Showing posts with label algos. Show all posts
Showing posts with label algos. Show all posts

Friday, October 4, 2024

Robots Using Robots

Sometimes you have to give it to these scientists, the stuff they come up with is pretty smart. 

Who wrote this? Engineers discover novel method to identify AI-generated text
Mar 2024, phys.org

First, an interesting note:
"Stubbornness" is when LLMs show a tendency to alter human-written text more readily than AI-generated text, and it happens because LLMs often regard AI-generated text as already optimal and thus make minimal changes.

Next, the purpose:
Raidar (geneRative AI Detection viA Rewriting) - identifies whether text has been written by a human or generated by AI or LLMs, without needing access to a model's internal workings. 

Finally, the clever part:
It uses a language model to rephrase a given text and then measures how many edits the system makes to the given text. Many edits mean the text is likely written by humans, while fewer modifications mean the text is likely machine-generated.

via Columbia University School of Engineering and Applied Science: Chengzhi Mao et al, Raidar: geneRative AI Detection viA Rewriting, arXiv (2024). DOI: 10.48550/arxiv.2401.12970



Random robots are more reliable: New AI algorithm for robots consistently outperforms state-of-the-art systems
May 2024, phys.org

Maximum Diffusion Reinforcement Learning (MaxDiff RL) - an algorithm that encourages robots to explore their environments as randomly as possible in order to gain a diverse set of experiences; "designed randomness"; improves the quality of the data collected

If the robots move randomly, instead of some highly calculated, optimized trajectories, somehow the resulting data they collect on the world around them is better. Like when randomness is the base, it makes way better structures. I'm immediately thinking of watching a baby learn to move their body parts, or their vocal chords; underneath those first recognizable attempts is an endless iteration of random movements that are sometimes just now starting to get it right. 

via Northwestern McCormick School of Engineering: Maximum diffusion reinforcement learning, Nature Machine Intelligence (2024). DOI: 10.1038/s42256-024-00829-3

Post Script: It's funny to think of this, the MaxDiff RL, as an "algorithm", since it's kind of getting rid of any algorithms, that's the point here. When the algorithm is random, it's not an algorithm anymore; randomness is the anti-algorithm?


Researchers test AI systems' ability to solve the New York Times' connections puzzle
May 2024, phys.org

Chain of thought prompting:

The researchers found that explicitly prompting GPT-4 to reason through the puzzles step-by-step significantly boosted its performance to just over 39% of puzzles solved.

"Our research confirms prior work showing this sort of 'chain-of-thought' prompting can make language models think in more structured ways. Asking the language models to reason about the tasks that they're accomplishing helps them perform better."

via NYU Tandon School of Engineering: Graham Todd et al, Missed Connections: Lateral Thinking Puzzles for Large Language Models, arXiv (2024). DOI: 10.48550/arxiv.2404.11730


New ransomware attack based on an evolutional generative adversarial network can evade security measures
Jun 2024, phys.org

GAN-based architectures consist of two artificial neural networks that compete against each other to generate increasingly "better" results on a specific task. 

You already know it as the way we get hyperrealistic image generation or convincing conversation from a robot, and it's now being used to make malware attacks more effective. 

These scientists tested a version of this attack-enhancing approach, and found their framework capable of bypassing the majority of available anti-virus systems.

via Texas A&M University and Ho Technical University: Daniel Commey et al, EGAN: Evolutional GAN for Ransomware Evasion, 2023 IEEE 48th Conference on Local Computer Networks (LCN) (2023). DOI: 10.1109/LCN58197.2023.10223320


New technique improves the reasoning capabilities of large language models
Jun 2024, phys.org

Their approach, called natural language embedded programs (NLEPs), involves prompting a language model to create and execute a Python program to solve a user's query, and then output the solution as natural language.

NLEPs also improve transparency, since a user could check the program to see exactly how the model reasoned about the query and fix the program if the model gave a wrong answer.

via MIT: Tianhua Zhang et al, Natural Language Embedded Programs for Hybrid Language Symbolic Reasoning, arXiv (2023). DOI: 10.48550/arxiv.2309.10814

Posted by at No comments:
Labels: , , , , ,

Wednesday, September 4, 2024

Quantum Surprise


You can't escape quantum this and quantum that while perusing science headlines, but these articles in particular are examples of moments when quantum experiments produced surprising results, or even better, results that look really cool but we don't even know what to do with them yet. That's the best kind of surprise. 

Promising quantum state found during error correction research
Sep 2023, phys.org

A team of Cornell researchers unexpectedly discovered the presence of "spin-glass" quantum state while conducting a research project designed to learn more about quantum algorithms and, relatedly, new strategies for error correction in quantum computing.

The researchers emphasized that they weren't simply trying to generate a better error protection scheme when they began this research. Rather, they were studying random algorithms to learn general properties of all such algorithms.

"Interestingly, we found nontrivial structure," Mueller said. "The most dramatic was the existence of this spin-glass order, which points toward there being some extra hidden information floating around, which should be useable in some way for computing, though we don't know how yet."

via Cornell's Laboratory of Atomic and Solid State Physics: Vaibhav Sharma et al, Subsystem symmetry, spin-glass order, and criticality from random measurements in a two-dimensional Bacon-Shor circuit, Physical Review B (2023). DOI: 10.1103/PhysRevB.108.024205



Redefining quantum machine learning
Mar 2024, phys.org

The team has discovered that neuronal quantum networks can not only learn but also memorize seemingly random data. 

"Our experiments show that these quantum neural networks are incredibly adept at fitting random data and labels, challenging the very foundations of how we understand learning and generalization."

via Free University of Berlin: Elies Gil-Fuster et al, Understanding quantum machine learning also requires rethinking generalization, Nature Communications (2024). DOI: 10.1038/s41467-024-45882-z


Research demonstrates a new mechanism of order formation in quantum systems
Apr 2024, phys.org

This means something for materials science which is already leaping past us. Also RIKEN:

Active matter agents change from a disordered to an ordered state in what is called a "phase transition." As a result, they move together in an organized fashion without an external controller.

They created a theoretical model in which spins of subatomic particles align in one direction just like how flocking birds face the same direction while flying. They found that the ordering can appear without elaborate interactions between the agents in the quantum model.

"It was different from what was expected based on biophysical models."

via University of Tokyo and RIKEN: Activity-induced ferromagnetism in one-dimensional quantum many-body systems, Physical Review Research (2024). dx.doi.org/10.1103/PhysRevResearch.6.023096

Posted by at No comments:
Labels: , , , , ,

Friday, August 2, 2024

Signs You May Have a Problem

Wait a second - are we really coming up with substance abuse treatments for cell phone usage? Like as if it was nicotine or heroin? If the healthier way to "use" a phone involves an override program that literally makes it more difficult to use, then I think we have a problem here. 

Managing screen time by making phones slightly more annoying to use
May 2024, phys.org

Once the user's designated screen limit has been reached, InteractOut can delay the phone's response to a user's gesture, shift where tapping motions are registered or slow the screen scrolling speed.

The strength of the delays and shifts continues to increase each time the user touches the phone, up to a pre-set maximum, and the user can decide how the app interferes with their phone use. The app's gradual interference allows users to continue using their phone, but with a little extra difficulty.

"If we just continuously add a little bit of friction to the interaction with the phone, eventually the user becomes more aware of what they are doing because there's a mismatch between what they expect to happen and what actually happens. That makes using smartphones a more deliberate process." 
via University of Michigan: Tao Lu et al, InteractOut: Leveraging Interaction Proxies as Input Manipulation Strategies for Reducing Smartphone Overuse, Proceedings of the CHI Conference on Human Factors in Computing Systems (2024). DOI: 10.1145/3613904.3642317

Image credit: I think this is an early generation of AI image generation, pre 2022 pre Stable Diffusion-era?


Companies may still buy consumer genetic information despite its modest predictive power
May 2024, phys.org

At-home DNA tests have provided millions with insights into their family history and health risk by simply spitting in a tube. Genetic risk screening of embryos at in vitro fertilization clinics is now available. These tests rely on polygenic scores—a tally of variations in human genes that influence a certain trait. However, while powerful at predicting traits in large populations, these scores are rather weak at an individual level.

"Firms operate under a lot of uncertainty and any little bit of information that they have about you is worthwhile."

Using an economic model, the researchers found that companies might be willing to pay because the information may raise profit and is relatively cheap.

The researchers argue that current laws and policies are inadequate to address the ethical, privacy, and legal concerns surrounding the potential corporate use of polygenic scores. While the US Genetic Information Nondiscrimination Act prohibits discrimination in health coverage and employment based on genetic information, there are loopholes. The act only covers health insurance, excluding long-term, disability, life, and other insurances. It also doesn't apply to employers with fewer than 15 people, which accounts for 85% of US companies.

"I don't think people realize that when they give up their genetic information today, they're not giving it up just for today; they're giving it up forever"

via Johns Hopkins University Geisinger College of Health Sciences: Potential Corporate Uses of Polygenic Indexes: Starting a Conversation about the Associated Ethics and Policy Issues, The American Journal of Human Genetics (2024). DOI: 10.1016/j.ajhg.2024.03.010. 

Posted by at No comments:
Labels: , , , , , , ,

Wednesday, July 31, 2024

Social Math and the Particle Physics of People


As a teenager I spent my lunch hour staring at strangers in the mall, trying to figure out their relationship to each other based on no criteria whatsoever, mostly because I was reading Deepak Chopra, the Celestine Prophecy, and Carlos Castaneda, and I thought maybe you really could develop superpowers and see invisible energy fields connecting people together. I did not develop Kirlian vision, but I did learn how to read mostly invisible body language and subtle interpersonal dynamics. Maybe there was some rudimentary social network mapping going on as well:


Math can help people identify the bonds of friendship
Apr 2023, phys.org

People use statistical information to determine bonds between people. It found that children as young as five can enter a room and use statistics on social cues to determine whether two people are friends.

Researchers presented participants with diagrams of social networks that showed lines drawn between two main characters and other people in the group. Participants were told that these lines indicated friendship, and importantly, researchers did not show a line connecting the two main characters. Researchers then asked participants how likely it was that the two main characters were friends.

Both children and adults thought the two main characters were friends when they had a lot of overlap in their social connections in common. This pattern was found in children as young as five years old.

Participants were also able to conclude the strength of social connections from the size of the network. For example, when a character had many mutual connections with someone with a smaller network, participants saw this as more meaningful than having many mutual connections with someone with a larger social network.

via University of Waterloo: Claudia G. Sehl et al, The social network: How people infer relationships from mutual connections., Journal of Experimental Psychology: General (2022). DOI: 10.1037/xge0001330

Totally unrelated image credit: AI Art - Artificial Meat_2 - 2024


Spike in major league home runs tied to climate change
Apr 2023, phys.org

With enough data, you can see all kinds of things

While the researchers attribute only 1% of recent home runs to climate change, they found that rising temperatures could account for 10% or more of home runs by 2100 if greenhouse gas emissions and climate change continue unabated.

"There's a very clear physical mechanism at play in which warmer temperatures reduce the density of air. Baseball is a game of ballistics, and a batted ball is going to fly farther on a warm day," said senior author Justin Mankin, an assistant professor of geography.

The researchers analyzed more than 100,000 major league games and 220,000 individual hits to correlate the number of home runs with the occurrence of unseasonably warm temperatures. 

via Dartmouth: Christopher W. Callahan et al, Global warming, home runs, and the future of America's pastime, Bulletin of the American Meteorological Society (2023). DOI: 10.1175/BAMS-D-22-0235.1


The more stakeholders are included in policy planning, the better their policies protect them
Jun 2023, phys.org

This relates to network science in that it relies on a big complex system (of people and decision making) for maximum effect

"It's a big deal that we found empirical evidence that stakeholder integration leads to better protection. There are very few published papers that show this connection empirically."

"I was pretty stunned. I thought this was going to be a different paper." (Surprise!)

via University of California Santa Barbara: Debra Perrone et al, Stakeholder integration predicts better outcomes from groundwater sustainability policy, Nature Communications (2023). DOI: 10.1038/s41467-023-39363-y

AI Art - Artificial Meat_3 - 2024

Ridesourcing platforms thrive on socio-economic inequality, say researchers
Apr 2024, phys.org

Platforms that offer rides to passengers, such as Uber and DiDi, thrive on socio-economic inequality. It explains why in some cities ridesourcing services can be big players in the mobility system, while in other cities they don't get off the ground. 

"In cities like Amsterdam, with relatively low inequality, short travel distances and well-established bicycle and public transport networks, Uber is unlikely to flourish," researcher Arjan de Ruijter explains. "Therefore, transport authorities in such cities should rather focus on providing shared bikes and scooters to improve station access."

Conversely, in cities marked by significant inequality, like Johannesburg or Rio de Janeiro, Uber-like ridesourcing platforms thrive. Various explanations, taking into account driver's and passenger's behavior, emerge in the study. The platform capitalizes on a workforce willing to accept lower wages, leading to a service with limited waiting times for passengers. Moreover, it acknowledges the demand for mobility on demand among the affluent segments of unequal societies, willing to pay for a premium-like service.

"In a society with high inequality, companies can charge higher commissions to drivers, as drivers have limited alternative labor opportunities."

"On the other hand, in societies with low inequality, all else being equal, pricing strategies must attract more selective job seekers, resulting in lower commission rates."

via Delft University of Technology: Arjan de Ruijter et al, Ridesourcing platforms thrive on socio-economic inequality, Scientific Reports (2024). DOI: 10.1038/s41598-024-57540-x


Families will change dramatically over the years to come, says study
Jan 2024, phys.org

Gentle reminder 

  • The number of relatives that an individual has is expected to decrease by more than 35% in the near future. 
  • In 1950, a 65-year-old woman had an average of 41 living relatives. By 2095, a woman of the same age will have an average of only 25 living relatives.
  • The number of cousins, nieces, nephews and grandchildren will decline sharply, while the number of great-grandparents and grandparents will increase significantly.
  • The effects will be felt less in North America and Europe and more in South America and the Caribbean. 

via Max Planck Institute for Demographic Research in Rostock, University of Buenos Aires, and University of Amsterdam: Diego Alburez-Gutierrez et al, Projections of human kinship for all countries, Proceedings of the National Academy of Sciences (2023). DOI: 10.1073/pnas.2315722120

Posted by at No comments:
Labels: , , , ,

Wednesday, June 26, 2024

For Food's Sake


In a perfect world, this would destroy the artificial food industry

Want to know how processed your food is? There's an algorithm for that
Jun 2023, phys.org

The fingerprint of food processing:

"In the paper what we do is really say that we believe that nutritional information, so the chemicals that are measured as nutrients in the nutritional facts, somehow encode the fingerprint of food processing," says Giulia Menichetti, senior research scientist at Northeastern's Network Science Institute and lead author of the research. "Because when we process a food, when we modify some staple ingredients, we change its chemistry in many different ways."

That "fingerprinting" is the way that researchers can glean insight into just how many chemical alterations have been made to a given food using the U.S. Department of Agriculture's Food and Nutrient Database for Dietary Studies.

The machine learning classifier called FoodProX is from the Foodome project. 

The researchers note how the NOVA system, which splits foods into four classifications, from "unprocessed or minimally processed" to ultra-processed, is fundamentally limiting because it doesn't account for the different gradations of processing within each separate category.
More than 73% of the U.S. food system is ultra-processed. And tobacco companies own the food companies, in case you didn't know, like I didn't know. Companies never die, they only change their names, and the people who run them. Companies are transcendental intelligent entities, a sentient superorganism. In this case, it is a superorganism that is literally devouring the human population, eating the years off of our collective life expectancies in order to survive. Or are we all just cells that must be sloughed, part of a normal-functioning healthy system?

via Center for Complex Network Research at Northeastern University: Giulia Menichetti et al, Machine learning prediction of the degree of food processing, Nature Communications (2023). DOI: 10.1038/s41467-023-37457-1

(I believe the Center for Complex Network Research at Northeastern University was the original home of the Barabasi Lab [link])


To Repeat: The food industry is really the tobacco industry in disguise, and that most Americans eat one extra meal a day in junk food.
--US tobacco companies selectively disseminated hyper-palatable foods into the US food system: Empirical evidence and current implications. TL Fazzino, D Jun, L Chollet-Hinton, K Bjorlie. Addiction v119i1pp62-71 Jan 2024, 08 September 2023 https://doi.org/10.1111/add.16332
--Snacks contribute considerably to total dietary intakes among adults stratified by glycemia in the United States. K Heitman, CA Taylor, et al. PLoS Global Public Health. October 26, 2023. https://doi.org/10.1371/journal.pgph.0000802

Bonus:
AI Art - Cigarette Food Robot Confusion - Three Ladies Smoking in the Kitchen - 2023

Image credit: Note the thumbnail image above, and how AI Art doesn't understand how we use our mouths, and anytime you want a picture of someone eating, or smoking a cigarette for that matter, it looks like this. Meanwhile, consider the these image generator models are trained on The Internet, which we all know is 99% porn. Now remove the porn images from the trained model, i.e., "unlearn" the porn, and this is what you get - a robot that doesn't understand what a mouth is. 


Posted by at No comments:
Labels: , , , , ,

Wednesday, June 19, 2024

Algorithmic Crack


'Temu is as addictive as sugar': How the ecommerce retailer drives a shopping frenzy
Apr 2024, BBC News

China was the first (and to date the only) to use the term "electronic drugs"
--Shares slide after China brands online games 'electronic drugs'. Aug 2021, BBC News https://www.bbc.com/news/business-58066659 

This article came at a significant moment, one where I notice an inordinate amount of friends self-diagnosing as afflicted with ADHD (formerly known as ADD, after we realized that half the problem was the diet, but before we were allowed to call out the food industry for it, hence we still refuse to accept the affliction for what it is). 

We eat an entire extra meal's worth of calories in snacks every day. Americans that is, and half of us. We are, again likely the majority, all suffering from sleep deprivation due mostly to social jetlag and a bit of just-in-time algorithmic gig-worker flex-scheduling. 

But now, let's add this to the onslaught - we're all addicted to algorithms, groveling at the feet of the eternal, anti-dimensional Al-Khwarizmi. Pick your program - on a scale from simple social media to shopping to straight-gambling. Dark patterns, endless scrolls, rabbit holes, notifications specifically designed to bypass your fatigue thresholds, your fragile cerebral manifold. 

These programs, software, applications - basically anything you click or touch - they are engineered to game your brain. We learned the 4 rules of game theory a very long time ago, and have been plugging them into every user interface possible to extract every grain of attention from your fully-assembled, indivisible self that we possibly can. 

And then we come to the conclusion that we're the problem, because we have ADHD. The solution? More fucking drugs. If you don't think drug companies pay social media influencers to convince you that you have a mental disease that requires medication ,you're really not paying attention. (Granted, the influencers don't know what they're doing, or who they're really doing it for.) 

I can imagine it being hard for people to believe - I'm not injecting this into my veins. It's just a phone. 

No, it's not just a phone. It's neuromimetic virus that's been engineered to override your endocrine system. It hijacks your dopamine system to force you into paying attention, very close attention, to whatever it wants. 

You might have attention-deficit disorder all right. But after you're done watching that superhot 24 year old explain the epigenetic precursors for prefrontal cortex regulation dysfunction, or whatever she's saying that sounds really scientific and well-researched and must totally be highly credible information, stop and think about how endless scroll works. Or how sex sells. Or how humans can't not turn their head when they smell something rotten. We have reflexes, they're hardwired into us. We have biological limits, and we understand those limits really really well, and we know how to overcome and manipulate them. Do you know what an abusive relationship looks like? This is what it looks like. Maybe it will be easier to see it this way once the algorithms become people. We're almost there:


Post Script, on features of post-consumerism:
"You're either going to get something really good and an absolute bargain. Or you get something that's a bit naff, but won't actually return it, because it's less than £10 so it's not worth your time". [link]


And finally, just for good measure:
"Crack cocaine of algorithms"
--Imran Ahmed, the Center for Countering Digital Hate, TikTok sets 60-minute daily screen time limit for under-18s, Feb 2023, BBC News https://www.bbc.com/news/technology-64813981

AI Art - Scruffy British Man Sitting on a Dirty Couch - 2024

See further: 
On the Imagery of Addictions as Seen through the Eyes of the Internet, 2022

Electronic Drugs and Addiction by Design, 2022


Posted by at No comments:
Labels: , ,

Thursday, December 21, 2023

When Algorithms Attack - Thought Hygiene and Social Behavior in the Digital Age


Here's some things about how algorithms and social behavior interact, starting with a good background on how we as individuals treat bad information:

No simple answer for why people believe in conspiracy theories
Jun 2023, phys.org
"Conspiracy theorists are not all likely to be simple-minded, mentally unwell folks—a portrait which is routinely painted in popular culture," said Bowes. "Instead, many turn to conspiracy theories to fulfill deprived motivational needs and make sense of distress and impairment."
  • 170 studies, 158,000 participants from the United States, UK and Poland, measuring motivations or personality traits associated with conspiratorial thinking.
  • overall, people were motivated to believe in conspiracy theories by a need to understand and feel safe in their environment and a need to feel like the community they identify with is superior to others
  • people were more likely to believe specific conspiracy theories when they were motivated by social relationships
  • participants who perceived social threats were more likely to believe in events-based conspiracy theories (Sept. 11), rather than an abstract theory in general (governments plan to harm their citizens to retain power)
  • people who are motivated by a desire to feel unique are more likely to believe in general conspiracy theories about how the world works
  • The Big Five personality traits (extraversion, agreeableness, openness, conscientiousness and neuroticism) had a much weaker relationship with conspiratorial thinking
  • may be related to recent theory about social identity motives and conspiracy theories 
via Emory University: The Conspiratorial Mind: A Meta-Analytic Review of Motivational and Personological Correlates, Psychological Bulletin (2023). DOI: 10.1037/bul0000392

Image credit: AI Art - Is Your Data Under the Weather - 2023


Perception of Russia-Ukraine conflict linked to endorsement of false news about adversary
Mar 2023, phys.org
Analysis of the survey responses showed that participants who perceived a higher level of conflict between Ukraine and Russia were less likely to believe in and want to share the false stories about the European Union, but were more likely to endorse the false stories about Russia. Stories about Tanzania (a neutral control) were least likely to be endorsed.

These findings suggest that people's tendency to endorse false news does not depend simply on their group identity; it also depends on perceptions of the level of conflict between their group and another group. This implies that conflict de-escalation could help prevent the spread of misinformation.
via Leiden University in the Netherlands and Public Library of Science: Information battleground: Conflict perceptions motivate the belief in and sharing of misinformation about the adversary, PLoS ONE (2023). DOI: 10.1371/journal.pone.0282308


Social media 'trust' or 'distrust' buttons could reduce spread of misinformation
Jun 2023, phys.org

(You mean like a "dislike" button?)
The addition of "trust" and "distrust" buttons on social media, alongside standard "like" buttons, could help to reduce the spread of misinformation, finds a new experimental study led by University College London (UCL) researchers.
via UUCL Psychology & Language Sciences, Max Planck UCL Center for Computational Psychiatry and Aging Research, and Massachusetts Institute of Technology: Laura K Globig et al, Changing the incentive structure of social media platforms to halt the spread of misinformation, eLife (2023). DOI: 10.7554/eLife.85767

Also: Valentina Vellani et al, The illusory truth effect leads to the spread of misinformation, Cognition (2023). DOI: 10.1016/j.cognition.2023.105421


Controversy in Facebook posts linked to speed of spread among users
Jun 2023, phys.org

  • 57 million posts published across about 2 million Facebook pages and groups from 2018 to 2022. 
  • Posts that went viral were more likely to be associated with negative or controversial reactions among users, regardless of topic.
  • Posts that did not go viral were associated with more positive reactions.
via Sapienza Università di Roma: Etta G, Sangiorgio E, Di Marco N, Avalle M, Scala A, Cinelli M, et al. Characterizing engagement dynamics across topics on Facebook, PLoS ONE (2023). DOI: 10.1371/journal.pone.0286150

Image credit: AI Art - Isometric Maze by Jesper Ejsing, Rhads, Makoto Shinkai Lois van Baarle, and Ilya Kuvshinov - 2022

Researchers build an AI system to identify social norm violations
Jul 2023, phys.org
Built the system using GPT-3, zero-shot text classification, and automatic rule discovery. The system used a binary of ten social emotions: competence, politeness, trust, discipline, caring, agreeableness, success, conformity, decency, and loyalty.

DARPA commissioned The Computational Cultural Understanding (CCU) program to create cross-cultural language understanding technologies to improve a Department of Defense operator's situational awareness and interactional effectiveness. Cross-cultural miscommunication not only derails negotiations, but also can be a contributing factor leading to war, according to DARPA's explanation of the rationale for the program.
via Department of Cognitive and Brain Sciences Ben-Gurion University of the Negev: Yair Neuman et al, AI for identifying social norm violation, Scientific Reports (2023). DOI: 10.1038/s41598-023-35350-x


Sharing on Facebook reveals two very different news environments
Aug 2023, phys.org

Low-Cred
Here it is: "There's a little more nuance to what it takes to be a well-informed person. You can't just be reading multiple news stories. You have to be reading sources that are well known to be credible."
(I think we're having a problem with the "well-known" part; well-known by who? It gets even harder  when more and more of our online infodiets come from robots.)
  • 2 million online news stories that were shared at least 100 times by Facebook users from February 2017 to April 2019
  • both high- and low-credibility publishers tended to put out bursts of coverage at the same time -- but they often were about different topics
  • low-credibility publishers more often converged on stories about politics and the government
  • high-credibility publishers' topics ranged more widely
via Ohio State: Ceren Budak et al, Bursts of contemporaneous publication among high- and low-credibility online information providers, New Media & Society (2023). DOI: 10.1177/14614448231183617


Social media algorithms exploit how humans learn from their peers
Aug 2023, phys.org

They can't change to create a sense of community because that's not their purpose and doesn't yield profit.
Humans are biased to learn from others in a way that typically promotes cooperation and collective problem-solving, which is why they tend to learn more from individuals they perceive as a part of their ingroup and those they perceive to be prestigious. In addition, when learning biases were first evolving, morally and emotionally charged information was important to prioritize, as this information would be more likely to be relevant to enforcing group norms and ensuring collective survival.

In contrast, algorithms are usually selecting information that boosts user engagement in order to increase advertising revenue. This means algorithms amplify the very information humans are biased to learn from, and they can oversaturate social media feeds with what the researchers call Prestigious, Ingroup, Moral, and Emotional (PRIME) information, regardless of the content's accuracy or representativeness of a group's opinions.
This last part is great part because it sounds like how you have to talk about an alien race taking over the planet or a new parasite to wipe out the population -- 
"It's not that the algorithm is designed to disrupt cooperation. It's just that its goals are different."
via Kellogg School of Management at Northwestern: Algorithm-mediated social learning in online social networks, Trends in Cognitive Sciences (2023). DOI: 10.1016/j.tics.2023.06.008


No evidence linking Facebook adoption and negative well-being
Aug 2023, phys.org

According to the article below this, it doesn't matter because it's already broken, but also because nobody is going to believe this.
"largest independent scientific study ever conducted" (nearly a million people across 72 countries over 2008 to 2019)
Reasons -- "Much of the past research into social media use and well-being has been hampered by an exclusive focus on well-being data in the Global North and a reliance on inaccurate self-reports of social media engagement."
There was another study that says the positive effects are mediated by affluence, so that the wealthier the country the more negative the effects and vice versa? I don't see it here, but I think it just came out in the last few months. Also, and this isn't so much a swipe at the authors, but the state of affairs in general, that their first reference, where they introduce their "2.94 billion monthly active users" number, is sourced from quarterly earnings data from the company itself. We know private companies blackbox lots of their data, and that we can't use with confidence the numbers they give us (the public, or university researchers), and yet one of the base numbers of their whole calculation comes from the company, lowering the confidence of all that follows. 

via University of Oxford's Internet Institute: Matti Vuorre et al, Estimating the association between Facebook adoption and well-being in 72 countries, Royal Society Open Science (2023). DOI: 10.1098/rsos.221451


Facebook's design makes it unable to control misinformation, research suggests
Sep 2023, phys.org
Facebook's efforts were undermined by the core design features of the platform itself. "Our results show that removing content or changing algorithms can be ineffective if it doesn't change what the platform is designed to do."
via George Washington and Johns Hopkins Universities: David Broniatowski, The Efficacy of Facebook's Vaccine Misinformation Policies and Architecture During The COVID-19 Pandemic, Science Advances (2023). DOI: 10.1126/sciadv.adh2132



And finally, feed this to your model-bot:

High rate of mental health problems and political extremism found in those who bought firearms during COVID pandemic
Sep 2023, phys.org

Not about the article itself, but this is the statement that defines you as a conspiracy theorist in scientific studies:

"The government, media, and financial worlds in the U.S. are controlled by a group of Satan-worshiping pedophiles who run a global child sex trafficking operation." (I agree, I do not agree)

via University of Michigan: Brian M. Hicks et al, Who bought a gun during the COVID-19 pandemic in the United States?: Associations with QAnon beliefs, right-wing political attitudes, intimate partner violence, antisocial behavior, suicidality, and mental health and substance use problems, PLOS ONE (2023). DOI: 10.1371/journal.pone.0290770

Posted by at No comments:
Labels: , , , , , ,

Friday, September 30, 2022

Dematerialization Incoming


New climate modeling predicts increasing occurrences of flash flooding across most of the US
May 2022, phys.org

It realizes the concept of 'Digital Twin in Earth System Science'

via University of Oklahoma, NOAA National Severe Storms Laboratory, National Center for Atmospheric Research: Zhi Li et al, The conterminous United States are projected to become more prone to flash floods in a high-end emissions scenario, Communications Earth & Environment (2022). DOI: 10.1038/s43247-022-00409-6

UN Climate Change Reports



Google Maps “immersive view” is the ultimate graphics mode for Google Maps
May 2022, Ars Technica

There are simulated cars that drive through the roads, and birds fly through the sky. Clouds pass overhead and cast shadows on the world. The weather is simulated, and water has realistic reflections that change with the camera. London even has an animated Ferris wheel that spins around. Google can't possibly be tracking things like the individual positions of birds (yet!), but a lot of this is real data. The cars represent the current traffic levels on a given street. The weather represents the actual weather, even for historical data. The sun moves in real time with the time of day.

Google Maps: Immersive View


NeuroMechFly: A digital twin of Drosophila
May 2022, phys.org

It's not an actual robot but a model:

EPFL scientists have developed a digital model of the fruit fly, Drosophila melanogaster, that realistically simulates the movements of the animal. The twin is a big step towards reverse engineering the neuromechanical control of animal behavior, and developing bioinspired robots.

Morphologically realistic biomechanical model and a kinematic database of real limb movements of the fly from deep-learning pose estimation motion-capture software

A model that integrates what we know about the fly's nervous system and biomechanics to test if it is enough to explain its behavior."

Drosophila is the most commonly used insect in the life sciences.

via EPFL Ecole Polytechnique Federale de Lausanne: NeuroMechFly, a neuromechanical model of adult Drosophila melanogaster, Nature Methods (2022). DOI: 10.1038/s41592-022-01466-7


Why you may have a thinking digital twin within a decade
Jun 2022, BBC News

Dr Steve Levin's daughter was born with congenital heart disease, and a few year's back, when she was in her late 20s and at high risk of heart failure, he decided to recreate her heart in virtual reality so it can be tested and analysed, allowing surgeons to play out a series of "what if" scenarios for the organ, using various procedures and medical devices.

US software firm Nvidia runs a platform called Omniverse to create digital twins. One of its most ambitious projects is to build a digital doppelganger of the Earth, called Earth-2.

In March this year, the European Commission, in conjunction with the European Space Agency among others, announced its own plans to make a digital twin of the planet, dubbed Destination Earth.


Algorithm predicts crime a week in advance, but reveals bias in police response
Jul 2022, phys.org

Just here for the paraself (shoutout to Ted Chiang)

"We created a digital twin of urban environments. If you feed it data from happened in the past, it will tell you what's going to happen in future. 

via University of Chicago: Ishanu Chattopadhyay, Event-level prediction of urban crime reveals a signature of enforcement bias in US cities, Nature Human Behaviour (2022). DOI: 10.1038/s41562-022-01372-0


Twin physically unclonable functions (PUFs) based on carbon nanotube arrays to enhance the security of communications
Jul 2022, phys.org

"To achieve a low-cost and hardware-based secure communication, we introduced a new technology, twin physically unclonable functions (PUFs)," Zhang said.

"The basic idea behind PUFs is to utilize random physical imperfections existing in a physical entity caused by the fabrication process, and these imperfections cannot be predicted or cloned, even by the original manufacturer."

Due to their unique design, PUF devices are unclonable and unpredictable. This makes them incredibly effective at generating safe secret keys for encryption.

via Peking University and Jihua Laboratory: Donglai Zhong et al, Twin physically unclonable functions based on aligned carbon nanotube arrays, Nature Electronics (2022). DOI: 10.1038/s41928-022-00787-x

Posted by at No comments:
Labels: , , , , ,

Monday, September 19, 2022

Human Meat Restaurant


Unpaid social media moderators perform labor worth $3.4 million a year on Reddit alone
Jun 2022, phys.org

"2.8% of Reddit's 2019 revenue"

"Putting a price tag on the labor that people—in this case, content moderators on Reddit—have subsidized is leverage those moderators could wield when asking platforms for better resources and tools to help them monitor more effectively," Li said.

People, Space and Algorithms (PSA) Research Group at Northwestern. The group's overall mission is to "identify and address societal problems that are created or exacerbated by advances in computer science."

Li said a key part of the PSA Group's work involves re-framing user contributions to sites like Google, Facebook, Twitter and Reddit as "work"—not passive participation in online space—because companies use data and time that users provide to generate profit: to train their algorithms, better target advertising, recruit new users and ultimately earn more revenue.

This reframing led them to coin the term "data labor subsidy" when placing a dollar value on the contributions of tech platform users.

Not only do the users offer their data for free, but they also do the work of maintaining the platform for free; clever business model.

via People, Space and Algorithms Research Group at Northwestern: Hanlin Li, Brent Hecht, Stevie Chancellor, Measuring the Monetary Value of Online Volunteer Work. arXiv:2205.14528v1 [cs.HC], arxiv.org/abs/2205.14528

Also: Hanlin Li, Brent Hecht, Stevie Chancellor, All That's Happening behind the Scenes: Putting the Spotlight on Volunteer Moderator Labor in Reddit. arXiv:2205.14529v1 [cs.HC], arxiv.org/abs/2205.14529

Image credit: AI Art - Human Meat Restaurant, 2022
Prompt: human meat restaurant, horror, nightmare, cook, food, cooking. https://lexica.art/prompt/0a5b4f3f-f3e5-41b5-8942-d374325546e3


Human-like features in robot behavior: Response time variability can be perceived as human-like
Jul 2022, phys.org

Features of human behavior, namely response timing, can be translated into the robot in a way that humans cannot distinguish whether they are interacting with a person or a machine.

The human brain has sensitivity to extremely subtle behavior which manifests humanness," says Agnieszka Wykowska. "In our non-verbal Turing test, human participants had to judge whether they were interacting with a machine or a person, by considering only the timing of button presses during a joint action task."

The results showed that people interacting with the robot were not able to tell whether the robot was human-controlled or pre-programmed in the condition when the robot was in fact pre-programmed. This suggests that the robot passed this version of the non-verbal Turing test in this specific task.

via Italian Institute of Technology: F. Ciardo et al, Human-like behavioral variability blurs the distinction between a human and a machine in a nonverbal Turing test, Science Robotics (2022). DOI: 10.1126/scirobotics.abo1241.


How Facebook clickbait draws users into engaging with posts
Jul 2022, phys.org
  • The team collected 4,000 posts from seven consecutive days in late 2017 from ten U.S. and U.K. news outlets' Facebook pages, including "reputable" and "tabloid" sources. 
  • User engagement as measured by shares, comments and reactions
Results:
  • Unusual punctuation in the headline got 2.5 times more engagement.
  • Unusual punctuation in the text, however, got a decrease in engagement.
  • Questions in either the headline or the text did not get increased engagement.
  • Longer words in headlines got reduced engagement. 
  • The opposite was true for text.
  • Doubling the number of headline words led to 23.7% fewer comments, but no difference in reactions or shares.
  • The opposite for text, where all engagement increased with a doubling of the word count.
  • Common clickbait phrases in headlines — like "this will blow your mind" — were associated with a loss of around a quarter of engagement in comparison to those without such phrases.
  • In the sentiment analysis, negative wording in posts can increase comments
  • But for headlines, positive tone increases comments.
In other words, make sure your headlines have unusual punctuation and positive wording sentiment, and your text has longer words, more words, and negative wording sentiment.

And make sure your headlines DO NOT have questions, longer words, more words, or phrases "this will blow your mind", and that your text DOES NOT have unusual punctuation or questions.

via University of Duisburg-Essen: Click me…! The influence of clickbait on user engagement in social media and the role of digital nudging, PLoS ONE (2022). DOI: 10.1371/journal.pone.0266743

Posted by at No comments:
Labels: , , , , , ,

Wednesday, September 7, 2022

Falling Down


AKA Data Miners Go on Strike

Amazon targets 10,000 fake review Facebook groups
Jul 2022, BBC News
 
Amazon is taking legal action against the administrators of more than 10,000 Facebook groups, the BBC has learned.

It says the groups are generating fake reviews on Amazon marketplaces in the US, UK, Germany, France, Italy, Spain, and Japan.

The tech sector has finally started to cannibalize itself. 


First ever sales fall for Facebook owner Meta
Jul 2022, BBC News
https://www.bbc.com/news/business-62322862

"When you look at the ad space right now...growth is deteriorating at a very fast pace, much faster than most people had anticipated," Mr Zino said.

Meta's reliance on small and medium-sized businesses - which are "spooked" about the economy - makes it particularly vulnerable to any market slowdown, said Nikhil Lai senior analyst for performance marketing at Forrester Research.

The company's fine-tuned model for targeting adverts was also disrupted last year when Apple overhauled its privacy settings.

Let's just accept it now, in a pretty unexpected turn of events, the Roe decision has pushed the American populace off the privacy cliff. 

Let's also recall the 2014 paper that followed the disease model on a social network and predicted Facebook would run out of users by 2018 (which then happened, although the balance was being replaced by bots at the same time so nobody cared to acknowledge it).
-Epidemiological modeling of online social network dynamics, John Cannarell, Joshua A. Spechler. Department of Mechanical and Aerospace Engineering, Princeton University. arxiv:1401.4208v1 [cs.SI] 17 Jan 2014. https://arxiv.org/pdf/1401.4208v1.pdf

Posted by at No comments:
Labels: , , , , , ,

It Knows


AKA The Great Recognizer

Neural network can read tree heights from satellite images
Apr 2022, phys.org

We need to be reminded of how powerful data can be when it's f**king massive. It doesn't even have to be related. Like for example I can tell which zip code you grew up in, even the street and maybe even the house, based on really really fine-grained data about your teethbrushing habits, delivered by your smart toothbrush of course.

It sounds crazy, but given enough data, nothing is crazy. 

We don't even have to know how to do it, just give a neural network enough data, and it will figure out the problem for you:

"Since we don't know which patterns the computer needs to look out for to estimate height, we let it learn the best image filters itself."

All you need are some training data, so in this case that means a bunch of trees for which we do know the height. Then we take the (otherwise flat) satellite data, mash it with the height data for the known trees to teach the network, and then unleash the network on the unknowns.  

And why do we want to know how tall trees are? "Because whenever we cut down trees, we release carbon into the atmosphere, and we don't know how much carbon we are releasing."

via ETH Zurich: Nico Lang, Walter Jetz, Konrad Schindler, Jan Dirk Wegner, A high-resolution canopy height model of the Earth. arXiv:2204.08322v1 [cs.CV], arxiv.org/abs/2204.08322

Image source: The 4 Trends That Prevail on the Gartner Hype Cycle for AI, 2021, Gartner, Sep 2021
via: Intelligent Sensing: Enabling the Next “Automation Age” by Marco Cassis of STMicroelectronics at International Solid-State Circuits Conference 2022


Anxious individuals identified by analyzing their walking gait
May 2022, phys.org

The best method for identifying anxious individuals was walking. The team successfully identified people who were anxious with 75% accuracy.

They had to complete a balance test and a two-minute walk while wearing sensors. Based on these data, the team determined the young people who report being anxious walk in a way that's very similar to older adults who are fearful of falling. They find that young, anxious adults are constantly scanning for threats from side to side while walking and have trouble turning. The researchers also reported that anxious people have worse balance than those who are anxious.

via Clarkson University: Maggie Stark et al, Identifying Individuals Who Currently Report Feelings of Anxiety Using Walking Gait and Quiet Balance: An Exploratory Study Using Machine Learning, Sensors (2022). DOI: 10.3390/s22093163


Using electric signals from human brains, new software can perform computerized image editing
Jun 2022, phys.org

"All the existing software has been previously trained with labeled input. So, if you want an app which can make people look older, you feed it thousands of portraits and tell the computer which ones are young, and which are old.

Here, the brain activity of the subjects was the only input.

This is an entirely new paradigm in artificial intelligence—using the human brain directly as the source of input."

"All the existing software has been previously trained with labeled input. So, if you want an app which can make people look older, you feed it thousands of portraits and tell the computer which ones are young, and which are old. Here, the brain activity of the subjects was the only input. This is an entirely new paradigm in artificial intelligence—using the human brain directly as the source of input."

But alas:
"Collecting individual brain signals does involve ethical issues..."

We are the data, and we have lots of it.

Be a real shame if someone were to save our brainwaves and then sell them on the data market, where they could in turn unintentionally feed back to us our own biases after having been amplified by some behaviorally-exploitative algorithm...

via University of Copenhagen and University of Helsinki: Brain-Supervised Image Editing. Keith M. Davis III et al. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun 2022. 


People with similar faces likely have similar DNA
Aug 2022, phys.org

They recruited human doubles from the photographic work of François Brunelle, a Canadian artist who has been obtaining worldwide pictures of look-alikes since 1999. They obtained headshot pictures of 32 look-alike couples. 

Physical traits such as weight and height, as well as behavioral traits such as smoking and education, were correlated in look-alike pairs. Taken together, the results suggest that shared genetic variation not only relates to similar physical appearance, but may also influence common habits and behavior.

Eventually the all-seeing omnibot will be able to create a complete living human population gene network based only on our faces (and it will predict our habits and behaviors too).

via Barcelona Supercomputing Center and Josep Carreras Leukaemia Research Institute Barcelona: Manel Esteller, Look-alike humans identified by facial recognition algorithms show genetic similarities, Cell Reports (2022). DOI: 10.1016/j.celrep.2022.111257


Algorithm predicts crime a week in advance, but reveals bias in police response
Jul 2022, phys.org

Algorithm that forecasts crime by learning patterns in time and geographic locations from public data on violent and property crimes. The model can predict future crimes one week in advance with about 90% accuracy.

Not like that!
In a separate model, the research team also studied the police response to crime by analyzing the number of arrests following incidents and comparing those rates among neighborhoods with different socioeconomic status. They saw that crime in wealthier areas resulted in more arrests, while arrests in disadvantaged neighborhoods dropped. Crime in poor neighborhoods didn't lead to more arrests, however, suggesting bias in police response and enforcement.

via University of Chicago: Ishanu Chattopadhyay, Event-level prediction of urban crime reveals a signature of enforcement bias in US cities, Nature Human Behaviour (2022). DOI: 10.1038/s41562-022-01372-0


AI reveals unsuspected math underlying search for exoplanets
May 2022, phys.org

I forget about the singularity sometimes, but it's happening right now; we're already in the singularity:
Artificial intelligence (AI) algorithms trained on real astronomical observations now outperform astronomers in sifting through massive amounts of data to find new exploding stars, identify new types of galaxies and detect the mergers of massive stars, accelerating the rate of new discovery in the world's oldest science.

But that's already been the case, now it appears the AI has discovered "unsuspected connections hidden in the complex mathematics arising from general relativity". 
The algorithm figured out new rules of gravitational microlensing:
"I argue that they constitute one of the first — if not the first — time that AI has been used to directly yield new theoretical insight in math and astronomy."-Joshua Bloom, UC Berkeley professor of astronomy

"Keming's machine learning algorithm uncovered this degeneracy that had been missed by experts in the field toiling with data for decades. This is suggestive of how research is going to go in the future when it is aided by machine learning, which is really exciting." -Scott Gaudi, professor of astronomy at Ohio State

via University of California - Berkeley: Keming Zhang et al, A ubiquitous unifying degeneracy in two-body microlensing systems, Nature Astronomy (2022). DOI: 10.1038/s41550-022-01671-6


Post Script:
"AI software has collaborated with mathematicians to successfully develop a theorem about the structure of knots, but the suggestions given by the code were so unintuitive that they were initially dismissed. Only later were they discovered to offer invaluable insight. The work suggests AI may reveal new areas of mathematics where large data sets make problems too complex to be comprehended by humans."
-DeepMind AI collaborates with humans on two mathematical breakthroughs, Matthew Sparkes, New Scientist, Dec 2021 [link]

They're talking about knots. Instead of feeding it facebook photos or product reviews, they gave it math problems...about knots. The "connection between algebraic and geometric invariants of knots".

And why knots? Because theories about knots can also be applied to quantum field theory. Don't ask me why, but they do, and it's called topology.

via Google DeepMind: Advancing mathematics by guiding human intuition with AI. Davies, A., Veličković, P., Buesing, L. et al. Nature 600, 70–74 (2021). DOI: 10.1038/s41586-021-04086-x


Post Post Script, On Synthetic Data:
'Fake' data helps robots learn the ropes faster
Jul 2022, phys.org

How to be a human:
For the rope-looping simulation and experiment, Mitrano and Berenson expanded the data set by extrapolating the position of the rope to other locations in a virtual version of a physical space—so long as the rope would behave the same way as it had in the initial instance. Using only the initial training data, the simulated robot hooked the rope around the engine block 48% of the time. After training on the augmented data set, the robot succeeded 70% of the time.
via  University of Michigan: Data Augmentation for Manipulation, arXiv:2205.02886v3 [cs.RO]  https://doi.org/10.48550/arXiv.2205.02886


Posted by at No comments:
Labels: , , , , , , , , , , , , ,
Subscribe to: Posts (Atom)