Feng (Jeff) Liang, 梁丰

I am a PhD student at UT Austin, fortunately working with Prof. Diana Marculescu. I obtained my master and bachelor degree from Tsinghua University and Huazhong University of Science and Technology, respectively.

My current research interests lie in vision-language models and generative AI with a special interest in efficiency. If you find any research interests that we might share, feel free to drop me an email. I am always open to potential collaborations.

I’m graduating in Spring 2025 and actively seeking full-time Research Scientist or Engineer positions in GenAI and vision-language models. I’d love to connect if you have any relevant opportunities!

Email  /  CV  /  Google Scholar  /  Linkedin  /  Zhihu  /  Twitter

We present StreamV2V to support real-time video-to-video translation for streaming input.

We leverage the temporal optical flow clue within video to enhance the temporal consistency for text guided video-to-video synthesis.

For the first time, we show open-vocabulary generalist models match the performance of supervised specialist models without dataset-specific adaptations.

SupMAE extends MAE to a fully-supervised setting by adding a supervised classification branch, thereby enabling MAE to effectively learn global features from golden labels.

We propose Data efficient CLIP (DeCLIP), a method to efficiently train CLIP via utilizing the widespread supervision among the image-text data.

we propose the ANT framework to harness these redundancies for reducing the computational cost of video processing. The proposed ANT adapts a purpose-fit network by inspecting the semantic differences between frames.

We propose RePre to extends contrastive frameworks by adding a branch for reconstructing raw image pixels in parallel with the existing contrastive objective.

We present CRNAS that can learn computation reallocation strategies across different feature resolution and spatial position diectly on the target detection dataset.

We present Once Quantization-Aware Training (OQAT), a novel framework that searches for quantized efficient models and deploys their quantized weights at the same time without additional post-process.

We proposed a new mutant of dilated convolution, namely inception (dilated) convolution where the convolutions have independent dilation among different axes, channels and layers.

We apply our techniques to produce fully quantized 4-bit detectors based on RetinaNet and Faster RCNN, and show that these achieve state-of-the-art performance for quantized detectors.

We propose NASGEM which stands for Neural Architecture Search via Graph Embedding Method. NASGEM is driven by a novel graph embedding method equipped with similarity measures to capture the graph topology information.

We present ScaleNAS, a one-shot learning method for exploring scale-aware representations. Scale-NAS solves multiple tasks at a time by searching multi-scalefeature aggregation.

• MLCommons ML and Systems Rising Stars by MLCommons 2024.
• Qualcomm Innovation Fellowship Finalist by Qualcomm 2024.
• UT Austin Engineering Fellowship by UT Austin, 2021 & 2023.
• Excellent Student Leader by Tsinghua University, 2018.
• National Scholarship by Ministry of Education of China, 2014 & 2015.

Thanks to Jon Barron