Not all users are turned away by chatbots for making suspicious requests. Your identity plays a key role when chatbot decides whether to refuse your potentially problematic inquiry.

Our new study shows that the ChatGPT guardrails sensitive knowledge to users with different gender, racial, and political demographics. In particular, younger, female, and Asian-American users are more likely to trigger the refusal from the ChatGPT when querying sensitive information. We proposed a new evaluation framework to identify such bias in the chatbot's refusal behaviors.


Paper Preprint (to appear at EMNLP 2024 Main)

Have you ever thought about if chatbot LLMs are internally modeling your profile? If they are, how might this model of you influence the answers they give to your questions?

Our experiments provide evidence that a conversational AI is internally profling its user during the chat. We design an end-to-end prototype—TalkTuner—that expose this internal user model to the users. User study shows this new chatbot UI design impacts the user's trust in AIs and expose biases of LLM systems.


Paper Preprint

Recent work from Gurnee et al. show that the activations of neural language models have high correlation with the spatial and temporal properties of their inputs. However, the work didn't estabilish a causal link between two. Our experiments, filling a part of this blank, show that editing LLM's spatial representations can improve the model's performance on a simple spatial task.


Paper Preprint

Does the 2D image generative model has an internal model of 3D geometry? Can a 2D neural network see beyond the X-Y dimension of a matrix of pixels? Our project found controllable representations of 3D geometry inside diffusion model.


Paper

Youtube Video (55K Views)

Wonder how attention flows inside your Vision Transformer? What visual patterns are recognize by machine's attention? Does machine's attention resemble human's visual cognition? Collaboration with Catherine Yeh. My main contribution is the visualization of learnt attention in vision transformer models.


Paper

KALMUS is a Python package for the computational analysis of colors in films. It provides quantitative tools to study and compare the use of film color. This package serves two purposes: (1) various ways to measure, calculate and compare a film’s colors and (2) various ways to visualize a film’s color.

We aim to further improve the accuracy and clinical applicability of echocardiography video segmentation by extending the analysis from half-heartbeat (End-diastolic to End-systolic phases) to multi-heartbeat video. We proposed a sliding window data augmentation technique for efficiently learning moition tracking and semantic segmentation from sparesely annotated echo videos (only 2 annotations per video). Paper .

We assessed a 3D-UNet's performance on jointly segmenting sparsely annotated half-heartbeat echocardiography videos and estimating cardiac structure's motions. The 3D-UNet was trained on CAMUS dataset, and we evaluated its generalizability on Stanford's EchoNet dataset. Comparing with traditional frame-based segmentation method, our results show that the joint learning of motion tracking and segmentation enhances the segmentation performance on video data. The video model also has better generalizability on unseen dataset than frame-based model.

Designing a Dashboard for Transparency and Control of Conversational AI, https://arxiv.org/abs/2406.07882

Linear probe found representations of scene attributes in a text-to-image diffusion model

The implementation of CLAS-FV described in "Fully automated multi-heartbeat echocardiography video segmentation and motion tracking".

KALMUS Color Toolkit for color analysis and visualization in film studies.