Alright learning crew, Ernis here, ready to dive into some seriously cool research! Today, we're cracking open a paper about making computers smarter by helping them reason better using something called Knowledge Graphs. Think of Knowledge Graphs as massive digital webs of information, like a super-powered Wikipedia that understands how things are connected.
Now, these Knowledge Graphs are packed with information – not just facts, but also numbers and attributes. Imagine you're looking at a graph about movies. You'd see things like the movie title, the director, the actors, but also numerical data like the budget, the box office revenue, and the IMDb rating. Being able to reason with these numbers is super important.
The problem is, current methods, like Graph Neural Networks (GNNs) and Knowledge Graph Embeddings (KGEs), are like detectives who only look at the immediate neighbors of a clue. They're good, but they often miss the bigger picture – the logical paths that connect seemingly unrelated pieces of information. It’s like only looking at the fingerprints on a doorknob and missing the getaway car speeding away.
That's where ChainsFormer comes in. This is a brand-new approach that's all about tracing those logical paths, or "chains" of reasoning, within the Knowledge Graph. Think of it like following a breadcrumb trail to solve a mystery!
What makes ChainsFormer so special? Well, it does a few key things:
Builds Explicit Chains: Instead of just looking at immediate neighbors, ChainsFormer actively constructs logical chains of information.
Goes Deep: It doesn't just stop at one hop; it explores multiple steps in the chain, allowing for deeper, more complex reasoning.
Introduces RA-Chains: This is a special type of logic chain called "Relation-Attribute Chains" that model sequential reasoning patterns. Imagine following a chain like: "Movie A directed by Director B, Director B won award for Best Director, Best Director award given in year Year C." That's an RA-Chain in action!
Learns Step-by-Step: ChainsFormer uses a technique called "sequential in-context learning" to understand the reasoning process step-by-step along these RA-Chains. It's like learning a recipe one ingredient at a time.
Filters Out Noise: Not all chains are created equal. Some are misleading or irrelevant. ChainsFormer uses a "hyperbolic affinity scoring mechanism" to identify and select the most relevant logic chains. This is like sifting through clues to find the ones that really matter.
Highlights Critical Paths: Finally, it uses an attention-based numerical reasoner to pinpoint the most important reasoning paths, making the whole process more transparent and accurate.

"ChainsFormer significantly outperforms state-of-the-art methods, achieving up to a 20.0% improvement in performance."

So, why should you care? Well, this research has implications for a ton of different areas:
For the Techies: This is a big step forward in improving the accuracy and efficiency of knowledge graph reasoning, which is crucial for building more intelligent AI systems.
For the Business Folks: Better knowledge graph reasoning can lead to better recommendations, more accurate market analysis, and more effective decision-making.
For Everyone: Think about smarter search engines, more personalized experiences online, and AI assistants that can actually understand your questions. This research is helping to make that a reality.
The researchers have even made their code available on GitHub (https://github.com/zhaodazhuang2333/ChainsFormer), so you can check it out for yourself!
Now, this all sounds pretty amazing, right? But it also brings up some interesting questions:
How do we ensure that these "logical chains" are actually logical and not just based on biased or inaccurate data?
As these AI systems become more sophisticated, how do we maintain transparency and understand why they're making the decisions they are?
Food for thought, learning crew! Until next time, keep exploring and keep questioning!Credit to Paper authors: Ze Zhao, Bin Lu, Xiaoying Gan, Gu Tang, Luoyi Fu, Xinbing Wang

Alright learning crew, Ernis here, ready to dive into some seriously cool tech that's making our video-understanding AI a whole lot smarter! Today, we're unpacking a paper that tackles a tricky problem: How do we teach AI to really "see" what's happening in a video, not just identify objects?
Think of it like this: You're watching a movie scene where a character puts a key in a lock and opens a door. A standard AI might recognize the key, the lock, and the door. But does it understand the relationship between them? Does it grasp that the key caused the door to open? That's where things get complicated.
Turns out, even these fancy "Video-LLMs" (fancy talk for AI that can understand both video and language) struggle with this. They're not great at understanding spatial relationships (where things are in relation to each other), temporal ordering (what happens first, second, third), or cross-frame continuity (how things change smoothly from one moment to the next).
Imagine showing the AI a video of someone juggling. It might see the balls, the hands, and the person. But does it understand the pattern of the juggling? The cause and effect of the throws and catches? Probably not as well as we'd like.
That's where this awesome new framework called VideoPASTA comes in. Now, I know what you're thinking: "VideoPASTA? What's with the name?" Honestly, I don't know! But what I do know is that it's a clever approach to making these Video-LLMs much better at understanding video.
The core idea behind VideoPASTA is to train the AI to distinguish between good video understanding and bad video understanding. They do this by creating "adversarial examples" – basically, trick videos designed to fool the AI. These videos deliberately mess up the spatial, temporal, or cross-frame relationships.
Think of it like showing the AI a video where a glass magically floats off a table before someone touches it. It violates our understanding of cause and effect, right? VideoPASTA uses these kinds of "impossible" scenarios to teach the AI what shouldn't be happening.
"VideoPASTA trains models to distinguish accurate video representations from carefully generated adversarial examples that deliberately violate spatial, temporal, or cross-frame relations."
What's really cool is how they do this. They use a technique called "Direct Preference Optimization." It sounds complicated, but essentially, they're showing the AI pairs of video understandings: one good, one bad. And the AI learns to prefer the good one. What is impressive is that they only used around 7,000 pairs of videos, which is not a lot in the grand scheme of AI training.
And guess what? It works! The researchers tested VideoPASTA on some standard video benchmarks, and the results were impressive. The AI performed significantly better on tasks that required understanding spatial relationships, temporal ordering, and cross-frame continuity.
The paper highlights performance gains on benchmarks like VideoMME, NeXTQA, and LongVideoBench, improving over the baseline Qwen2.5-VL model. This shows the method's effectiveness in enhancing video understanding capabilities.
But here's the kicker: VideoPASTA achieves these improvements without requiring massive amounts of training data or complex architectural changes. In fact, it's incredibly efficient. They only used 32-frame sampling, compared to the 96-frame setups used by other researchers. This means it's a "plug-and-play" solution that can be easily integrated with existing models.
So, why does this matter? Well, for starters, it means we're getting closer to AI that can truly understand the world around us through video. This has huge implications for:
Robotics: Imagine robots that can understand complex tasks by watching videos.
Self-driving cars: Better video understanding means safer autonomous navigation.
Medical diagnosis: AI that can analyze medical videos to detect diseases earlier.
Content creation: Tools that can automatically generate summaries, captions, and even edits for videos.
This research offers a scalable and efficient way to improve video-language models. The targeted alignment with adversarial examples proves to be more effective than relying solely on large-scale pretraining or complex architectural modifications.
It really makes you wonder: Is targeted training more effective than just throwing tons of data at a problem?
Here are a couple of thought-provoking questions that come to my mind after reading this paper:
Could this same approach be used to improve AI's understanding of other types of data, like audio or text?
How can we ensure that these "adversarial examples" don't inadvertently teach the AI to be biased or discriminatory?
Credit to Paper authors: Yogesh Kulkarni, Pooyan Fazli

Hey PaperLedge crew, Ernis here, ready to dive into another fascinating paper! Today, we're tackling something super cool: editing text directly into images, even if that text needs to be twisted, turned, or warped to fit perfectly. Think of it like Photoshopping text onto a curved sign, but way smarter!
The paper introduces something called DanceText. Now, the name might sound a bit whimsical, but the tech behind it is seriously impressive. The core problem they're tackling is this: existing AI models can generate images with text, but they often struggle when you want to edit text that's already in an image, especially if you need that text to, say, curve around a bottle or slant along a building.
Imagine trying to change the label on a bottle of soda in a photo. Regular AI might just slap the new label on top, making it look flat and totally out of place. DanceText, on the other hand, tries to make the edit look like it was always there.
So, how does it work? The key is a clever, layered approach. Think of it like this: DanceText first carefully separates the text from the background image. It's like carefully cutting out a sticker from a page. Then, it applies the geometric changes – the rotations, scaling, warping – only to the text layer. This gives you much more control. Think of it like using a stencil where the text is on a separate layer and can be moved around and edited without affecting the background.
But that's not all! Just changing the shape of the text isn't enough. It also needs to blend seamlessly with the background. That's where their depth-aware module comes in. It figures out the 3D structure of the scene to make sure the lighting and perspective of the text match the background perfectly. It's like making sure the sticker appears to be part of the original image itself and cast the right shadows.
"DanceText introduces a layered editing strategy that separates text from the background, allowing geometric transformations to be performed in a modular and controllable manner."
The really cool thing is that DanceText is "training-free." This means it doesn't need to be specifically trained on tons of examples of text edits. Instead, it cleverly uses existing, pre-trained AI models to do its job. This makes it much more flexible and easier to use in different situations.
They tested DanceText on a big dataset called AnyWord-3M, and it performed significantly better than other methods, especially when dealing with large and complex text transformations. This means more realistic and believable edits.
So, why does this matter? Well, for artists and designers, this could be a game-changer for creating realistic mockups or editing product labels. For advertisers, it opens up new possibilities for creating eye-catching visuals. Even for everyday users, it could make editing text in photos much easier and more fun.
Think about the possibilities! Imagine quickly updating signage in a photo to reflect new information, or realistically adding custom text to a product image without any clunky Photoshop work.
Here are a couple of things that jumped into my head:
How easily could this be integrated into existing photo editing software?
Could this technology be adapted to edit other objects in images, not just text?
Food for thought, learning crew! Until next time!Credit to Paper authors: Zhenyu Yu, Mohd Yamani Idna Idris, Pei Wang, Yuelong Xia

Hey PaperLedge learning crew, Ernis here, ready to dive into some seriously cool science! Today, we're tackling a paper that looks at how things influence each other even when they're far apart – think of it like the butterfly effect, but on a more mathematical level.
So, what's this paper about? Well, imagine you're watching a flock of birds. They all seem to move together, right? Even though one bird can't directly tell every other bird what to do, there's a kind of collective behavior going on. This is similar to what scientists call nonlocal interactions. These are interactions where what happens in one place affects things in another, sometimes distant, place.
These nonlocal interactions pop up all over the place! From patterns forming in nature (like the stripes on a zebra) to how brain cells fire, and even how cells move around in our bodies. Scientists use math equations to try and understand these things, and often these equations include something called an integral kernel. Think of it as a recipe that describes how much one thing influences another, based on how far apart they are.
Now, here's the tricky part: these nonlocal equations are hard to solve! Because everything is connected to everything else, it makes the math super complicated. That's where this paper comes in. The researchers have developed a clever trick to simplify things.
Their idea is to approximate these nonlocal interactions with something called a reaction-diffusion system. Imagine you have a bunch of chemicals spreading out and reacting with each other. This is a local interaction – things only directly affect what's right next to them. The researchers found a way to show that certain types of nonlocal interactions can be mimicked by a bunch of these local reaction-diffusion systems working together!
Think of it like this: instead of a single, complicated network influencing everything at once (nonlocal), you have a bunch of smaller, simpler networks that pass information along step-by-step (local). It's like breaking down a big problem into smaller, more manageable pieces.
"Our results establish a connection between a broad class of nonlocal interactions and diffusive chemical reactions in dynamical systems."
The key to their approach is finding the right "recipe" (or kernel) that can be approximated by these reaction-diffusion systems. They focus on a specific type of recipe that can be broken down into simpler parts, called Green functions, especially in high-dimensional spaces.
So, why does this matter? Well, it makes it much easier to study these complex systems! By turning nonlocal interactions into local ones, scientists can use simpler mathematical tools to understand things like:
How patterns form in nature
How our brains work
How diseases spread
This research essentially builds a bridge between the world of nonlocal interactions and the more familiar world of local reactions and diffusion. It gives us a new way to think about and analyze these fascinating phenomena!
And that connection between seemingly different worlds of science is what makes this work so exciting. It's not just about simplifying equations; it's about uncovering the underlying connections that govern how things work in the universe!
But here are a couple of things I'm wondering about. If you're thinking about this too, let me know!

Could this approximation method be used to design new materials with specific properties, by controlling how things interact at a distance?


What are the limitations of this approach? Are there certain types of nonlocal interactions that can't be approximated in this way?

Credit to Paper authors: Hiroshi Ishii, Yoshitaro Tanaka

Alright learning crew, Ernis here, ready to dive into some fascinating research! Today, we're tackling a paper about how we can trust the answers we get from those super-smart AI language models, like the ones that write emails for us or answer our burning questions online.
Think of it this way: Imagine you're writing a research paper, but instead of hitting the library, you have a super-powered AI assistant. This assistant uses something called Retrieval-Augmented Generation, or RAG for short. Basically, RAG lets the AI look up information in a bunch of documents – like a digital library – and then use that information to answer your questions, with citations, just like a real research paper!
Now, here's the kicker: how do we know if the AI is actually telling the truth, or if it's just making things up? This is what researchers call hallucination, and it's a big problem. We want to make sure that the information in those citations actually supports the AI's answer.
This paper dives deep into how we can evaluate whether the AI's answer is backed up by solid evidence. They looked at something called the TREC 2024 RAG Track, which is like a big competition where different teams submit their RAG systems. The researchers compared how well an AI judge (GPT-4o, a really powerful version of GPT) agreed with human judges on whether the AI's answers were supported by the cited documents.
Imagine it like this: you have a statement, say "Dogs make great pets because they are loyal." Now you have a source document that says "Dogs are known for their unwavering loyalty to their owners." Does the source document support the statement? That's the sort of thing these judges, both human and AI, are trying to determine.
They did this in two ways:
From scratch: Human judges read the AI's answer and the cited document, and then decided whether the document supported the answer.
Post-editing: The AI judge gave its opinion first, and then the human judges could either agree with it or change it if they thought it was wrong.
So, what did they find? Well, in over half the cases (56%), the AI judge (GPT-4o) and the human judges agreed perfectly from the start! And when the human judges could edit the AI's predictions, they agreed even more often (72%). That's pretty impressive!
But here's the really interesting part. The researchers found that when the human and AI judges disagreed, another independent human judge actually agreed more often with the AI judge than with the original human judge! This suggests that the AI judge might actually be pretty good at this, maybe even as good as, or in some cases better than, human judges at determining support.
The researchers concluded that "LLM judges can be a reliable alternative for support assessment."
Why does this matter?
For researchers: This helps us understand how to build better AI systems that are more trustworthy.
For businesses: This could lead to better AI-powered tools for research, customer service, and more.
For everyone: As AI becomes more and more integrated into our lives, it's crucial that we can trust the information it provides.
This research is a step towards making AI more reliable and transparent. By understanding how well AI can assess its own answers, we can build systems that are less prone to errors and more helpful to everyone.
So, what does this all mean for the future of AI? Here are a couple of questions that popped into my head:
Could we eventually rely solely on AI judges for tasks like this, freeing up human experts to focus on more complex problems?
How can we ensure that these AI judges are fair and unbiased, especially when dealing with sensitive topics?
That's all for today's deep dive, learning crew! Stay curious, and keep questioning!Credit to Paper authors: Nandan Thakur, Ronak Pradeep, Shivani Upadhyay, Daniel Campos, Nick Craswell, Jimmy Lin

Hey PaperLedge crew, Ernis here, ready to dive into another fascinating piece of research! Today, we're tackling something that's becoming super relevant in our increasingly digital world: teaching AI to write better code.
Think of those fancy AI tools that can whip up code for you - code-generating Large Language Models (LLMs). They're like having a super-helpful, if sometimes a little quirky, coding assistant. This paper explores how we can make these assistants even better.
The core idea is to use a technique called Reinforcement Learning. Imagine training a dog: you give it treats when it does something right. Reinforcement Learning is similar. The AI generates code, and then gets feedback on how good that code is. This feedback helps it learn to write even better code next time.
Now, the tricky part is how we give the AI that feedback. That's where Direct Preference Optimization comes in. Instead of just saying "good" or "bad," we're basically saying, "This version of the code is better than that version." It's like showing the AI two different answers to a problem and letting it figure out which one is superior.
But here's where things get really interesting. The researchers realized that the data they were using to train the "feedback giver" (what they call the reward model) wasn't as good as it could be. It was like trying to teach the dog based on incomplete instructions. So, they used a cool technique called symbolic execution to create a more comprehensive and objective dataset. Think of symbolic execution like running the code in a simulated environment, exploring all the possible paths and outcomes.
Imagine you are testing a math problem:
You can solve it step by step with real numbers to check if your program gives the right answer.
Or you can use symbolic execution to solve all the different possible paths of the code to check it.
The benefit is it allows you to test every single corner and edge case that your program can have, making it more robust.
This is important because with better data, the reward model becomes a much better "judge" of code quality. And a better "judge" means the AI can learn to write even more efficient and bug-free code.
"With symbolic execution, we create a custom dataset that better captures the nuances in code evaluation."
So, what did they find? Well, the reward models trained with this new, improved data were significantly better at judging code quality compared to previous methods. And, the code-generating AIs trained using this feedback were able to achieve similar performance to a well-established benchmark called CodeRL. This means they're on the right track to building truly powerful coding assistants.
Why does this matter?
For developers: This could mean less time spent debugging and more time building amazing things.
For businesses: Faster software development translates to faster innovation and a competitive edge.
For everyone: More efficient and reliable software powers everything from our smartphones to our cars.
Now, this raises some interesting questions for our discussion:
If AI can write code, what does this mean for the future of programming jobs? Will programmers become more like "AI wranglers," guiding and refining the code generated by these models?
Could this technology be used to create more accessible and inclusive coding tools, allowing people with less technical expertise to build software?
What are the ethical implications of using AI to generate code? Could it lead to unintended consequences, like the creation of malicious software or the perpetuation of biases?
I'm eager to hear your thoughts on this research, PaperLedge crew! Let's dive in and explore the exciting world of AI-powered coding.Credit to Paper authors: Marina Sakharova, Abhinav Anand, Mira Mezini

Hey PaperLedge learning crew, Ernis here, ready to dive into some fascinating research that tackles a real head-scratcher in the world of AI. We're talking about Large Language Models, or LLMs – those brainy algorithms powering things like ChatGPT. They're amazing at general knowledge, but what happens when you need them to be experts in, say, rocket science or tax law? That's where things get tricky.
The paper we're unpacking today is all about making these powerful LLMs even more powerful by giving them a smart study buddy. Think of it like this: imagine you're putting together a presentation on a complex topic. You might start with a basic outline from a classmate who's got some background knowledge, and then you, with your broader understanding, take that outline and turn it into something truly spectacular. That's the essence of what this research is doing with LLMs.
See, fine-tuning these giant LLMs for every single specialized task is like trying to teach a golden retriever every single trick in the dog training manual. It's expensive, time-consuming, and sometimes just plain impossible, especially when we don't have full access to the inner workings of these models – they're often "black boxes".
So, these researchers came up with a clever workaround: a collaborative framework. They pair a strong, general LLM (the one with all the broad knowledge) with a weak, specialized model (the one with deep expertise in a specific area). The weak model acts like that classmate, generating initial drafts and background info relevant to the task at hand. Then, the strong model steps in, using its advanced reasoning skills to polish, refine, and expand on that foundation. It's like having a junior researcher give you the groundwork, and then you, the senior researcher, bring it all together.
Think of it like this:
Weak Model: A specialist doctor who deeply understands one rare disease but has limited general medical knowledge.
Strong Model: A general practitioner with broad medical knowledge but lacks the specialist's in-depth understanding of the rare disease.
Collaboration: The general practitioner consults with the specialist, leveraging their combined knowledge to provide the best possible diagnosis and treatment plan for the patient.
But here's the really cool part: the researchers didn't just leave it at that. They developed a way to give the weak model feedback, so it gets better and better at helping the strong model. They call it "collaborative feedback." Essentially, it's a system that figures out how much the weak model's contributions actually influenced the final result, and then uses that information to guide the weak model's learning. It's like saying, "Hey, weak model, that paragraph you wrote was really helpful in getting the strong model to the right answer. Do more of that!"
This is achieved using preference pairs which tell the weak model, "This output was better than that output in terms of how well it helped the stronger model achieve the final result."
"By leveraging complementary strengths, the collaboration significantly outperforms each model alone."
The researchers tested this framework across three different areas, and the results were impressive. The collaborative approach consistently outperformed either model working alone. And, even more impressively, tuning the weak model using this collaborative feedback boosted performance even further. This means the system wasn't just good; it was getting better over time.
So, why does this matter? Well, for starters, it offers a way to extend the capabilities of LLMs without requiring massive amounts of computing power or access to the inner workings of these models. This is huge for businesses that want to use LLMs for specialized tasks but don't have the resources to fine-tune them from scratch. It's also important for researchers who want to explore the potential of LLMs in different domains.
But beyond that, this research highlights the power of collaboration in AI. It shows that by combining the strengths of different models, we can create systems that are more powerful and adaptable than any single model could ever be on its own. This has implications for how we design AI systems in the future, suggesting that a collaborative, modular approach might be the key to unlocking even greater potential.
This study has got me thinking...
Could this collaborative approach be applied to other types of AI systems, not just LLMs?
How could we design even more effective ways to provide feedback to the weak model, so it learns even faster?
Does this strategy reinforce existing knowledge biases or help to overcome them?
I'm really curious to hear your thoughts on this one, learning crew! Let me know what you think in the comments. Until next time, keep learning and keep exploring!Credit to Paper authors: Yizhu Jiao, Xuchao Zhang, Zhaoyang Wang, Yubo Ma, Zhun Deng, Rujia Wang, Chetan Bansal, Saravan Rajmohan, Jiawei Han, Huaxiu Yao

Alright learning crew, Ernis here, ready to dive into some seriously cool tech! Today, we're talking about how to make construction sites safer and more efficient using...wait for it...exoskeletons powered by AI brains!
Now, imagine a construction worker. They're constantly moving, lifting heavy things, climbing ladders – it's a tough job. And unlike a robot on an assembly line, their environment is constantly changing. That means wearing an exoskeleton, those robotic suits that help you lift and move, can be tricky. The suit needs to know what the worker is about to do to provide the right kind of assistance.
That's where this research comes in. These researchers asked a really important question: How can we get exoskeletons to anticipate what a worker is going to do before they do it, so the suit can provide the right support at the right time?
Their solution? They built an AI "brain" for the exoskeleton, using the same kind of tech that powers ChatGPT – Large Language Models or LLMs. But they didn't stop there; they gave it a memory too!
Think of it like this: imagine you're teaching a dog a new trick. At first, you give very clear commands: "Sit!" and you might even physically help them. But over time, the dog learns. You can use shorter commands or even just a gesture, and the dog remembers what to do because they have a short term memory and a long term memory.
That's what this AI does. It uses a few key parts:
Perception Module: This is like the AI's eyes and ears. It uses smart glasses to "see" what the worker sees and "hear" what they say – even simple spoken commands.
Short-Term Memory (STM): This is like the AI remembering what just happened. Did the worker just pick up a brick? That influences what they're likely to do next.
Long-Term Memory (LTM): This is where the AI stores information about the worker's habits and the general tasks they're performing. For example, it might learn that when a worker says "mortar," they're likely about to lay bricks.
Refinement Module: This part takes all the information and makes the best guess about what the worker is going to do next.
So, how well does it work?
The researchers tested the AI by having it predict what the worker would do next. Without any memory (just the perception module), it was right about 73% of the time. Not bad, but not great. Adding the short-term memory boosted it to 81%. But the real magic happened when they added both short-term and long-term memory. The AI was then able to predict the worker's actions correctly a whopping 90% of the time!
What's really impressive is that it did especially well with commands that were vague or related to safety. For example, if the worker said "Careful!" the AI was better able to predict what kind of hazard they were responding to.
They also measured how confident and accurate the AI was in its predictions. They found that by adding the short term and long term memories, the AI's predictions became much more reliable and trustworthy. This is super important because we want the exoskeleton to only assist when it's really needed.
So, why does all this matter?
This research is a big step towards making construction sites safer and more efficient. By anticipating a worker's needs, exoskeletons can provide support exactly when it's needed, reducing strain and preventing injuries. Plus, workers can focus on their tasks without having to constantly adjust the exoskeleton.
But it's not just about construction. This technology could be used in all sorts of dynamic industries, from manufacturing to disaster relief. Imagine firefighters wearing exoskeletons that anticipate their movements as they navigate a burning building, or warehouse workers effortlessly lifting heavy boxes all day long!
This research points to a future where humans and machines work together seamlessly, each enhancing the other's capabilities.
Here are some things that crossed my mind:
How do you ensure the AI doesn't become too reliant on past behavior and miss something new or unexpected? What safety measures are in place to prevent the exoskeleton from making a wrong move?
Could this technology be adapted to other wearable devices, like augmented reality headsets, to provide real-time information and guidance to workers?
What are the ethical considerations of using AI to predict human behavior in the workplace? How do we protect worker privacy and autonomy?
That's all for today, learning crew! Until next time, keep those neurons firing!Credit to Paper authors: Ehsan Ahmadi, Chao Wang