number of server-side copies by augmenting the RPC mechanism with side buffers.

Applied computational geometry for electronic design automation: implemented Fortune's algorithm for Voronoi diagram generation on large chip design point map inputs, used to accelerate voltage calculations with nearest neighbor queries.

Advised by Professors Ion Stoica, Joseph E. Gonzalez, Michael I. Jordan. Dropped out.

Certificate program in Statistics and Machine Learning.