(Press-News.org) CAMBRIDGE, MA – Because machine-learning models can give false predictions, researchers often equip them with the ability to tell a user how confident they are about a certain decision. This is especially important in high-stake settings, such as when models are used to help identify disease in medical images or filter job applications.
But a model’s uncertainty quantifications are only useful if they are accurate. If a model says it is 49% confident that a medical image shows a pleural effusion, then 49% of the time, the model should be right.
MIT researchers have introduced a new approach that can improve uncertainty estimates in machine-learning models. Their method not only generates more accurate uncertainty estimates than other techniques, but does so more efficiently.
In addition, because the technique is scalable, it can be applied to huge deep-learning models that are increasingly being deployed in health care and other safety-critical situations.
This technique could give end users, many of whom lack machine-learning expertise, better information they can use to determine whether to trust a model’s predictions or if the model should be deployed for a particular task.
“It is easy to see these models perform really well in scenarios where they are very good, and then assume they will be just as good in other scenarios. This makes it especially important to push this kind of work that seeks to better calibrate the uncertainty of these models to make sure they align with human notions of uncertainty,” says lead author Nathan Ng, a graduate student at the University of Toronto who is a visiting student at MIT.
Ng wrote the paper with Roger Grosse, an assistant professor of computer science at the University of Toronto; and senior author Marzyeh Ghassemi, an associate professor in the Department of Electrical Engineering and Computer Science and a member of the Institute of Medical Engineering Sciences and the Laboratory for Information and Decision Systems. The research will be presented at the International Conference on Machine Learning.
Quantifying uncertainty
Uncertainty quantification methods often require complex statistical calculations that don’t scale well to machine-learning models with millions of parameters. These methods also require users to make assumptions about the model and data used to train it.
The MIT researchers took a different approach. They use what is known as the minimum description length principle (MDL), which does not require the assumptions that can hamper the accuracy of other methods. MDL is used to better quantify and calibrate uncertainty for test points the model has been asked to label.
The technique the researchers developed, known as IF-COMP, makes MDL fast enough to use with the kinds of large deep-learning models deployed in many real-world settings.
MDL involves considering all possible labels a model could give a test point. If there are many alternative labels for this point that fit well, its confidence in the label it chose should decrease accordingly.
“One way to understand how confident a model is would be to tell it some counterfactual information and see how likely it is to believe you,” Ng says.
For example, consider a model that says a medical image shows a pleural effusion. If the researchers tell the model this image shows an edema, and it is willing to update its belief, then the model should be less confident in its original decision.
With MDL, if a model is confident when it labels a datapoint, it should use a very short code to describe that point. If it is uncertain about its decision because the point could have many other labels, it uses a longer code to capture these possibilities.
The amount of code used to label a datapoint is known as stochastic data complexity. If the researchers ask the model how willing it is to update its belief about a datapoint given contrary evidence, the stochastic data complexity should decrease if the model is confident.
But testing each datapoint using MDL would require an enormous amount of computation.
Speeding up the process
With IF-COMP, the researchers developed an approximation technique that can accurately estimate stochastic data complexity using a special function, known as an influence function. They also employed a statistical technique called temperature-scaling, which improves the calibration of the model’s outputs. This combination of influence functions and temperature-scaling enables high-quality approximations of the stochastic data complexity.
In the end, IF-COMP can efficiently produce well-calibrated uncertainty quantifications that reflect a model’s true confidence. The technique can also determine whether the model has mislabeled certain data points or reveal which data points are outliers.
The researchers tested their system on these three tasks and found that it was faster and more accurate than other methods.
“It is really important to have some certainty that a model is well-calibrated, and there is a growing need to detect when a specific prediction doesn’t look quite right. Auditing tools are becoming more necessary in machine-learning problems as we use large amounts of unexamined data to make models that will be applied to human-facing problems,” Ghassemi says.
IF-COMP is model-agnostic, so it can provide accurate uncertainty quantifications for many types of machine-learning models. This could enable it to be deployed in a wider range of real-world settings, ultimately helping more practitioners make better decisions.
“People need to understand that these systems are very fallible and can make things up as they go. A model may look like it is highly confident, but there are a ton of different things it is willing to believe given evidence to the contrary,” Ng says.
In the future, the researchers are interested in applying their approach to large language models and studying other potential use cases for the minimum description length principle.
END
When to trust an AI model
More accurate uncertainty estimates could help users decide about how and when to use machine-learning models in the real world
2024-07-12
ELSE PRESS RELEASES FROM THIS DATE:
Research shows gamified investment sites have risks for novice investors
2024-07-12
TORONTO - What happens when online investment trading platforms start to resemble games that keep people playing for hours, with badges and exploding confetti to reward investors for their engagement?
For those who know what they’re doing, it won’t make much of a difference. New research from the University of Toronto engaging nearly 1,000 volunteers in artificial investment scenarios shows that more informational features such as price change notifications might even help savvy investors execute ...
Specially equipped natural killer cells show effectiveness against the most common form of ovarian cancer
2024-07-12
RESEARCH SUMMARY
Study Title: CAR memory-like NK cells targeting the membrane proximal domain of mesothelin demonstrate promising activity in ovarian cancer
Publication: Science Advances
Dana-Farber Cancer Institute authors include: Rizwan Romee, MD, senior author; and Mubin Tarannum, PhD, KhanhLinh Dinh, and Juliana Vergara, MD, MMSc, co-first authors
Summary: Natural killer, or NK, cells endowed with memory-like abilities and armed with a novel chimeric antigen receptor (CAR) have generated encouraging results in experiments in epithelial ovarian cancer ...
Entering the golden age for antibody-drug conjugates in gynecologic cancer
2024-07-12
“We are optimistic that the incorporation of ADCs into the treatment of aggressive tumors and treatment refractory gynecologic cancers will improve quality of life and survival outcomes in our patients.”
BUFFALO, NY- July 12, 2024 – A new editorial paper was published in Oncoscience (Volume 11) on May 20, 2024, entitled, “Entering the golden age for antibody-drug conjugates in gynecologic cancer.”
In this new editorial, researchers Michelle Greenman, Blair McNamara, Levent Mutlu, and Alessandro D. Santin from Yale University School of Medicine discuss gynecologic cancers. Biologically aggressive ...
Judge: Texas university must release records on research study that resulted in deaths of dozens of animals
2024-07-12
SAN ANGELO, Texas —Tom Green County District Court Judge Barbara L. Walther ruled Thursday, July 11, 2024, that Angelo State University must release public records relating to an experiment conducted on dozens of mice that resulted in the animals’ unnecessary suffering and death, reportedly to study the impact of the foster care system on human children.
The ruling overturns Texas Attorney General Ken Paxton’s Nov. 17, 2022 decision to side with the university in denying the records.
On July 13, 2023, the Physicians Committee for Responsible Medicine, a Washington, D.C. based health advocacy group of more than 17,000 doctor members that encourages higher standards ...
UMass Amherst food scientist rises to the challenge of giving marbled fatty feel and taste to plant-based meat
2024-07-12
One of the challenges of creating realistic-looking and delectable plant-based meat is mimicking the marbled effect of animal fat that many carnivores expect and enjoy.
A University of Massachusetts Amherst food scientist has a plan to tackle this quandary by developing new technology supported by a $250,000 grant from the Good Food Institute. The not-for-profit think tank promotes plant-based alternatives to meat, dairy and eggs, as well as cultivated “clean meat” grown from animal cells in a facility.
The technology proposed ...
Complex impact of large wildfires on ozone layer dynamics unveiled by new study
2024-07-12
In a revelation that highlights the fragile balance of our planet's atmosphere, scientists from China, Germany, and the USA have uncovered an unexpected link between massive wildfire events and the chemistry of the ozone layer. Published in Science Advances, this study reveals how wildfires, such as the catastrophic 2019/20 Australian bushfires, impact the stratosphere in previously unseen ways.
The ozone layer, a crucial shield protecting life on Earth from harmful ultraviolet (UV) radiation, has been on a path to ...
Brain inflammation triggers muscle weakness after infections
2024-07-12
Infections and neurodegenerative diseases cause inflammation in the brain. But for unknown reasons, patients with brain inflammation often develop muscle problems that seem to be independent of the central nervous system. Now, researchers at Washington University School of Medicine in St. Louis have revealed how brain inflammation releases a specific protein that travels from the brain to the muscles and causes a loss of muscle function.
The study, in fruit flies and mice, also identified ways to block this process, which could have ...
Research alert: All stem cell therapies are not created equal
2024-07-12
Researchers from University of California San Diego have found that two of the most frequently administered stem cell therapies, which are often used interchangeably, actually contain completely different types of cells. The results challenge the current “one-cell-cures-all” paradigm in orthopedic stem cell therapeutics and highlight the need for more informed and rigorous characterization of injectable stem cell therapies before they are marketed for use in patients.
The researchers analyzed cell populations of autologous bone marrow aspirate concentrate (BMAC) and adipose-derived ...
Complex impact of large wildfires on ozone layer dynamics
2024-07-12
The ozone layer, a crucial shield protecting life on Earth from harmful ultraviolet (UV) radiation, has been on a path to recovery thanks to the Montreal Protocol. This landmark international treaty, adopted in 1987, successfully led to phasing out the production of numerous substances responsible for ozone depletion. Over the past decades, the ozone layer has shown significant signs of healing, a testament to global cooperation and environmental policy.
However, the stability of this vital atmospheric layer is now facing a new and unexpected challenge. During the 2019/20 Australian wildfires, ...
AI found to boost individual creativity – at the expense of less varied content
2024-07-12
Stories written with AI assistance have been deemed to be more creative, better written and more enjoyable.
A new study published in the journal Science Advances finds that AI enhances creativity by boosting the novelty of story ideas as well as the ‘usefulness’ of stories – their ability to engage the target audience and potential for publication.
It finds that AI “professionalizes” stories, making them more enjoyable, more likely to have plot twists, better written and less boring.
In ...
LAST 30 PRESS RELEASES:
Antiretroviral drugs for treatment and prevention of HIV in adults: 2024 recommendations of the International Antiviral Society–USA Panel
Virginie McNamar appointed President and Chief Operating Officer (COO) of SynGAP Research Fund (SRF) dba Cure SYNGAP1
Planetary boundaries: Confronting the global crisis of land degradation; Potsdam institute report opens UNCCD COP 16
VUMC and Philips landmark emissions assessment of a radiology department published in premier industry journal
When devices can read human emotions without a camera
Warming temperatures impact immune performance of wild monkeys, U-M study shows
Fine particulate air pollution may play a role in adverse birth outcomes
Sea anemone study shows how animals stay ‘in shape’
KIER unveils catalyst innovations for sustainable turquoise hydrogen solutions
Bacteria ditch tags to dodge antibiotics
New insights in plant response to high temperatures and drought
Strategies for safe and equitable access to water: a catalyst for global peace and security
CNIO opens up new research pathways against paediatric cancer Ewing sarcoma by discovering mechanisms that make it more aggressive
Disease severity staging system for NOTCH3-associated small vessel disease, including CADASIL
Satellite evidence bolsters case that climate change caused mass elephant die-off
Unique killer whale pod may have acquired special skills to hunt the world’s largest fish
Emory-led Lancet review highlights racial disparities in sudden cardiac arrest and death among athletes
A new approach to predicting malaria drug resistance
Coral adaptation unlikely to keep pace with global warming
Bioinspired droplet-based systems herald a new era in biocompatible devices
A fossil first: Scientists find 1.5-million-year-old footprints of two different species of human ancestors at same spot
The key to “climate smart” agriculture might be through its value chain
These hibernating squirrels could use a drink—but don’t feel the thirst
New footprints offer evidence of co-existing hominid species 1.5 million years ago
Moral outrage helps misinformation spread through social media
U-M, multinational team of scientists reveal structural link for initiation of protein synthesis in bacteria
New paper calls for harnessing agrifood value chains to help farmers be climate-smart
Preschool education: A key to supporting allophone children
CNIC scientists discover a key mechanism in fat cells that protects the body against energetic excess
Chemical replacement of TNT explosive more harmful to plants, study shows
[Press-News.org] When to trust an AI modelMore accurate uncertainty estimates could help users decide about how and when to use machine-learning models in the real world