Computers Math

The era of artificial-intelligence chatbots that seem to understand and use language the way we humans do has begun. Under the hood, these chatbots use large language models, a particular kind of neural network. But a new study shows that large language models remain vulnerable to mistaking nonsense for natural language. To a team of researchers at Columbia University, it’s a flaw that might point toward ways to improve chatbot performance and help reveal how humans process language.
In a paper published online today in Nature Machine Intelligence, the scientists describe how they challenged nine different language models with hundreds of pairs of sentences. For each pair, people who participated in the study picked which of the two sentences they thought was more natural, meaning that it was more likely to be read or heard in everyday life. The researchers then tested the models to see if they would rate each sentence pair the same way the humans had.
In head-to-head tests, more sophisticated AIs based on what researchers refer to as transformer neural networks tended to perform better than simpler recurrent neural network models and statistical models that just tally the frequency of word pairs found on the internet or in online databases. But all the models made mistakes, sometimes choosing sentences that sound like nonsense to a human ear.
“That some of the large language models perform as well as they do suggests that they capture something important that the simpler models are missing,” said Dr. Nikolaus Kriegeskorte, PhD, a principal investigator at Columbia’s Zuckerman Institute and a coauthor on the paper. “That even the best models we studied still can be fooled by nonsense sentences shows that their computations are missing something about the way humans process language.”
Consider the following sentence pair that both human participants and the AI’s assessed in the study:
That is the narrative we have been sold.
This is the week you have been dying. More

Capturing blur-free images of fast movements like falling water droplets or molecular interactions requires expensive ultrafast cameras that acquire millions of images per second. In a new paper, researchers report a camera that could offer a much less expensive way to achieve ultrafast imaging for a wide range of applications such as real-time monitoring of drug delivery or high-speed lidar systems for autonomous driving.
“Our camera uses a completely new method to achieve high-speed imaging,” said Jinyang Liang from the Institut national de la recherche scientifique (INRS) in Canada. “It has an imaging speed and spatial resolution similar to commercial high-speed cameras but uses off-the-shelf components that would likely cost less than a tenth of today’s ultrafast cameras, which can start at close to $100,000.”
In Optica, Optica Publishing Group’s journal for high-impact research, Liang together with collaborators from Concordia University in Canada and Meta Platforms Inc. show that their new diffraction-gated real-time ultrahigh-speed mapping (DRUM) camera can capture a dynamic event in a single exposure at 4.8 million frames per second. They demonstrate this capability by imaging the fast dynamics of femtosecond laser pulses interacting with liquid and laser ablation in biological samples.
“In the long term, I believe that DRUM photography will contribute to advances in biomedicine and automation-enabling technologies such as lidar, where faster imaging would allow more accurate sensing of hazards,” said Liang. “However, the paradigm of DRUM photography is quite generic. In theory, it can be used with any CCD and CMOS cameras without degrading their other advantages such as high sensitivity.”
Creating a better ultrafast camera
Despite a great deal of progress in ultrafast imaging, today’s methods are still expensive and complex to implement. Their performance is also limited by trade-offs between the number of frames captured in each movie and light throughput or temporal resolution. To overcome these issues, the researchers developed a new time-gating method known as time-varying optical diffraction.
Cameras use gates to control when light hits the sensor. For example, the shutter in a traditional camera is a type of gate that opens and closes once. In time-gating, the gate is opened and closed in quick succession a certain number of times before the sensor reads out the image. This captures a short high-speed movie of a scene. More

Scientists have confirmed that human brains are naturally wired to perform advanced calculations, much like a high-powered computer, to make sense of the world through a process known as Bayesian inference.
In a study published in the journal Nature Communications, researchers from the University of Sydney, University of Queensland and University of Cambridge developed a specific mathematical model that closely matches how human brains work when it comes to reading vision. The model contained everything needed to carry out Bayesian inference.
Bayesian inference is a statistical method that combines prior knowledge with new evidence to make intelligent guesswork. For example, if you know what a dog looks like and you see a furry animal with four legs, you might use your prior knowledge to guess it’s a dog.
This inherent capability enables people to interpret the environment with extraordinary precision and speed, unlike machines that can be bested by simple CAPTCHA security measures when prompted to identify fire hydrants in a panel of images.
The study’s senior investigator Dr Reuben Rideaux, from the University of Sydney’s School of Psychology, said: “Despite the conceptual appeal and explanatory power of the Bayesian approach, how the brain calculates probabilities is largely mysterious.”
“Our new study sheds light on this mystery. We discovered that the basic structure and connections within our brain’s visual system are set up in a way that allows it to perform Bayesian inference on the sensory data it receives.
“What makes this finding significant is the confirmation that our brains have an inherent design that allows this advanced form of processing, enabling us to interpret our surroundings more effectively.”
The study’s findings not only confirm existing theories about the brain’s use of Bayesian-like inference but open doors to new research and innovation, where the brain’s natural ability for Bayesian inference can be harnessed for practical applications that benefit society. More

When people feel that their resources are scarce — that they don’t have enough money or time to meet their needs — they often make decisions that favor short-term gains over long-term benefits. Because of that, researchers have argued that scarcity pushes people to make myopic, impulsive decisions. But a study published by the American Psychological Association provides support for a different, less widely held view: People experiencing scarcity make reasonable decisions based on their circumstances, and only prioritize short-term benefits over long-term gains when scarcity threatens their more immediate needs.
“This research challenges the predominant view that when people feel poor or live in poverty, they become impatient and shortsighted and can’t or don’t think about the future,” said study co-author Eesha Sharma, Ph.D., of San Diego State University. “It provides a framework, instead, for understanding that when people are experiencing financial scarcity, they’re trying to make the best decision they can, given the circumstances they’re in.”
The research was published in the Journal of Personality and Social Psychology.
Sharma and co-authors Stephanie Tully, Ph.D., of the University of Southern California, and Xiang Wang, Ph.D., of Lingnan University in Hong Kong, wanted to distinguish between two competing ideas: That people’s preference for shorter-term gains reflects impatience and impulsivity, or that it reflects more intentional, deliberate decision-making. To do so, they examined how people’s decisions change depending on the timeline of the needs that they feel they don’t have enough money for.
“Needs exist across a broad time horizon,” said Tully. “We often think about immediate needs like food or shelter, but people can experience scarcity related to future needs, too, such as replacing a run-down car before it dies, buying a house or paying for college. Yet research on scarcity has focused almost exclusively on immediate needs.”
In the current study, the researchers conducted five experiments in which they measured or induced a sense of scarcity in participants, and examined how the choices people made changed depending on whether that scarcity was related to a shorter- or longer-term need.
Overall, they found that when people feel that they don’t have enough resources to meet an immediate need, such as food or shelter, they are more likely to make decisions that offer an immediate payout, even if it comes at the expense of receiving a larger payout later. But when scarcity threatens a longer-term need, such as replacing a run-down car, people experiencing scarcity are no less willing to wait for larger, later rewards — and in some cases are more willing to wait — compared with people not experiencing scarcity. More

Recent advances in artificial intelligence and deep learning have revolutionized many industries, and might soon help recreate your neighborhood as well. Given images of a landscape, the analysis of deep-learning models can help urban landscapers visualize plans for redevelopment, thereby improving scenery or preventing costly mistakes.
To accomplish this, however, models must be able to correctly identify and categorize each element in a given image. This step, called instance segmentation, remains challenging for machines owing to a lack of suitable training data. Although it is relatively easy to collect images of a city, generating the ‘ground truth’, that is, the labels that tell the model if its segmentation is correct, involves painstakingly segmenting each image, often by hand.
Now, to address this problem, researchers at Osaka University have developed a way to train these data-hungry models using computer simulation. First, a realistic 3D city model is used to generate the segmentation ground truth. Then, an image-to-image model generates photorealistic images from the ground truth images. The result is a dataset of realistic images similar to those of an actual city, complete with precisely generated ground-truth labels that do not require manual segmentation.
“Synthetic data have been used in deep learning before,” says lead author Takuya Kikuchi. “But most landscape systems rely on 3D models of existing cities, which remain hard to build. We also simulate the city structure, but we do it in a way that still generates effective training data for models in the real world.”
After the 3D model of a realistic city is generated procedurally, segmentation images of the city are created with a game engine. Finally, a generative adversarial network, which is a neural network that uses game theory to learn how to generate realistic-looking images, is trained to convert images of shapes into images with realistic city textures This image-to-image model creates the corresponding street-view images.
“This removes the need for datasets of real buildings, which are not publicly available. Moreover, several individual objects can be separated, even if they overlap in the image,” explains corresponding author Tomohiro Fukuda. “But most importantly, this approach saves human effort, and the costs associated with that, while still generating good training data.”
To prove this, a segmentation model called a ‘mask region-based convolutional neural network’ was trained on the simulated data and another was trained on real data. The models performed similarly on instances of large, distinct buildings, even though the time to produce the dataset was reduced by 98%.
The researchers plan to see if improvements to the image-to-image model increase performance under more conditions. For now, this approach generates large amounts of data with an impressively low amount of effort. The researchers’ achievement will address current and upcoming shortages of training data, reduce costs associated with dataset preparation and help to usher in a new era of deep learning-assisted urban landscaping. More

Researchers at the University of Washington have developed small robotic devices that can change how they move through the air by “snapping” into a folded position during their descent.
When these “microfliers” are dropped from a drone, they use a Miura-ori origami fold to switch from tumbling and dispersing outward through the air to dropping straight to the ground. To spread out the fliers, the researchers control the timing of each device’s transition using a few methods: an onboard pressure sensor (estimating altitude), an onboard timer or a Bluetooth signal.
Microfliers weigh about 400 milligrams — about half as heavy as a nail — and can travel the distance of a football field when dropped from 40 meters (about 131 feet) in a light breeze. Each device has an onboard battery-free actuator, a solar power-harvesting circuit and controller to trigger these shape changes in mid-air. Microfliers also have the capacity to carry onboard sensors to survey temperature, humidity and other conditions while soaring.
The team published these results Sept. 13 in Science Robotics.
“Using origami opens up a new design space for microfliers,” said co-senior author Vikram Iyer, UW assistant professor in the Paul G. Allen School of Computer Science & Engineering. “We combine the Miura-ori fold, which is inspired by geometric patterns found in leaves, with power harvesting and tiny actuators to allow our fliers to mimic the flight of different leaf types in mid-air. In its unfolded flat state, our origami structure tumbles chaotically in the wind, similar to an elm leaf. But switching to the folded state changes the airflow around it and enables a stable descent, similarly to how a maple leaf falls. This highly energy efficient method allows us to have battery-free control over microflier descent, which was not possible before.”
These robotic systems overcome several design challenges. The devices: are stiff enough to avoid accidentally transitioning to the folded state before the signal. transition between states rapidly. The devices’ onboard actuators need only about 25 milliseconds to initiate the folding. change shape while untethered from a power source. The microfliers’ power-harvesting circuit uses sunlight to provide energy to the actuator.The current microfliers can only transition in one direction — from the tumbling state to the falling state. This switch allows researchers to control the descent of multiple microfliers at the same time, so they disperse in different directions on their way down.
Future devices will be able to transition in both directions, the researchers said. This added functionality will allow for more precise landings in turbulent wind conditions.
Additional co-authors on this paper are Kyle Johnson and Vicente Arroyos, both UW doctoral students in the Allen School; Amélie Ferran, a UW doctoral student in the mechanical engineering department; Raul Villanueva, Dennis Yin and Tilboon Elberier, who completed this work as UW undergraduate students studying electrical and computer engineering; Alberto Aliseda, UW professor of mechanical engineering; Sawyer Fuller, UW assistant professor of mechanical engineering; and Shyam Gollakota, UW professor in the Allen School.
This research was funded by a Moore Foundation fellowship, the National Science Foundation, the National GEM Consortium, the Google fellowship program, the Cadence fellowship program, the Washington NASA Space Grant fellowship Program and the SPEEA ACE fellowship program. More

Researchers at Moorfields Eye Hospital and UCL Institute of Ophthalmology have developed an artificial intelligence (AI) system that has the potential to not only identify sight-threatening eye diseases but also predict general health, including heart attacks, stroke, and Parkinson’s disease.
RETFound, one of the first AI foundation models in healthcare, and the first in ophthalmology, was developed using millions of eye scans from the NHS. The research team are making the system open-source: freely available to use by any institution worldwide, to act as a cornerstone for global efforts to detect and treat blindness using AI. This work has been published in Nature today.
Progress in AI continues to accelerate at a dizzying pace, with excitement being generated by the development of ‘foundation’ models such as ChatGPT. A foundation model describes a very large, complex AI system, trained on huge amounts of unlabelled data, which can be fine-tuned for a diverse range of subsequent tasks. RETFound consistently outperforms existing state-of-the-art AI systems across a range of complex clinical tasks, and even more importantly, it addresses a significant shortcoming of many current AI systems by working well in diverse populations, and in patients with rare disease.
Senior author Professor Pearse Keane (UCL Institute of Ophthalmology and Moorfields Eye Hospital) said: “This is another big step towards using AI to reinvent the eye examination for the 21st century, both in the UK and globally. We show several exemplar conditions where RETFound can be used, but it has the potential to be developed further for hundreds of other sight-threatening eye diseases that we haven’t yet explored.
“If the UK can combine high quality clinical data from the NHS, with top computer science expertise from its universities, it has the true potential to be a world leader in AI-enabled healthcare. We believe that our work provides a template for how this can be done.”
AI foundation models have been called “a transformative technology” by the UK government in a report published earlier this year, and have come under the spotlight with the launch in November 2022 of ChatGPT, a foundation model trained using vast quantities of text data to develop a versatile language tool. Taking a comparable approach with eye images in a world-first, RETFound has been trained on millions of retinal scans to create a model that can be adapted for potentially limitless uses.
One of the key challenges when developing AI models is the need for expert human labels, which are often expensive and time-consuming to acquire. As demonstrated in the paper, RETFound is able to match the performance of other AI systems whilst using as little as 10% of human labels in its dataset. This improvement in label efficiency is achieved by using an innovative self-supervising approach in which RETFound masks parts of an image, and then learns to predict the missing portions by itself. More

A new simulation model that can predict flooding during an ongoing disaster more quickly and accurately than currently possible has been developed by University of Melbourne researchers.
Published in Nature Water, researchers say the new model has major potential benefits for emergency responses, reducing flood forecasting time from hours and days to just seconds, and enabling flood behaviour to be accurately predicted quickly as an emergency unfolds.
University of Melbourne PHD student Niels Fraehr, alongside Professor Q. J. Wang, Dr Wenyan Wu and Professor Rory Nathan, from the Faculty of Engineering and Information Technology, developed the Low-Fidelity, Spatial Analysis and Gaussian Process Learning (LSG) model to predict the impacts of flooding.
The LSG model can produce predictions that are as accurate as our most advanced simulation models, but at speeds which are 1000 times faster.
Professor Nathan said the development had enormous potential as an emergency response tool.
“Currently, our most advanced flood models can accurately simulate flood behaviour, but they’re very slow and can’t be used during a flood event as it unfolds,” said Professor Nathan, who has 40 years’ experience in engineering and environmental hydrology.” Professor Nathan said.
“This new model provides results a thousand times more quickly than previous models, enabling highly accurate modelling to be used in real-time during an emergency. Being able to access up-to-date modelling during a disaster could help emergency services and communities receive much more accurate information about flooding risks and respond accordingly. It’s a game-changer.”
When put to the test on two vastly different yet equally complex river systems in Australia, the LSG model was able to predict floods with a 99 per cent accuracy on the Chowilla floodplain in Southern Australia in 33 seconds, instead of 11 hours, and the Burnett River in Queensland in 27 seconds, instead of 36 hours, when compared to presently-used advanced models. More