Berivan Isik

I am a research scientist at Google, working on efficient and trustworthy AI.  

My current interests are efficient training/finetuning of large models, pretraining data valuation and scaling laws for LLMs, differential privacy, and unlearning. 

I completed my PhD at Stanford University, advised by Tsachy Weissman and Sanmi Koyejo, where I was affiliated with the SAIL and StatsML groups. My research was supported by Stanford Graduate Fellowship (2019-2023), Google Ph.D. Fellowship (2023-2026), and a Meta research grant. 

If you have similar interests and would like to collaborate, please reach out via email

berivan [at] google [dot] com

LinkLinkedInGitHubTwitter

Publications

Differentially Private Adaptation of Diffusion Models via Noisy Aggregated Embeddings [PDF]

Pura Peetathawatchai, Wei-Ning Chen, Berivan Isik, Sanmi Koyejo, Albert No

Preprint, 2024.



Lottery Ticket Adaptation: Mitigating Destructive Interference in LLMs [PDF] [code]

Ashwinee Panda, Berivan Isik, Xiangyu Qi, Sanmi Koyejo, Tsachy Weissman, Prateek Mittal

Preprint, 2024.

Preliminary version accepted at the Workshop on Advancing Neural Network Training (WANT): Computational Efficiency, Scalability, and Resource Optimization, ICML, 2024. [Best Paper Award, Oral]

Preliminary version accepted at the Workshop on Efficient Systems for Foundation Models (ES-FoMo), ICML, 2024. [Oral]



Scaling Laws for Downstream Task Performance of Large Language Models [PDF]

Berivan Isik, Natalia Ponomareva, Hussein Hazimeh, Dimitris Paparas, Sergei Vassilvitskii, Sanmi Koyejo

Preprint, 2024.

Preliminary version presented at the Workshop on Navigating and Addressing Data Problems for Foundation Models (DPFM), ICLR, 2024. 

Preliminary version presented at the Workshop on Mathematical and Empirical Understanding of Foundation Models (ME-FoMo), ICLR, 2024.



Sketching for Distributed Deep Learning: A Sharper Analysis [PDF]

Mayank Shrivastava, Berivan Isik, Qiabo Li, Arindam Banerjee,  Sanmi Koyejo

Conference on Neural Information Processing Systems (NeurIPS), 2024.

Preliminary version accepted at the Workshop on Bridging the Gap Between Practice and Theory in Deep Learning, ICLR, 2024.



On Fairness of Low-Rank Adaptation of Large Models [PDF] [code]

Zhoujie Ding, Ken Liu, Pura Peetathawatchai, Berivan Isik, Sanmi Koyejo

Conference on Language Modeling (COLM), 2024.

Preliminary version presented at the Workshop on Mathematical and Empirical Understanding of Foundation Models (ME-FoMo), ICLR, 2024. 

Preliminary version presented at the Workshop on Secure and Trustworthy Large Language Models (SeT LLM), ICLR, 2024.



Improved Communication-Privacy Trade-offs in L2 Mean Estimation under Streaming Differential Privacy [PDF]

Wei-Ning Chen, Berivan Isik, Peter Kairouz, Albert No, Sewoong Oh, Zheng Xu

International Conference on Machine Learning (ICML), 2024.



Adaptive Compression in Federated Learning via Side Information [PDF] [code]

Berivan Isik*, Francesco Pase*, Deniz Gunduz, Sanmi Koyejo, Tsachy Weissman, Michele Zorzi 

International Conference on Artificial Intelligence and Statistics (AISTATS), 2024.

Preliminary version presented at the Workshop on Federated Learning and Analytics, ICML, 2023.



Exact Optimality of Communication-Privacy-Utility Tradeoffs in Distributed Mean Estimation [PDF] [code]

Berivan Isik, Wei-Ning Chen, Ayfer Ozgur, Tsachy Weissman, Albert No

Conference on Neural Information Processing Systems (NeurIPS), 2023.

Preliminary version presented at the ICML 2023 Federated Learning Workshop and the Theory and Practice of Differential Privacy (TPDP) Workshop in 2023.



Sparse Random Networks for Communication-Efficient Federated Learning  [PDF] [code] 

Berivan Isik*, Francesco Pase*, Deniz Gunduz, Tsachy Weissman, Michele Zorzi 

International Conference on Learning Representations (ICLR), 2023.

Preliminary version presented at the Workshop on Federated Learning, NeurIPS, 2022. [Oral]



An Information-Theoretic Justification for Model Pruning [PDF] [code] 

Berivan Isik, Tsachy Weissman, Albert No

International Conference on Artificial Intelligence and Statistics (AISTATS), 2022.

Preliminary version presented at the Sparsity in Neural Networks Workshop, 2021. [Spotlight talk]

GPT-Zip: Deep Compression of Finetuned Large Language Models [PDF]

Berivan Isik*,  Hermann Kumbong*, Wanyi Ning*, Xiaozhe Yao*, Sanmi Koyejo, Ce Zhang

Preliminary version presented at the Workshop on Efficient Systems for Foundation Models, ICML, 2023.



Sandwiched Compression: Repurposing Standard Codecs with Neural Network Wrappers [PDF][code]

Onur Guleryuz, Philip A. Chou, Berivan Isik, Hugues Hoppe, Danhang Tang, Ruofei Du, Jonathan Taylor, Philip Davidson, Sean Fanello

Submitted to the IEEE Transactions on Image Processing (TIP), 2024.



Mutual Information Upper Bounds for Uniform Inputs through the Deletion Channel [PDF]

Francisco Pernice, Berivan Isik, Tsachy Weissman

IEEE Transactions on Information Theory (TIT), 2024.



Lossy Compression of Noisy Data for Private and Data-Efficient Learning [PDF]

Berivan Isik, Tsachy Weissman

IEEE Journal of Selected Areas in Information Theory (JSAIT), 2023.



Sandwiched Video Compression: Efficiently Extending the Reach of Standard Codecs with Neural Wrappers [PDF] [code]

Berivan Isik, Onur Guleryuz, Danhang Tang, Jonathan Taylor, Philip A. Chou

IEEE International Conference on Image Processing (ICIP), 2023.



Neural Network Compression for Noisy Storage Devices [PDF]

Berivan Isik, Kristy Choi, Xin Zheng, Tsachy Weissman, Stefano Ermon, H.-S. Philip Wong, Armin Alaghi

ACM Transactions on Embedded Computing Systems (TECS), 2023.

Preliminary version presented at the Deep Learning through Information Geometry Workshop,  NeurIPS, 2020.



Learning under Storage and Privacy Constraints [PDF1] [PDF2 (longer)]

Berivan Isik, Tsachy Weissman

IEEE International Symposium on Information Theory (ISIT), 2022.



LVAC: Learned Volumetric Attribute Compression for Point Clouds using Coordinate Based Networks [PDF] [code] 

Berivan Isik, Philip A. Chou, Sung Jin Hwang, Nick Johnston, George Toderici

Journal of Frontiers in Signal Processing (FSP), 2022.



Neural 3D Compression via Model Compression  [PDF]

Berivan Isik

WiCV, CVPR, 2021.



rTop-k: A Statistical Estimation Approach to Distributed SGD [PDF] 

Berivan Isik*Leighton P. Barnes*,  Huseyin A. Inan*,  Ayfer Ozgur

IEEE Journal on Selected Areas in Information Theory (JSAIT), 2020. 

Preliminary version presented at the Federated Learning for User Privacy and Data Confidentiality, ICML, 2020. [Long Oral]


PhD Thesis

Statistical methods for efficient and trustworthy machine learning (2024) 


Patents

 "Learned Volumetric Attribute Compression Using Coordinate-Based Networks", US Patent 17/708,628.

Philip A. Chou, Berivan Isik, Sung Jin Hwang, Nick Johnston, George Toderici 

Invited Talks

Involvements

Hobbies

I like photography and reading