Xijun Wang

PhD student @umdcs @gammaumd Applied Scientist intern @amazon/a9

[GAMMA UMD @gammaumd]

🎓 We’re thrilled to announce that Xijun Wang @xijunwang_cs, Tianrui Guan @terryguan97, and Pooja Guhan @GuhanPooja have successfully defended their PhD theses! Huge congratulations to all three. Your hard work and dedication have truly paid off. We’re so proud of you! 👏🥳

[Xijun Wang @xijunwang_cs]

LLMs know how to make use of short cuts! This is a good exploration on the potential LLM redundancy. Check our paper for more details!

[Pedro Sandoval @psandovalsegura]

Attention sinks in LLMs are weird. There’s ~20% of heads that don’t seem to do anything.

Do these heads matter? Turns out that if we get rid of them, benchmark scores don’t change.

[Pedro Sandoval @psandovalsegura]

More details are in our paper, now on arXiv: https://t.co/CzU6q46n7V

Thanks to my collaborators: @xijunwang_cs, @PandaAshwinee, @micahgoldblum, @RonenBasri, @tomgoldsteincs, @davwiljac

[GAMMA UMD @gammaumd]

AUTOHALLUSION: Automatic Generation of Hallucination Benchmarks for VLMs
@wu_xiyang*, @terryguan97*, D Li, S Huang, X Liu, @xijunwang_cs, @RuiqiXian, @abhi2610, @furongh, @boydgraber, @zhoutianyi, @dmanocha

11/3; 16:00-17:30 - Riverfront Hall
Website: https://t.co/XCy4Xt2gze

[GAMMA UMD @gammaumd]

.@xijunwang_cs will be presenting “SCP: Soft Conditional Prompt Learning for Aerial Video Action Recognition”, which utilizes prompt learning for aerial video action recognition. ✈️

Thurs Oct 17
17:30 - 17:45
Room 6

[Xijun Wang @xijunwang_cs]

#ECCV2024 Visit our ViLA poster and Say Hi to Prof.
@MingCLinCS !

Time: October 3, 4:30pm - 6:30pm
Location: Poster session 6, ID 274

Arxiv: https://t.co/LjIpvuBBLy
Code: https://t.co/yohW1Rp0AL

Big shout-out to @shanyangmie!
@gammaumd

[GAMMA UMD @gammaumd]

Xijun Wang @xijunwang_cs will be presenting "ViLA: Efficient Video-Language Alignment for Video Question Answering", which addresses both efficient frame sampling and effective cross-modal alignment! 🗣️

Time: October 3, 4:30pm - 6:30pm
Location: Poster session 6, ID 274

[Xijun Wang @xijunwang_cs]

@laurayuzheng sounds like latte without coffee and milk☠️

[Xiyang Wu (On the Job Market) @wu_xiyang]

📢Thrilled to share our latest work #AutoHallusion, an automatic hallucination pipeline that scales up the benchmark generation process.

🌐 Webpage & Examples: https://t.co/7wnpjh1gpV… 📷 Arxiv: https://t.co/cpyfebl3jg

👀Please stay tuned for our future updates!

Thanks a lot for co-leading
@terryguan97
and other wonderful collaborators!
@cookielee77
@ShuaiyiH
@xiaoyu_liu_1231
@xijunwang_cs
@RuiqiXian
@abhi2610
@furongh
@boydgraber
@zhoutianyi
@dmanocha
from
@umdcs
@umiacs
@ml_umd
@gammaumd
@UMDResearch
@UMDscience
#VLMs #hallucination #benchmark #gpt4v #Gemini #Claude #LLaVa #CognitiveScience #brain

[Xijun Wang @xijunwang_cs]

@rohanchandra30 @UVA Congratulations 🎉🎊

[Xijun Wang @xijunwang_cs]

@laurayuzheng @Waymo Awesome! Have a good time and make a splash there!👍👍

[Xijun Wang @xijunwang_cs]

Join us at #93 of the Poster session on Thursday, February 22! #AAAI24 @gammaumd

Say hi to @shanyangmie there!

Our ICAR can recommend items with good visual compatibility including similarity (color, geometry, texture, etc.) and complementarity (like table vs chair).

[Ruiqi Xian @RuiqiXian]

Join us at Naupaka #74, Saturday 17:15-19:15! #WACV2024

We proposed PMI Sampler for aerial footage, it efficiently selects key frames using patch-wise mutual information. Ideal for moving camera footage with dynamic backgrounds.

@xijunwang_cs @DivyaKRaman1 @dmanocha @gammaumd

[Xijun Wang @xijunwang_cs]

See you Saturday 17:15-19:17 at Naupaka #51 #WACV2024

For aerial videos, we proposed MITFAS to focus on the regions corresponding to salient motions and find the more informative frames by using mutual information.

@RuiqiXian @dmanocha @gammaumd

[Xijun Wang @xijunwang_cs]

@laurayuzheng @saizhang0 Hahaha, 4 for me😹

[Tianrui Guan @terryguan97]

📢 Sharing a couple of interesting observations and results of #GeminiAI Pro Vision on our #HallusionBench (Big thanks to @GoogleDeepMind for making those API available! 🙏):

1. Language hallucination is a significant issue, often resulting in outputs that include irrelevant information not found in the question prompt or the accompanying image. A few examples are provided.

2. In terms of accuracy on #HallusionBench, Gemini Pro Vision demonstrates inferior performance compared to #gpt4V and #LLaVA, as detailed in the Table 2 included.

3. Our benchmark results indicate that Gemini Pro Vision exhibits the least bias between 'Yes' and 'No' responses compared to all other models tested, including #gpt4V. See the Table 3 included.

4. Accessibility -- All of the results are obtained using Gemini API. We are facing some "internal error" issue while using Google AI Studio on the web page.

I can't wait to see the release of Gemini Ultra and evaluate whether those problems are alleviated or fixed. Keep an eye on our GitHub page (https://t.co/q2FFw3g42Q) for updates on more results, and stay tuned for the arXiv update as those models are released!

[Fuxiao Liu @Kongkong]

🔥CoBig shoutout to @Google for launching their groundbreaking large multimodal model #Gemini.
🚩However, there are still obvious hallucinations with Gemini. Here are a few examples with Gemini Pro.

Want to learn more? Look at our HallusionBench Paper: https://t.co/3qjsR3Drh7

[Xijun Wang @xijunwang_cs]

How to strengthen both CNNs and Transformers?

Check our work “SCSC: Spatial Cross-scale Convolution Module to Strengthen both CNNs and Transformers” on Oct. 2nd (1:30pm-6pm) at #ICCV2023 Workshop on New Ideas in Vision Transformers
@room P01

ArXiv: https://t.co/Ozc5PGk7zB

[Xijun Wang @xijunwang_cs]

Big thanks to the leading author @terryguan97 and all the coauthors!

[Aswath Muthuselvam @selvamaswath]

How to localize a robot(equipped with a LiDAR) in a Large-scale point cloud map?

We present "CrossLoc3D - Aerial-Ground Cross-Source 3D Place Recognition"

Code: https://t.co/nuDdbW9UWj
PDF: https://t.co/jYZQzRTg6B

#ICCV2023
🧵

[Xijun Wang @xijunwang_cs]

@deedzgao @JaminFong 👍