Zheng Zhan

I am a final year PhD student in Machine Learning at Northeastern University, under the supervision of Prof. Yanzhi Wang. I also work closely with Prof. Stratis Ioannidis, Prof. Pu Zhao, and Prof. Jennifer G. Dy. My research interests focus on Efficient Generative AI, Model Compression, and Continual Learning.

Email  /  Google Scholar  /  LinkedIn  /  Resume

• Efficient Deep Learning: E²GAN (ICML'24), SparCL(NeurIPS'22), ICCV'22, LOTUS (DAC'24), Condense (DAC'23), All-in-One (ICCAD'22), NeurIPS'24.
• Generative AI: E²GAN (ICML'24), ICCV'22, NeurIPS'24.
• Efficient LLMs: EMNLP'24.
• Continual (Lifelong) learning: SparCL (NeurIPS'22), DualHSIC (ICML'23).

E²GAN construct a lightweight GAN model and employ Low-Rank Adaptation (LoRA) with a simple yet effective rank search process, rather than fine-tuning the entire base model. We also investigate the minimal amount of data necessary for fine-tuning, further reducing the overall training time.

DualHSIC presents a method to improve rehearsal based approach for continual learning. The basic idea proposed here is to leverage inter-task relationships using two concepts related to the Hilbert Schmidt independence criterion (HSIC). One component (HSIC-Bottleneck for Rehearsal (HBR)) helps reduce interference between tasks and the other component (HSIC Alignment - HA) helps share task-invariant knowledge.

Propose a two-level algorithm for obtaining subnets with arbitrary ratios in a single model with theoretical proof. Develop a framework which leverages the DVFS and compression techniques to get multiple subnetworks in one neural network to lower the variance of inference runtime for different hardware frequency levels. It’s a much more automatic framework.

SparCL explores sparsity for efficient continual learning and achieves both training acceleration and accuracy preservation through the synergy of three aspects: weight sparsity, data efficiency, and gradient sparsity.

All-in-One, a highly representative pruning framework to work with dynamic power management using DVFS. The framework can use only one set of model weights and soft masks to represent multiple models of various pruning ratios. By re-configuring the model to the corresponding pruning ratio for a specific execution frequency (and voltage), we can keep the difference in speed performance under various execution frequencies as small as possible.

We propose a compiler-aware SR neural architecture search (NAS) framework that conducts depth search and per-layer width search with adaptive SR blocks. A speed model incorporated with compiler optimizations is leveraged to predict the inference latency of the SR block with various width configurations for faster convergence.

We proposes to use neural architecture search and network pruning to find a highly efficient network for image super-resolution. we are the first to achieve real-time SR inference (with only tens of milliseconds per frame) for implementing 720p resolution with competitive perceptual performance on mobile platforms.

We prove that the ”ideal” SCNNs and BNNs satisfy the universal approximation property with probability 1 (due to the stochastic behavior), we further prove that SCNNs and BNNs exhibit the same energy complexity.