Research
LLM Alignment with Model Editing
Collaborated with Lei Gao, Tingting Tang, Salman Avestimehr, Murali Annavaram
In this project, we aim to align and rectify language model,
and reduce potential bias, toxic information, privacy leakage.
We first analyze model weights in an orthogonal space and identify the part that
encode undesired knowledge.
Then we directly edit a pre-trained model and remove weights that encode undesirable knowledge
such as bias and toxic information.
Read More ...
Asymmetric Learning for Privacy-Preserving ML
Collaborated with Ramy E. Ali, Saurav Prakash, Salman Avestimehr
In this project, we explore a new privacy-preserving learning and inference framework: private learning with asymmetric
data flows.
We show that exploring low-rank structure of data in machine learning is very important in terms of complexity
reduction and privacy protection.
Along this line of research, we leverage low-rank structure of data and propose asymmetric learning framework that
achieves privacy-preserving model learning and inference.
Read More ...
Federated Learning at the Resource-Constrained Edge
Collaborated with Saurav Prakash, Souvik Kundu, Sunwoo Lee, Salman Avestimehr
Federated learning of large neural nets at the edge faces significant challenges due to
the limited computation and memory on-device resources.
Sub-model training methodology presents a promising solution.
We investigate a new sub-model training method that reduces computations at the edge while still attaining a full model.
Read More ...
High-Order Optimization on Large-Scale NN Training
Collaborated with Zalan Fabian, Sunwoo Lee, Mahdi Soltanolkotabi, Salman Avestimehr
High-order/Newton methods face significant challenges in current large-model training era, due to
their quadratic computation and memory complexities.
In this project, we aim to convey an effective quasi-Newton method, L-BFGS, to large-scale model training.
To address convergence instability of L-BFGS in stochastic optimization, we introduce a momentum to the Hessian approximation.
Read More ...