@doomie @willccbb Another maybe important question: which 75% training samples are selected? Are they randomly selected?

That's actually not true. It uses JSON representation to represent all the data. ```messages=[ {"role": "user", "content": [ {"type": "text", "text": prompt}, {"type": "image_url", "image_url": { "url": f"data:image/png;base64,{base64_image}"} } ]} ],```. It requires user to encode the raw image as base64 and then converts this base64 string into a UTF-8 string. e.g. ```import base64 def encode_image(image_path): with open(image_path, "rb") as image_file: # Read the image in binary mode return base64.b64encode(image_file.read()).decode("utf-8") ```. I think this can later be read using OpenCV.

@chriswolfvision Follow your analogy, switching to Camtasia is like hopping onto an electric bike: straightforward, smooth, and still powerful enough to get you where you need to go without breaking a sweat. No need for pilot training—just start pedaling and enjoy the ride!

@nntsn Congratulations on the new role! I know nothing about nuclear fusion. Which introduction machine learning papers do you recommend to read that is also related to nuclear fusion? Or r u going to write one?

@YiMaTweets I want to agree. But how can you prove that claim? One direction could go wrong is that the existing mathematical languages that human invented are not sufficient to explain the progress made in the AI through “engineering hacks”.

@mtrainier2020 感觉主要是因为，千里马常有而伯乐不常有。可能这里伯乐就是工程师。当前AI最火的一个方向，diffusion model就是基于一个微分方程求解的论文，BDO Anderson 1982. 那个paper的引用几十年都是几十个，无人问津，图像生成sora出来之后，开始暴涨。可能���力学好微积分某天也能用来做Netflix电影。

Text2image is the coolest app in town! 🚀 Just when I thought trainable MRFs were irrelevant, they prove otherwise, making text2image inference efficient/cheaper! Check out this #CVPR2024 paper: MarkovGen: Structured Prediction with Trainable MRFs.

Excited to head to Seattle for #CVPR Sun-Fri! Looking forward to connect and chat about the latest in foundational models and autonomous robotics. DM me to meet up—looking forward to catching up with old friends and making new ones!

Love the Google Scholar PDF Reader for its quick access to paper abstracts while reading! It's great but lacks support for arXiv links like chrome-extension://xxx/ Currently, you must download and reopen the PDF to activate this feature.

My toddler ate paper today. To convince him that paper isn’t food, I used #DALLE3 to generate some cute images. They’re quite convincing. Check them out! The second one isn’t as compelling as the first.

@chriswolfvision Is this using the same trick like the good old integral image in HOG?

I asked chatGPT what are the favorite CV/ML papers before 2010. Turns out he mentioned: Eigenfaces for Recognition (1991), A Few Useful Things to Know About Machine Learning (1997), t-SNE (2008), Conditional Random Fields for Object Recognition (2005), and SIFT (1999).

@ducha_aiki Definitely “ A discriminatively trained, multiscale, deformable part model” Cvpr 2008.

@eric_brachmann “ Wer mit Ungeheuern kämpft, mag zusehn, dasser nicht dabei zum Ungeheuer wird. Und wenn du lange in einen Abgrund blickst, blickt der Abgrund auch in dich hinein.” Lol

@ftm_guney Very cool work! A bit disappointed, no discussion like unknown unknown. Would be useful to have that uncertainty measure.

Yes, with 8 USD per month, @elonmusk twitter needs have at least markdown/latex support, so that we can tweet code/equation and more.

@sqcai It is not totally new, people can be IEEE paid member, with 210+ USD per year fee. That is 17.5 USD per month.

I really like parrots, who can speak whatever language after some training. Maybe it is a path for AI. Checkout the latest workshop @icdm2022 Foundation Models in Vision and Language and our paper (w code)

Holger Caesar from @tudelft covers the talk titled autonomous vehicles from imperfect and limited labels. A new autonomous vehicle dataset called nuPlan is introduced in this talk.

Yu Cheng from @MSFTResearch presents the talk title Towards data efficient vision-language (VL) models. It covers the methods such as FewVLM and Grounded-FewVLM.