The sheer volume of research publications appearing today makes it impossible for any individual to stay fully current, a challenge particularly acute in machine learning. This discipline now influences – and generates scholarly work within – nearly every sector and organization. This feature intends to gather and clarify the significance of the most impactful recent findings and papers, with a primary focus on, but not limited to, advancements in artificial intelligence.

The focus this week leans toward foundational research rather than direct consumer-facing applications. While machine learning offers numerous benefits readily experienced by end-users, its potential for significant change extends to fields such as seismology and biology. These areas possess substantial, previously untapped datasets that can be utilized to develop artificial intelligence models or serve as a source for valuable data analysis.

Inside earthshakers

We are constantly surrounded by natural events that remain incompletely understood—while the origins of earthquakes and storms are generally known, the precise mechanisms of their development are still being investigated. How do these events spread, and what additional consequences arise when different measurements are compared? To what extent can we foresee these occurrences in advance?

Several recently completed research initiatives have employed machine learning techniques in an attempt to gain a more comprehensive understanding of, or improve predictions for, these natural phenomena. The availability of data spanning many decades offers opportunities for new discoveries across various fields—provided that seismologists, meteorologists, and geologists can secure the necessary funding and specialized knowledge.

A recent finding, originating from researchers at Los Alamos National Labs, utilizes a novel data source alongside machine learning to document previously unobserved activity along fault lines during “slow slip events.” By analyzing synthetic aperture radar imagery obtained from satellites—capable of penetrating cloud cover and operating at night to provide consistent, detailed ground surface mapping—the team was able to directly observe “fracture propagation” for the first time, specifically along the North Anatolian Fault in Turkey.

“The deep learning methodology we created allows for the automated identification of the subtle and fleeting ground deformations occurring on faults with exceptional precision, opening the door to a systematic investigation of the relationship between slow and conventional earthquakes on a worldwide basis,” explained Bertrand Rouet-Leduc, a geophysicist at Los Alamos.

Another ongoing project, based at Stanford and continuing for several years, assists Earth science researcher Mostafa Mousavi in addressing the challenge of separating meaningful signals from noise within seismic data. After repeatedly reviewing data processed by existing software, he recognized the need for a more effective approach and has since dedicated years to developing various methods. His latest innovation is a technique for identifying evidence of minor earthquakes that previously went undetected, yet still left traces in the data.

The “Earthquake Transformer” (a name derived from a machine learning concept, rather than robotic devices) was trained using years of manually categorized seismographic data. When applied to recordings from Japan’s magnitude 6.6 Tottori earthquake, it identified 21,092 distinct events—more than double the number initially discovered through conventional analysis—and achieved this using data from fewer than half of the recording stations.

Image Credits: Stanford University

This tool does not independently predict earthquakes; however, a more thorough understanding of the true and complete nature of these events could potentially lead to improved predictive capabilities. “By enhancing our ability to detect and pinpoint these very small earthquakes, we can gain a more detailed understanding of how earthquakes interact, spread along faults, initiate, and ultimately cease,” stated co-author Gregory Beroza.

When “good enough” is better

Predicting the weather is another well-known challenge with inherent unpredictability. A significant difficulty lies in the vast quantities of data required for accurate forecasting. While a meteorologist can reasonably estimate conditions for the next few days based on local patterns and measurements, longer-range predictions are considerably more complex, often relying on the convergence of years-long trends within incredibly intricate systems.

Image showing how the globe is divided into easier geometrical units for simulation by the UW system. Image Credits: Weyn et al./UW

Researchers at the University of Washington have found an innovative solution, or at least a substantial advancement, by adjusting their expectations and leveraging the strengths of machine learning. They developed a model using 40 years of historical weather data, without incorporating information about physics or the underlying mechanisms, and discovered that although it doesn't match the performance of leading forecasting systems, it can generate adequately reliable predictions using significantly fewer computational resources.

This approach might not be ideal for routine weekly forecasts, but its speed allows meteorologists to execute numerous simulations in the time typically required for a single run, enabling broader coverage and consideration of a wider range of potential scenarios. This capability is particularly valuable in areas like hurricane prediction. Currently, the model is still undergoing refinement, but the methodology appears very encouraging (even if it isn't intended for direct consumer applications).

Machine learning models also demonstrate proficiency in efficiently performing “good enough” versions of repetitive tasks, specifically in transforming one type of image into another. This is evident in applications like converting sketches into fantastical creatures or rendering photographs in the style of paintings, all of which showcase the ability to rapidly convert data from one format to another.

A specialized application of this concept involves converting two-dimensional scans of bodies and organs into three-dimensional representations. For instance, when evaluating a medication designed to prevent infections from causing pancreatic inflammation, researchers need to determine the pancreas’s location relative to other organs within a scan, and assess its size and volume. This process can be lengthy and prone to inaccuracies.

Image Credits: Astrid Eckert/TUM

New software developed by the Technical University of Munich automates this process, enabling an AI to estimate the positions and dimensions of bones and organs. “We only required approximately 10 full-body scans before the software could independently and successfully analyze the image data—and do so in a matter of seconds. A human would typically take hours to accomplish this,” explained TUM researcher Oliver Schoppe. The resulting estimations were even found to be more precise than those made by human analysts. The findings of this research were published in the journal Nature Communications.

The limits of the black box

In all of these instances, artificial intelligence demonstrates effectiveness and utility, but the methods by which it arrives at its conclusions often lack clear explanation. While this isn't a significant concern when reconstructing mouse organs, it becomes increasingly vital when an algorithm is making decisions with greater implications, to ensure there is no inherent bias or manipulation within the process.

Image Credits: Stanford University

Researchers at Duke University are developing a system that allows an AI to reveal its reasoning, essentially showing its work. For example, when a computer vision algorithm identifies a photograph as depicting a bedroom, what specific elements led to that conclusion? Was it the presence of a bed? The bed linens? Perhaps a nightstand with an alarm clock? Or are more subtle factors influencing the outcome? The most apparent answer (the bed itself!) isn't always accurate, as illustrated by the well-known case of the “sheep in a grassy field” (an AI incorrectly identified grassy areas as containing sheep, due to its association of grass with sheep).

Employing the Duke University approach, the AI would not only provide its final determination but also indicate which parts of its “memory” were engaged during the decision-making process. This would allow for the identification of misleading information or flawed logical steps, which would be reported alongside the final result.

Furthermore, some specialists believe that regardless of whether AI operates as a “black box,” there should be defined boundaries regarding its applications. Bryan Ford of EPFL asserted that inherent limitations in AIs render them unsuitable for tasks such as policy creation.

“The realm of policy concerning human governance should remain exclusively within the domain of human decision-making,” he stated during the Governance of and by Digital Technology virtual conference. “This is because machine-learning algorithms learn from data sets reflecting past experiences, meaning AI-driven policy is fundamentally limited by the assumption that the past accurately represents the ideal or only foundation for future decisions. However, we recognize that all past and present societies are imperfect, and therefore, genuine societal improvement requires governance that is both visionary and forward-thinking.”

Stuart Russell, a long-standing and highly respected researcher in the field and author of a notable book on the subject (I spoke with him earlier this year), highlights that nearly all AI systems treat users as both customers and commodities:

“Currently, AI from fifty different companies resides within your mobile device, actively collecting your information and monetizing it as quickly as possible, without any software dedicated to representing your interests. Could we restructure this so that the software on your phone prioritizes your needs and negotiates with these other entities to protect your data?”

At a minimum, exploring this possibility is worthwhile.