news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Falsifiability is both an angel and a devil. Ultimately, falsifiability is a handshake opportunity for the world to align.

Author: Zhang Yijia Sukhoi

Ten years ago, a European academician said to me:There are no less than a hundred sub-tracks in artificial intelligence, and the United States will define one or two popular tracks, and then entrepreneurs and investors around the world rushed to catch up with these one or two tracks.”

In the decade that followed, the trend of artificial intelligence changed, from non-consensus to small consensus, and from small consensus to large consensus, but it never escaped this sentence.

This conversation with Serge began with a discussion of a paper he co-wrote two years ago.Searching for structure in unfalsifiable claimsUnfortunately, this article has received little attention so far. Google ScholarCitation count: only 2 times。

This is surprising.

"Jia Zi Guang Nian" believes thatthis essaySeriously underestimated。

Reason one: the problem this article faces is extremely critical (the underlying narrative of unfalsifiable social media data);

Reason two, this article provides important theoretical innovations (ternary annotation method and SNaCK set);

Reason three, this article provides practical engineering results (data set construction and detailed experimental comparison).

Authors: Peter Ebert Christensen, Frederik Warburg, Menglin Jia, and Serge Belongie; ARXIV2022

This paper is not difficult to understand.FalsifiabilityThis concept begins to be introduced.

Falsifiability is also called refutability. Philosophy of science often uses strictFalsificationTo judge whether a theory is scientific,“These conclusions must allowlogicThe existence of counterexamples to。

Karl Popper proposed in 1934 that a theory or hypothesis is valid if it can be logically disproven by empirical tests of existing techniques.FalsifiableIf an expression is so impeccable that the world has no room for comment, it will often only make everyone stay away from it, which is not helpful for promoting scientific progress.

The problems that scientists can or should study must have more or less loopholes, that is, they give others the opportunity to attack or even overturn them. The purpose of falsifiability is to make the theory predictive and testable.So it is useful in practice。

Serge's PaperSearching for structure in unfalsifiable claimsdiscussedThe complexity of interpreting unfalsifiable claims on social media。

Thesis Summary:

Social media platforms are flooded with posts and comments, and many claims cannot be disproven.Inadequate fact-checking tools, lack of structure in social network discussions, difficulty identifying narratives, and lack of quality in public discussionsetc., causing a lot of trouble.

The paper studies how to identify and understand thoseUnfalsifiable claims, and summarizes these claims asA limited number of narratives, in order to better facilitate discussion and debate on social media.

Interestingly, the author constructed aPAPYERA dataset containing 600 short text excerpts, 31 narratives, and 4 supercategories on the debate over hand drying methods in public restrooms to understand and discover dominant narratives in online discussions.

This paper introducesA new approach that goes beyond the capabilities of existing fact-checking technology, provide an important contribution to managing and understanding the impact of unfalsifiable claims in digital communication environments - using this pipeline to discover dominant narratives, and showing that this pipeline outperforms recent large-scale transformation models and state-of-the-art unsupervised topic models.

through experiments,The authors foundUse Modern Sentence Converter(such as T5 model)Initial Sentence Embeddingis the keyThey also found that the sampling strategy is crucial to generate high-quality embeddings, especially“Distance-Rnd” strategyBest performance.

Experimental results show that combining human annotationsTripletCan reveal complianceCrystallized NarrativeInteresting clustering of .

Only 2 citations

"Jia Zi Guang Nian" believes that the possible reasons why this paper has not received much attention in the industry so far include but are not limited to:

(1) The theoretical analysis is relatively thin, and the analysis of experimental results remains qualitative (only half a page of formulas in the 11-page text);

(2) There is almost no detailed introduction to the algorithms used for comparison;

(3) There may not be a unified academic data set in this field, which results in the lack of “out-of-circle” in the academic community;

(4) The author emphasized the importance of T5 but did not clearly describe its algorithmic superiority.

Serge Belongie's academic sharing at CVPR2024, source: "Jia Zi Guang Nian"

Although the above papers are little known, Serge himself is an influential scientist in the field of computer vision and machine learning, mainly studying object recognition and image segmentation. His papers have been cited a total of 178,971,000 times.

Serge Belongie is a professor of computer science at the University of Copenhagen and director of the Danish Pioneer Centre for Artificial Intelligence. Previously, he was associate dean and Andrew H. and Ann R. Tisch Professor of Computer Science at Cornell Tech.

The most worth introducing isSerge is the main author of MSCOCO。

The MSCOCO dataset is one of the most famous large-scale datasets for computer vision.In 2000, Serge and Jitendra Malik (now a professor of computer science at the University of California, Berkeley and a famous scholar in the field of computer vision)Together they proposed the concept of "Shape Context".It is a shape feature description method that is widely used in the fields of computer vision and object recognition.

In 2004, Serge was named a young technology innovator under 35 by MIT Technology Review; in 2007, he and Jitendra Malik received an honorary nomination for the Marr Prize; in 2015, Serge won the ICCV Helmholtz Prize, which is mainly awarded to authors of papers that have made fundamental contributions to the field of computer vision.

Serge is also the co-founder of several companies including Digital Persona (merged with CrossMatch in 2014), CarCode (acquired by Transport Data Systems), Anchovi Labs (acquired by Dropbox in 2012), and Orpix.

Currently, Serge TeamOpening up new dimensions of social network analysis——Starting from a large number of trivial statements that have not been paid attention to before and are not suitable for traditional fact-checking,Analyze the topic setting and“Narrative manipulation”。

This is of special significance at this point in time:

After the World Anti-Fascist War, no matter whether technological breakthroughs or bottlenecks alternate, they all go through ups and downs in the historical picture scroll with the passage of time. Just like the "Along the River During the Qingming Festival" unfolding in the time and space coordinate system, it is filled with thousands of scenes and all kinds of people from ancient times to the present.

The following is a conversation between Ms. A and Serge.

Follow the public account "Jia Zi Guang Nian" and reply "falsifiable" to obtain the information mentioned in the article.Searching for structure in unfalsifiable claimsandMSCOCO datasetTwo papers.

1. Falsifiability is being challenged

Researchers in practice are often influenced by narratives they like or dislike – somewhat similar to trending topics on Instagram.

Ms. A:"If it cannot be falsified, it cannot be science" has become a general consensus in the scientific community. However, many philosophers have questioned this, believing that the principle of falsifiability may lead to endless scientific debates.Is falsifiability a necessary condition for scientific progress?

Serge：According to popular opinion, a scientific theory must be falsifiable.

Ms. A:This is a popular view, but is falsifiability the prevailing paradigm today?

Serge：The machine learning literature has exploded in the past 15 years, with a large number of papers published and cited every day. In these papers, the relevant parts of the work often cite other papers, but the citations may not be the most relevant to their work. This is because the number of papers is huge.The researchers are actually responding to the dominant narrative in the field.

We usually think of ourselves as scientists in the tradition of Karl Popper, influenced only by falsifiable assertions. However, there are also trends in scientific research, such as techniques like Generative Adversarial Networks and Transformers. Although the goal of these papers is to follow the scientific tradition,Researchers in practice are often influenced by narratives they like or dislike——It’s a bit like trending topics on Instagram.。

Ms. A:Do you mean that, since machine learning, scientists have begun to deviate from the norm of falsifiability?

Serge：Scientists often claim to be unaffected by these influences and to be objective, but they are human after all and can be swayed by these popular opinions.This is something we consider not scientific, but more of a gut feeling and opinion.

Ms. A:How do you define unfalsifiable claims in social media?

Serge：We first need to discuss the literature on fact-checking. For example, Professor Isabelle Augenstein of the University of Copenhagen has developed a method that starts with determining the fact-checking value of a claim. We take a claim and put it up for verification.And determine its verification value in the range of 0 to 1。

For example, the statement that the capital of California is Sacramento is well suited for grammar and syntax checking because it can be found in multiple structured knowledge bases. We can check a statement like this: "The capital of California is Sacramento" and give it a verifiability score that might be close to 0.99. We then submit it to the structured knowledge base to confirm the answer. This deep learning-based verifiability system processes a large number of statements and training data to evaluate the verification value of different statements.

But some statements,For example, "Immigration to California is a bad thing" reflects more personal opinions and is not suitable for fact-checking.In contrast, statements such as “The number of immigrants in California has continued to increase since 2020” have a high verification value.

soWe are particularly concerned about claims that are difficult to verify.——These claims cannot be directly verified, but the discussion they have sparked on social media is significant.。Multiple checks may help us make better judgments.

Ms. A:In your research, what specific techniques or tools are used to identify and analyze unfalsifiable claims?

Serge：We use Natural Language Processing (NLP) techniques, clustering and grouping algorithms, and machine learning methods.

our target isCreating a Global Narrative Information Facility (GNIF), to research and organize social media content。

The combination of these techniques and tools enables us to better understand and process large amounts of narrative content.Indirectly helps identify unfalsifiable claims。

We are able to analyze all forms of text.Whether it's a tweet or a Reddit comment, we use NLP techniques to extract and understand the narratives and themes in these contents.

Second, we usedClustering and grouping algorithmsThese algorithms help us organize large amounts of social media content into distinct themes or narratives.

For example,We can find that out of millions of tweets, there are thousands of tweets that are very similar because they are all addressing the same basic narrative.

passNarrative clustering and assertion grouping, we organize large amounts of content into smaller clusters, allowing fact-checkers to process them more efficiently without having to check each item one by one. This way, even unfalsifiable claims can be identified and categorized through clustering and grouping for further analysis and processing.

We take two inputs, say two tweets, and measure their similarity based on different narrative aspects —These might cover topics such as the debate between nuclear power and green energy, or discussions of baby formula versus cow's milk.

There are many hotly debated topics online, often the result of disinformation campaigns。These activities can be very ambiguous. What we are trying to understand is how these different statements manifest themselves in the form of language or memes, which may contain images, text, audio statements, etc.Looks like a completely different piece of contentYou may have collected millions of discussions about a topic on a social media platform, but all the data may only contain dozens of opinions. We try to understand these phenomena through large language models, deep metric learning and other technologies.

Visualization of human-annotated pairs. Sub-image (a) shows positive pairs, i.e., pairs of similar or consistent narratives annotated by humans. Sub-image (b) shows negative pairs, i.e., pairs of dissimilar or inconsistent narratives annotated by humans. Source: Searching for Structure in Unfalsifiable Claims

2. Beyond “True or False”

Not all claims are worthy of fact-checking, and not all fact-checks result in true or false results.

Ms. A:The MSCOCO dataset you created is one of the most famous large-scale datasets for computer vision. How did you get started?

Serge：We started object detection research 15 years ago with a small dataset CUB200 containing more than 200 bird species. The COCO dataset was originally a summer internship project of my PhD student Tsung-Yi Lin at Microsoft Research, and his mentor at the time was another PhD student of mine, Piotr Dollá. This project gradually evolved into a consortium of researchers from academia and industry. They wanted to create a dataset that could depict everyday objects in natural environments in detail, and accurately annotate the names and spatial locations of these objects.

Ms. A:You named the dataset MSCOCO. I like Coco very much, and its English name is also Coco.

Serge：Yes, we all like the name "COCO", it is fun and easy to remember.

Ms. A:After the emergence of the MSCOCO dataset, the development of the field of computer vision has taken off like a rocket.

Serge：Well, we've organized a growing knowledge community around it, and COCO has been used by millions of people.We started small and ended up with a field of research that had a profound impact.

The first computer vision conference I attended was CVPR 1994, also in Seattle. That was thirty years ago, and there were about 300 people in attendance. Now, in 2024, there are 12,000 people in Seattle.

Ms. A:It's been 30 years. What drives your continued passion for computer vision and artificial intelligence research?

Serge：For as long as I can remember, I have been interested in patterns and categorizing things. In middle school, I did a class project on categorizing screws, bolts, and other fasteners. In college, I became interested in audio patterns, especially bioacoustics, such as bird or whale calls.When it comes to images, fingerprints and faces fascinate me.

I worked on lip reading from video. I was fascinated by all aspects of the problem: the fusion of audio and visual, the differences between different speakers, and the computational challenges. In the early 90s, digital cameras were just coming out, but they didn’t have any form of computational understanding. Today, you might take for granted face detection boxes in the viewfinder or photo album software that can intelligently organize your entire family photos, but that didn’t exist back then.

I felt at the time that there would be a huge demand for this technology.I also like the math behind the technology.I like the technology used in these fields, but I don't want to major in math or physics. For example, complex mathematical methods are used to solve problems in sound, video, and image processing.

I always feel that my mission in this world is to do this kind of work.

Ms. A:What academic presentations did you make at this year’s CVPR?

Serge：My team submitted several papers at the main CVPR conference, and I also gave presentations at two workshops. One of the presentations was about the history of computer vision research, mainly to help young scholars understand classic computer vision techniques, that is, techniques before deep learning and transformers. I also introduced the Visipedia project, which started with an extended version of the CUB200 dataset released in 2011. Currently, the research content of Visipedia has expanded to tens of thousands of plants, animals, and fungi, providing an important research foundation for object recognition in nature.

Another report is what I want to focus on sharing in today's interview. It is related to narrative, public opinion, and false information, especially in the context of the development of social media.

Ms. A:What innovations has your work brought to this field?

Serge: The classic problem in the world of disinformation and social media is fact-checking。For example, the University of Copenhagen has a lot of related work. The general approach is that for certain statements that need to be verified, we use artificial intelligence systems to search for relevant facts and predict a truthfulness score between 0 and 1 based on the facts.

Ms. A:What challenges does this approach present?

Serge：There is nothing wrong with this method itself; the challenge comes from the problem itself.Not all claims are worthy of fact-checking, and not all fact-checks result in true or false results.For example, "Pandas are China's national treasure" is a statement that can be verified through model training using a structured knowledge base and a large amount of data. However, "Moving to California" is not.

Ms. A:So you see research potential in statements like the latter?

Serge: These latter statements have not been studied much, but are equally important issues.These statements may not have a strict definition of true or false, but they will trigger a lot of discussion on social media. This challenge did not exist in the era when only news media were required to be fact-checked. However, in today's highly developed social media, a type of topic that triggers heated discussions, is difficult to scientifically characterize, or cannot be falsified has become very worthy of study.

Ms. A:Can you give an example of a case that had a real impact or even led to some intense conflict?

Serge：I'm glad you asked this question, let's take a fun example. When you wash your hands in a public restroom, you have two options for drying your hands. I'm not sure which is common in China, but in Europe, you can either pull out a paper towel or use a hot air dryer.

Ms. A:These two methods are also most common in China.

Serge：Dryer manufacturers and paper towel manufacturers can make a lot of money by signing contracts with any restaurant chain, and the entire market will probably generate billions of dollars in revenue. But many people in Europe now have very strong opinions about the difference between these two methods. Many people say that one method may spread diseases, while others say that using a lot of electricity or making paper will waste trees and cause environmental damage. Most people who hold these views are not public health or environmental experts.

Ms. A:Are these statements themselves true?

Serge: We don’t actually care about the authenticity of the speech, because many topics on social media cannot be strictly proved or falsified.But this topic was brought up because a small group of people wanted to convince the public that one method was better than the other. They probably created hundreds of thousands of bots to generate relevant content. If you search for discussions about paper towels and air dryers on social networks today, you will find millions of comments. Our research does not focus on rigorously comparing the pros and cons of the two methods and providing fact-checking results.We are more concerned with detecting these types of engineered issues.

3. COCO dataset in social media

Avoid making hasty decisions.

Ms. A: Your research has opened up another dimension. Traditional fact-checking focuses on the authenticity of semantics, while your research focuses on the authenticity of statements or claims.MingPragmatics——The goal of prediction is not limited to whether it is true or not, but extends to topic discussions on social networks created by some users or a large number of robots in order to achieve specific purposes.Do you know what this research means?

Serge：Yes, we are creating something completely new. Most of the research we know about is only focused on fact checking. But we are trying to use natural language technology related to topics to group and cluster discussions on social media to help individuals, businesses, and strategists understand what is happening on social media.We do not make value judgments on these topics and discussion contents, but only objectively show the form in which each issue is raised.

Ms. A:To achieve this goal, we first need a dataset. Building this dataset should be a big challenge. When you pioneered computer vision research, you expanded from a small bird dataset to COCO. How did you get started this time?

Serge：This type of social network behavior often has a characteristic: a topic may have a million tweets related to it, and it may appear that there are tens of thousands of accounts participating in the discussion.But through analysis, we might find that 100,000 of those tweets actually say exactly the same thing, with very similar or even identical narratives.But remember, this doesn't mean these discussions are right or wrong. We let users see clusters and groups of different opinions, making it easier for fact-checkers and social network analysts to process and understand large amounts of content without having to deal with a sudden influx of millions of tweets.

Ms. A:Can this system handle various controversial topics on social media in real time?

Serge：I think it can, and I hope so.Let's say two ships from Russia and the United States meet in the Mediterranean. Social networks start discussing and a set of narratives is born. Every few hours, new information appears, perhaps a statement from one captain or a cellphone recording from another. In this case, some narratives and issues gain attention, while others may become irrelevant.

Miss A: You hopereal timeCapture topics？

Serge：and other information.To help professional diplomats, we would like to create adash board（Dashboard）, providing comprehensive and relevant information, and also putting these events into the context of the world. Such a system canKeep people from making rash decisionsI want to emphasize that the system itself does not decide which side is right, but rather organizes information comprehensively.

Ms. A:What challenges need to be addressed in order to achieve reliable functionality?

Serge: There are both traditional and new challenges. Traditional challenges include the influence of language culture and emotional bias.

for example,Both The Little Mermaid and The Ugly Duckling are Danish works, but the Disney versions of the stories have been adapted to American culture. Due to the dominance of the American narrative over the Danish narrative on social networks, the original Andersen stories are almost unknown among Internet users in many other countries.

In the data annotation process, especially when annotating social network data, it is more affected by language and culture. For example, sentiment analysis is already an important part of fact-checking, and the sentiment prediction model itself may have a lot of bias and stereotypes in training.AI model training is a garbage in, garbage out process, and it is difficult to solve the problems caused by training data itself, so we must understand what training data the model uses.We can say that fact-checking without human involvement does not exist (is unreliable).

Ms. A:So what are the new challenges?

Serge: Fake content generated by language models is a new challenge we face.Previously, fake social media accounts often followed very simple patterns. But with GPT and image generation models, fake account creators can generate more complex and natural fake profiles, and then forge social media accounts that look real. These accounts are not easily found by traditional fake account identification models. These generative AI models also bring corresponding challenges to traditional fact-checking tasks. Therefore,Generative AI creating false information and identifying false information will be the cat-and-mouse game of this era.

4. AI Future

They (OpenAI) may not know our plans yet.

Ms. A:These challenges do not seem to be solvable simply through models, and may rise to the dimension of collaboration between AI and humans.youseemAlways discover new problems in new dimensions, and then solve them from a simple perspective.

Serge：Yes.Our new idea can be compared to Wikipedia.People used to think that Wikipedia only needed pages in different languages ​​with the same semantics for the same node. But the reality is that it's not just the language that's different.

The language, culture, values, traditions, all factors are mixed together on different pages of the same entry.For example, atomic energy and fossil fuels, it's covered very differently in different languages ​​and in different parts of the world. So this reminds us that the AI ​​system we're trying to build is not purely automated, nor is it a stand-alone model. It's a human-involved system, which means you need many different human communities around the world to annotate and organize the data and think about all the different parts.This is a big, deep problem because prejudice will always exist.

Ms. A:So, like MSCOCO, organizing data that is as comprehensive and fair as possible is the ambition of this research itself.

Serge：This is the process of organizing all the different types of communities. People of different ages and different parts of the world study literature, history, science, and so on, and each field has its own story. In order for the research I have described to be successful,We need a lot of annotations on various topicsBy。They don’t have to be experts, but they need to have some knowledge of the content they need to annotate, such as nuclear energy, entrepreneurship, or cryptocurrency, in order to know the similarities between narratives and topics. Therefore, the biggest challenge is the organization of the community, not the AI-based computing and storage facilities.

Ms. A:Did Sam Altman or Yann Lecun comment on your ideas?

Serge: They may not know our plan yet.

Ms. A:I seemed to be witnessing the first stages of a revolution:Discover problems at a higher dimension and find the most direct entry point.

Serge: If we develop this infrastructure around issue detection, like a lot of technology, it can be used for good or bad.Therefore, unlike many commercial AI,We seek to develop systems that are open, transparent and auditableTherefore, we will have a completely transparent knowledge base where users can see the editing history of the data, including when the data was collected and which annotators annotated it.

Ms. A:How to ensure the accuracy and objectivity of data?

Serge：The simple answer is,We cannot guarantee。

But the best we can do is to create a system that attracts tens of thousands of people interested in different discussion areas to annotate. As many annotators as possible can help us bring statistical objectivity. Wikipedia also has some mechanisms designed for transparency and accountability, and we will do the same.

Ms. A:How might this research impact policymakers, educators, and technologists?

Serge: You can think of what we are doing as a supplement to logical or factual reasoning.

Let's say a company wants to improve its diversity, equity, and inclusion. So their board of directors holds a meeting to discuss hiring more women or minorities. These kinds of discussions are common in many companies, for example, at a university, there may not be many women studying electrical engineering, and the department wants to take steps to change this. In these meetings, there may be a lot of discussions that are not supported by knowledge or information.

Some people will express biased opinions like women are not good at math. This is where a system is needed to help the department head, CEO, or teacher who needs to lead these discussions. They can extract a set of narratives from the system to structure the discussion. In addition, once the system starts working, it indexes and parses the statements into pre-existing narratives. In this way, the CEO, teacher, or meeting moderator can avoid low-quality or chaotic conversations and have an effective structure and classification system to guide the discussion and prevent redundant conversations.

Ms. A:What potential research directions do you think the technology will develop in future social media narratives and issue analysis?

Serge：Different fields have their own unique challenges. Some of them are classic problems, such as dealing with large amounts of data and how to label them, how to mitigate bias, etc. But in visualization, we also face big challenges.

We just mentioned the differences faced by different languages ​​and cultures. Each specific topic has many different perspectives, and different annotators will provide different annotations due to their own biases. From an information theory perspective, trying to compress these diverse accounts may result in loss or corruption of information. This type of problem will run through the entire project, and we will encounter these problems frequently.

Miss A: In your opinion, which recent advances in visual technology will have a profound impact on the future?

Serge：Now more and more researchers are beginning to pay attention to multimodal data, processing multiple data types such as images, text and audio in one model at the same time. This approach usually uses model architectures such as Transformer to solve complex practical problems. I believe this trend will continue, and newcomers in the field of artificial intelligence in the future will find that it is more natural to master multiple professional skills at the same time than to go deep into one field alone, such as natural language processing or computer vision.

Personally, I think that while some people claim that AI will completely replace doctors, this is overstated, I am sure that in areas such as radiology, dermatology, and histopathology, AI-assisted systems will become common and benefit everyone.

As for self-driving cars, despite past predictions that advances in vision technology and artificial intelligence would make self-driving cars commonplace, I think this is unlikely to happen. Unless the government takes steps to restrict conventional cars to certain lanes or ban them entirely, it is highly unlikely that self-driving cars will become the norm in the United States.

Miss A: I like your paper. I have similar ideas.Technological development is simultaneously unlocking new cognitive dimensions, and the most valuable methodology is precisely the one that has a minimal entry point but can radiate globally.

Serge：What methodology are you most interested in?

Ms. A:Let me give you a small example.Following the path of falsifiability, science will embark on an iterative path of negation of negation... returning to the structure of scientific revolution that we are so familiar with.

* Zhou Hang also contributed to this article.

Since this article involves academic discussion, here is a brief introduction of the author:

• Zhang Yiji, founder of Jiazi Light Year, graduated from the School of Mathematical Sciences of Peking University in 2013 with a double degree in economics from the National School of Development. He won the gold medal in the China Mathematical Olympiad and was selected for the national training team. His research areas are financial mathematics and game theory, and he also serves as a director of the School of Mathematical Sciences of Peking University.

• Zhou Hang, head of Jiazi Brain, graduated from the School of Mathematical Sciences of Peking University in 2019; his research interests are sparse optimization and non-convex optimization.

*References

• Searching for Structure in Unfalsifiable Claims.pdf

• 978-3-319-10602-1_48.pdf “Microsoft COCO: Common Objects in Context” MSCOCO dataset: Serge’s most cited paper.

• Thomas Samuel Kuhn, The Structure of Scientific Revolutions

Follow the public account "Jia Zi Guang Nian" and reply "falsifiable" to obtain the information mentioned in the article.Searching for structure in unfalsifiable claimsandMSCOCO datasetTwo papers.

|Miss ADialogue Series Review|