McGill researchers develop a more efficient way to identify when AI responses may need human review
New approach cuts memory and training costs while maintaining performance
McGill University
McGill University researchers have developed a more energy-efficient method of building AI systems that are better at measuring – and indicating – their own uncertainty. This will help users determine when human oversight is needed, when additional data should be collected and when a model is being asked to work beyond the conditions it was trained for, the researchers said.
“Artificial intelligence systems now play a central role in daily life, from medical diagnosis and content moderation to autonomous driving and AI agents that act on our behalf,” said Mame Diarra Touré, lead author and PhD Candidate in the Department of Mathematics and Statistics. “As these systems take on more responsibility, they need to become more trustworthy. They should recognize when they are uncertain, rather than giving confident answers in situations where they may be wrong.”
The research was supervised by David A. Stephens, Professor in the Department of Mathematics and Statistics.
More reliable estimates
Standard neural networks learn patterns from data and make predictions, but they typically provide a single answer without clearly indicating how confident they are in that response. Bayesian neural networks address this limitation by representing their internal settings as probabilities rather than fixed values, enabling them to estimate uncertainty, particularly when faced with unfamiliar data. However, this capability often requires significant computational and memory resources, making these networks difficult to deploy at the scale of modern AI systems.
The researchers found a way to make Bayesian neural networks substantially more efficient while maintaining strong predictive performance. In one experiment, their approach used about 33 times fewer parameters than a commonly used method for estimating uncertainty in AI systems.
The results suggest that reliable, uncertainty-aware AI can be made practical even for the large and complex systems in use today, Touré said.
The researchers are now exploring ways to automate the process of identifying which parts of a neural network are most important for a given task. This could help the approach work more effectively across different kinds of data and AI tools, they said.
About this study
"Singular Bayesian Neural Networks,” by Mame Diarra Touré and David A. Stephens, was presented at the Forty-Third International Conference on Machine Learning (ICML 2026).
NB: Machine learning research follows a different publication model from many other scientific fields. The top peer-reviewed conferences are often the main archival publication venues, rather than a preliminary step before journal submission.
ICML papers are reviewed through a double-blind peer-review process, and accepted papers are published through the Proceedings of Machine Learning Research.
Method of Research
Computational simulation/modeling
Subject of Research
Not applicable
Article Title
Singular Bayesian Neural Networks
AI-powered terrain recognition helps cyborg cockroaches navigate faster
Real-time terrain classification allows biohybrid insects to climb obstacles and cross holes with fewer detours and less steering stimulation
The University of Osaka
video:
Movement of the cyborg insect using the proposed navigation system.
view moreCredit: Mochammad Ariyanto et al., 2026, Biohybrid Navigation through Real-Time Terrain Recognition and Natural Climbing in Cyborg Insect, Device
Osaka, Japan - Cyborg insects combine the mobility of living organisms with miniature electronic devices, offering potential applications in search-and-rescue operations, infrastructure inspection, and exploration of environments that are difficult for conventional robots. However, most autonomous navigation systems are designed to avoid obstacles, even when an insect could naturally climb over them. This can result in longer routes and reduced exploration efficiency.
An international research team from the University of Osaka and Universitas Diponegoro has developed a new navigation system for cyborg insects that combines the cockroach’s natural climbing ability with AI-based real-time terrain recognition, enabling the insects to traverse obstacles faster and more efficiently than with conventional navigation methods.
Conventional systems direct cyborg insects around obstacles, even when the insects can climb over them. The team therefore developed a reactive-climbing strategy combining goal-seeking, obstacle avoidance, wall-following, and innate climbing behavior. However, because the controller could not identify terrain, it issued steering commands during climbing, causing hesitation and inefficient movement.
To address this problem, the researchers incorporated a multilayer-perceptron-based AI module that used onboard sensor data to recognize flat surfaces, ascents, descents, and holes in real time. The classifier achieved 92% accuracy in offline evaluation, enabling the controller to adjust stimulation according to the terrain, reduced unnecessary steering, and support sustained forward movement across challenging surfaces.
“The main challenge was to develop a system capable of recognizing terrain in real time without compromising the insect’s natural locomotor abilities,” explains Professor Keisuke Morishima of the University of Osaka. “In this study, we propose ‘biohybrid physical AI’ to enable efficient autonomous navigation. We hope these findings will inspire the development of robotic systems capable of operating in complex environments, including search-and-rescue sites.”
###
The article, “Biohybrid Navigation through Real-Time Terrain Recognition and Natural Climbing in Cyborg Insect,” will be published in Device at DOI: https://doi.org/10.1016/j.device.2026.101277
About The University of Osaka
The University of Osaka was founded in 1931 as one of the seven imperial universities of Japan and is now one of Japan's leading comprehensive universities with a broad disciplinary spectrum. This strength is coupled with a singular drive for innovation that extends throughout the scientific process, from fundamental research to the creation of applied technology with positive economic impacts. Its commitment to innovation has been recognized in Japan and around the world. Now, The University of Osaka is leveraging its role as a Designated National University Corporation selected by the Ministry of Education, Culture, Sports, Science and Technology to contribute to innovation for human welfare, sustainable development of society, and social transformation.
Website: https://resou.osaka-u.ac.jp/en
Journal
Device
Method of Research
Computational simulation/modeling
Subject of Research
Animals
Article Title
Biohybrid Navigation through Real-Time Terrain Recognition and Natural Climbing in Cyborg Insect
Article Publication Date
20-Aug-2026
Advanced locomotion control strategy based on terrain recognition. (A) MLP-based terrain classifier for identifying flat ground, uphill, downhill, and holes. (B) Adaptive stimulation strategy that adjusts locomotion control based on the recognized terrain to enhance locomotion and climbing efficiency.
Credit
Mochammad Ariyanto et al., 2026, Biohybrid Navigation through Real-Time Terrain Recognition and Natural Climbing in Cyborg Insect, Device
Climbing cyborg insect
Credit
Mochammad Ariyanto et al., 2026, Biohybrid Navigation through Real-Time Terrain Recognition and Natural Climbing in Cyborg Insect, Device
Monell researchers use machine learning to begin mapping complex scents
Monell Chemical Senses Center
PHILADELPHIA, PA – Aug. 20, 2026 – If you want to describe a particular color, you could look to the Pantone color wheel to find its exact hue, saturation and brightness, and how it compares to other colors. But nothing like that has existed for complex odors.
Now, research co-authored by scientists at the Monell Chemical Senses Center has gotten closer to that work, creating a means of using machine learning to discriminate among scents and how they relate to each other. A description of the work was published online Aug. 4 in the Proceedings of the National Academy of Sciences.
The work provides a validated metric and benchmark for comparing smells, laying a foundation for technologies such as digital olfaction, the ability to digitize scents, said study co-author Joel Mainland, Ph.D., a Member of the Monell Center.
“We’ve been interested for a long time in trying to digitize odors, to mathematically represent these in some way similar to what we have done with color vision and for hearing,” Mainland said.
In 2015, IBM ran an open investigator DREAM challenge to take a single odor molecule and predict what it smells like based on its chemical structure, he said. That drove a lot of science. But most odors we encounter in day-to-day life are complex mixtures of dozens or hundreds of molecules. “We have to understand how mixtures work if we want to digitize anything,” Mainland said.
Being able to quantitatively map odor mixtures has numerous applications, he continued. For example, some conditions like diabetes and liver failure have olfactory signatures, distinct scents that could be helpful in diagnosis. Mapping scents also could be helpful in quantifying flavors of foods, or trademarking particular aromas like the scents of brand-name laundry detergents. “The companies that make smells are doing a lot of trial and error, so the thought process is that if you could fix that part where it’s more mathematical, they could make products more efficiently,” he said.
Mainland and colleagues released their own DREAM challenge, inviting international teams to use machine learning to develop models to predict the amount of similarity between two scent mixtures. First, they standardized and compiled six datasets of odor-similarity measurements from three different studies into a new dataset, comprising 168 unique single molecules, 731 unique mixtures, and 507 mixture-pair measurements. The mixture pair distances were mapped onto a continuous perceptual scale from 0 (indistinguishable) to 1 (most distinct).
Next, over a three-month period, 26 teams competed to predict how similar the paired scents would be on a hidden test-set of 46 mixture pairs. The competition resulted in a four-way tie. Mainland and colleagues then built an ensemble model by averaging the predictions from the four winning teams with those from two additional high-performing models. Following the challenge, they validated the model on an independent set of 50 olfactory mixture pairs.
Their final model was found to be quite accurate. It achieved a median RMSE (root mean squared error) of 0.08, with 0 being a perfect score, meaning the model had good accuracy in predicting scent similarities. It also achieved a Pearson correlation of 0.57 on the test set, meaning the model had a moderate to strong positive ability to predict the characteristics of the dataset.
While many in the sensory science field believed that using science to predict similarity among mixtures was going to be much more difficult than in single molecules, that didn’t turn out to be the case, Mainland said: “This paper shows that if you’re already able to predict what a single component smells like, then you can make a pretty good prediction of mixtures out of that.”
Interestingly, he said, the machine learning models were more likely to use semantic language like “fruity and sweet” to describe scent mixtures rather than chemistry terms like ester or molecular weights. “Working with these semantic labels is a huge jump forward in being able to make predictions of what mixtures smell like,” he said. When a journal reviewer asked them to eliminate the semantic features to see how the models performed without that information, it was much more difficult to make predictions.
Investigators are now writing up results from a third DREAM Olfaction challenge to submit for journal publication. In that exercise, teams were given people’s impressions of what individual components of mixtures smelled like and were asked to use that information to predict how the complete mixture would smell.
“What we’re seeing is that, again, if you know what the components smell like, you know what the mixture smells like — it’s basically just an average of the components, which I think is really surprising to a lot of people in the field,” he said.
Co-authors of the current study include Xuebo Song, Tiffany Yang and Robert Pellegrino of the Monell Center. Other contributing authors were from State University of Londrina in Brazil, Texas A&M University, Yale University, Boston University, the University of Michigan, KU Leuven in Belgium, Université Cote d’Azur in France, University of Oxford in the United Kingdom, KTH Royal Institute of Technology in Sweden, Cornell University, Cold Spring Harbor Laboratory, Sage Bionetworks, University of the Basque Country in Spain, the Basque Foundation for Science in Spain, the University of California at Davis, the University of Toronto in Canada, The Rockefeller University, the Weizmann Institute of Science in Israel, the University of Pennsylvania, and IBM Research.
The work was supported in part by grants from the National Institutes of Health (R01 DC017757), the NOMIS Foundation, Pershing Square Philanthropies, the Stavros Niarchos Foundation, Rothberg Catalyzer, Paul Graham Foundation, Schmidt Futures, ERC SynGrant 101118977 D2Smell, William R. Miller Fellowship, French National Research Agency (ANR-19-CE07-0044), Fondation Roudnitska, and the Initiative of Excellence Université Côte d’Azur (ANR-15-IDEX-01). Additional funders include the Howard Hughes Medical Institute, the College of LSA at the University of Michigan, and the Flemish government.
Mainland serves on the scientific advisory board of Osmo Labs, PBC, and receives compensation for these activities.
# # #
About the Monell Chemical Senses Center
The Monell Chemical Senses Center is an independent nonprofit research institute in Philadelphia, Pennsylvania. It was founded in 1968 to advance and share discoveries in the science of the chemical senses of smell, taste, chemesthesis, and interoception to solve the world’s health, societal, and environmental challenges.
Journal
Proceedings of the National Academy of Sciences
Method of Research
Data/statistical analysis
Subject of Research
Not applicable
Article Title
A Semantic-Based Community Model for High-Fidelity Tuning of Olfactory Mixture Distances
Article Publication Date
21-Aug-2026
Perspectives highlight the strength and limitations of AI-powered, “self-driving” labs
Summary author: Walter Beckwith
In two Perspectives, authors highlight the strengths and limitations of “self-driving” laboratories, spaces that combine robotics, high-throughput experiments, artificial intelligence (AI), and automated analysis to transform scientific discovery. They were envisioned to help discovery go from a slow process of trial and error into a continuous, adaptive learning cycle. These systems are already being used to discover molecules and optimize materials, sometimes exploring enormous experimental spaces while using far less material and time than conventional approaches.
In one Perspective, Milad Abolhasani highlights a major challenge in using AI tools in laboratories. These tools can generate hypotheses far faster than physical laboratories can test them. Linking instruments, robots, software, and shared datasets could allow laboratories to learn from one another, uncover patterns hidden in complex systems, and reduce redundant experiments. At the same time, greater autonomy will require rigorous safety measures, transparent decision-making, standardized data, and broader access to prevent these technologies from becoming concentrated among a small number of institutions. Ultimately, Abolhasani envisions autonomous laboratories augmenting scientists rather than replacing them, with human researchers setting goals and interpreting evidence while intelligent systems handle much of the experimental exploration.
In another Perspective, Martin Burke and colleagues highlight “blocc” chemistry. Blocc chemistry is a modular approach to building small molecules from standardized chemical building blocks that could make organic synthesis faster, more automated, and accessible to nonspecialists. By enabling robots to repeatedly assemble carbon-carbon bonds, Burke suggests that the method could be used to generate large, standardized datasets that AI systems could use to predict and optimize the properties of new molecules, creating a feedback loop between automated synthesis, testing, and machine learning. According to the author, the approach has already produced promising materials for applications including organic electronics and solar cells. What’s more, blocc chemistry could also democratize molecular discovery and transform chemistry education. However, broader access will require standardized methods as well as safeguards to ensure that automated chemical innovation is used safely and responsibly.
Journal
Science
Article Title
The lab that learns
Article Publication Date
20-Aug-2026
No comments:
Post a Comment