Romain Cosentino
San Francisco,CA
Summary
I specialize in analyzing mathematical formulations of deep neural networks, gaining insights into their architecture and representation. This understanding enables me to develop novel algorithms that tackle both theoretical challenges and practical applications, driving advancements in artificial intelligence.
Overview
9
9
years of professional experience
Work History
Principal Scientist
Salesforce
09.2024 - Current
- Leading a team of Applied Research Scientists and Research Scientists.
- Conducting foundational and applied research on AI agents, with a focus on understanding transformers geometry, enhancing reasoning, and improving safety.
- Leading efforts to evaluate and develop Voice AI agents using Omni models.
- Driving machine learning initiatives to integrate the Tenyx Voice Agent solution into the Salesforce Agent environment.
- Overseeing the development of an auto-evaluation pipeline for Voice AI agents.
Head of Research
Tenyx
07.2022 - 09.2024
- Salesforce acquired Tenyx in September 2024. Key employee leading the research due diligence process.
- Developed a novel, efficient fine-tuning optimization algorithm, enabling continual learning for deep networks across any architecture.
Achievements: Two patents in progress.
Press Coverage: VentureBeat – Tenyx aims to fix LLMs' catastrophic forgetting problem.
- Led the development of the TenyxChat series of LLM models.
Achievements: First open-source model to outperform GPT-4 on reasoning tasks.
Press Coverage: VentureBeat – AI startup Tenyx's fine-tuned open-source Llama 3 model outperforms GPT-4.
- Initiated and led the theoretical development of novel geometrical insights into LLMs, highlighting their input space partitioning properties.
- Led the development of a novel local optimization paradigm to address the backpropagation locking problem.
- Contributed to the end-to-end production pipeline for LLM-based voice chatbot services, including data generation, fine-tuning, and evaluation.
Postdoctoral Research Associate
University of Southern California
09.2021 - 06.2022
- Geometrical Understanding of Transfer Learning
Developed theoretical and practical frameworks to analyze the geometry induced on deep neural networks' latent representations by self-supervised learning algorithms. Leveraged these insights to establish correlations between the geometry of learned representations and their effectiveness in specific transfer learning tasks.
Artificial Intelligence Research Intern
Intel Labs
06.2021 - 09.2021
- Learning Invariant Representations for Similarity Search - Self-supervised learning.
Research Associate
Simons Foundation
06.2020 - 08.2020
- A Dynamic Mode Decomposition Approach to the Auditory System.
Research Intern
Rice University
04.2016 - 08.2016
- An Adaptive Deep Scattering Spectrum for Seizure Prediction.
Education
Ph.D. - Electrical & Computer Engineering
Rice University
Houston, TX01.2021
M.S. - Applied Mathematics & Machine Learning
Ecole Normale Superieure Paris-Saclay
Paris, France01.2016
M.S. - Artificial Intelligence
Katholieke Universiteit Leuven
Leuven, Belgium01.2015
M.S. 1st Year - Applied Mathematics & Actuarial Sciences
Universite Libre De Bruxelles
Brussels, Belgium01.2014
M.S. 1st Year - Mathematics
Universite Paris VI Pierre Et Marie Curie
Paris, France01.2013
B.S. - Applied Mathematics and Economics
Universite Paris VII Diderot
Paris, France01.2012
Skills
- Python
- Pytorch
- Peft
- Deepspeed
Patents
Effect detection via convex hull similarity, Pending
R. Cosentino, S. Shekkizhar, Y. Salomon
Filed, January 2025, US 63 75 26 27
Gradient-free optimization of large language models, Pending
R. Cosentino, S. Shekkizhar
Filed, January 2025, US 63 752 618
Knowledge base for voice large language model applications, Pending
S. Shekkizhar, R. Cosentino,
Filed, January 2025, US 63 752 613
Domain manifold based model compression, Pending
R. Cosentino, D. Kalajdzievski, S. Shekkizhar, A. Earle
Filed, October 2024, US 189 05 761
Training a target activation sparsity in a neural network, Pending
D. Kalajdzievski, R. Cosentino, S. Shekkizhar, A. Earle,
Filed, August 2024, US 188 02 235
Endpoint detection, Pending
A. Earle, A. Ziaei, J. Weissenbeger, R. Cosentino
Filed, April 2024, US 186 36 671
Training multi-task neural network while minimizing catastrophic forgetting
R. Cosentino, A. Earle
Issued, US Patent 11 922 324
Fine-tuning Machine Learning Model while retaining accumulating knowledge, Pending
R. Cosentino, S. Shekkizhar, A. Earle, D. Kalajdzievski, J. Weissenbeger, I. Arel
Filed, October 2023, US 184 966 98
Publications
Reasoning in Large Language Models: A Geometric Perspective,
ArXiv, 2024, R. Cosentino*, S. Shekkizhar* (* equal contrib)
Characterizing Large Language Model Geometry Helps Solves Toxicity Detection and Generation,
ICML, 2023, R. Balestriero*, R. Cosentino*, S. Shekkizhar* (* equal contrib)
The Geometry of Self-supervised Learning Models and its Impact on Transfer Learning,
ArXiv, 2022, R. Cosentino*, S. Shekkizhar*, M. Soltanolkotabi, S. Avestimehr, A. Ortega (* equal contrib)
Toward a Geometrical Understanding of Self-supervised Contrastive Learning,
ArXiv, 2022, R. Cosentino, A. Sengupta, S. Avestimehr, M. Soltanolkotabi, A. Ortega, T. Willke, M. Tepper
Spatial Transformer K-means,
Asilomar, 2022, R. Cosentino, R. Balestriero, Y. Bahroun, A. Sengupta, R. Baraniuk, B. Aazhang
Manifold Approximating Graph Interpolation of Cardiac Local Activation Time,
IEEE Trans. on Biomedical Eng., 2022, Jennifer Hellar, Romain Cosentino, Mathews M John, Allison Post, Skylar Buchan, Mehdi Razavi, Behnaam Aazhang
Deep Autoencoders: From Understanding to Generalization Guarantees,
MSML, 2021, R. Cosentino, R. Balestriero, R. Baraniuk, B. Aazhang
Graph-Based Interpolation of Local Activation Time on the Cardiac Surface (Best Student Paper Award),
Asilomar, 2021, Jennifer Hellar, Romain Cosentino, Mathews M John, Allison Post, Skylar Buchan, Mehdi Razavi, Behnaam Aazhang
A novel convolutional neural network for reconstructing surface electrocardiograms from intracardiac electrograms and vice versa,
AI in Medicine, 2021, A. Banta, R. Cosentino, M. John, A. Post, S. Buchan, M. Razavi, B. Aazhang
Learnable Group Transform,
ICML, 2020, R. Cosentino, B. Aazhang
Universal Frame Thresholding,
IEEE Sig., 2020, R. Cosentino, R. Balestriero, R. Baraniuk, B. Aazhang
Sparse Multi-Family Deep Scattering Network,
ArXiv, 2020, R. Cosentino, R. Balestriero
The Geometry of Deep Networks: Power Diagram Subdivision,
NeuriPS, 2019, R. Balestriero, R. Cosentino, B. Aazhang, R. Baraniuk
Spline Filters For End-to-end Deep Learning,
ICML, 2018, R. Balestriero*, R. Cosentino*, H. Glotin, R. Baraniuk (* equal contrib)
Best Basis Selection Using Sparsity Driven Multi-Family Wavelet Transform,
GlobalSIP, 2016, R. Cosentino, R. Balestriero, B. Aazhang
Invited Speaker
- 03/01/25, TDX - San Francisco, Why Is Voice AI Different?
- 08/01/23, ADP AI Group - New York City, Efficient Finetuning of Large Language Models without Forgetting
- 02/01/19, Simons Foundation - Flatiron Institute - New York City, Learnable Time-frequency Analysis for End-to-end Deep Learning
Timeline
Principal Scientist
Salesforce
09.2024 - Current
Head of Research
Tenyx
07.2022 - 09.2024
Postdoctoral Research Associate
University of Southern California
09.2021 - 06.2022
Artificial Intelligence Research Intern
Intel Labs
06.2021 - 09.2021
Research Associate
Simons Foundation
06.2020 - 08.2020
Research Intern
Rice University
04.2016 - 08.2016
Ph.D. - Electrical & Computer Engineering
Rice University
M.S. - Applied Mathematics & Machine Learning
Ecole Normale Superieure Paris-Saclay
M.S. - Artificial Intelligence
Katholieke Universiteit Leuven
M.S. 1st Year - Applied Mathematics & Actuarial Sciences
Universite Libre De Bruxelles
M.S. 1st Year - Mathematics
Universite Paris VI Pierre Et Marie Curie
B.S. - Applied Mathematics and Economics
Universite Paris VII Diderot
Similar Profiles
Nigar ValoraNigar Valora
Success Guide at SalesforceSuccess Guide at Salesforce
LOUIS FENOGLIOLOUIS FENOGLIO
Senior Business Development Manager at SalesforceSenior Business Development Manager at Salesforce
Laura EscobarLaura Escobar
Mid Market Account Executive at SalesforceMid Market Account Executive at Salesforce
Joe PuliceJoe Pulice
INDUSTRY GO-TO-MARKET / INDUSTRIES MARKETING LEAD at SalesforceINDUSTRY GO-TO-MARKET / INDUSTRIES MARKETING LEAD at Salesforce
Naz FazliNaz Fazli
Senior Executive Assistant at Kraken Digital Asset ExchangeSenior Executive Assistant at Kraken Digital Asset Exchange