AI in Everything: Deep Science and its Impact

Recent Advances in Machine Learning and Artificial Intelligence

The volume of published research in machine learning is increasing at a rapid pace, making it challenging for professionals to stay current. This column aims to highlight and explain the significance of recent discoveries and papers, with a focus on artificial intelligence and its applications across various industries.

This week’s overview features several novel applications of machine learning, alongside a noteworthy instance where the methodology proved unsuitable for pandemic-related analysis.

Machine Learning in Government Regulation

The integration of machine learning into government regulation may seem unexpected, given the perceived technological lag in federal agencies. However, the U.S. Environmental Protection Agency (EPA) is collaborating with Stanford researchers to identify and address violations of environmental regulations using algorithmic methods.

The scale of the task necessitates this approach. EPA officials must analyze millions of permits and reports related to Clean Water Act compliance, including self-reported pollution data and independent lab findings. The Stanford-developed system identifies patterns, such as which facilities in specific locations pose the greatest risk to certain demographics. For example, wastewater treatment plants in urban areas may underreport pollution levels, disproportionately impacting communities of color.

The process of framing compliance as a computationally analyzable problem clarified the agency’s priorities. While the technique can pinpoint numerous minor violations, it also revealed that focusing on broader permit categories that enable significant pollution from multiple sources may be more effective.

Automating Scrap Metal Processing

Processing scrap metal is a substantial source of waste and expense. Much of the work in sorting and recycling centers is still performed manually, presenting both safety hazards and monotony for workers. Eversteel, a startup originating from the University of Tokyo, is developing an automated system to streamline this process, reducing the need for human intervention in initial sorting stages.

Eversteel employs a computer vision system to categorize incoming scrap into approximately two dozen types and to flag impure or unrecyclable materials for removal. Although still in its early phases, the lack of existing large datasets for training their models—necessitating the creation of a new dataset informed by steelworkers and imagery—demonstrated the potential for AI innovation in this traditionally tech-underdeveloped industry. Successful commercialization could attract the necessary funding for wider implementation.

Computer Vision for Soil Monitoring

Soil monitoring is a routine task for farmers, typically involving the assessment of water and nutrient levels. Current automated solutions often rely on complex and expensive equipment. Researchers from the University of South Australia and Middle Technical University in Baghdad propose a simpler, more cost-effective approach.

Their research indicates that a standard RGB digital camera can accurately estimate soil moisture by analyzing its color. Ali Al-Naji, a contributor to the study, stated, “We tested it at different distances, times and illumination levels, and the system was very accurate.” This technology could facilitate the development of affordable smart irrigation systems, improving crop yields for farmers with limited resources.

Non-Contact Heartbeat Monitoring with Smart Speakers

Smart speakers are ubiquitous, but researchers at the University of Washington have developed a unique application: remote heartbeat monitoring using ultrasound. This device can detect arrhythmia and potentially other cardiac issues with reasonable accuracy.

Anran Wang, the lead author and a UW graduate student, explained the challenge: “The motion from someone’s breathing is orders of magnitude larger on the chest wall than the motion from heartbeats, so that poses a pretty big challenge. And the breathing signal is not regular so it’s hard to simply filter it out.” By leveraging the multiple microphones in smart speakers, they created a beam-forming algorithm to isolate heartbeat signals.

The system utilizes a self-supervised machine learning model to extract the relevant signal from background noise. Shyam Gollakota, a UW professor, expressed his surprise at the system’s success. A low-cost, noninvasive home-based test could aid in the early detection of heart conditions.

Returning to Established Areas of Machine Learning

Machine learning excels at processing large datasets, and unsupervised learning—extracting patterns without human guidance—is gaining prominence. Researchers at Los Alamos National Labs have developed SmartTensors, a tool designed for this purpose, capable of handling datasets on the terabyte scale.

Unlike supervised learning, which requires labeled examples, unsupervised learning allows the system to discover its own patterns, potentially revealing insights that humans might overlook. The LANL system is applicable to diverse scientific disciplines, including seismology and text analysis.

The Julia-based factorization and feature finder frameworks are publicly available for those with the requisite expertise.

Interestingly, despite expectations, efforts to apply machine learning to COVID-19 diagnosis have largely failed. A Cambridge report analyzing hundreds of studies found that none were clinically viable due to methodological flaws and biases.

This is not necessarily a failure, but rather an indication that the approach requires refinement. Applying cutting-edge technology to a rapidly evolving pandemic inevitably involves numerous unsuccessful attempts. The researchers hope their recommendations will improve future success rates.

Deepfake Detection Through Eye Reflections

The ongoing arms race between creating and detecting deepfakes has become a distinct field of study. Researchers at the University of Buffalo have developed a technique for identifying deepfakes by analyzing the consistency of reflections in the eyes of the depicted individuals.

In deepfakes, these reflections often exhibit inconsistencies that generative adversarial networks fail to replicate. However, future generations of deepfakes will likely address this vulnerability, continuing the cycle of innovation and countermeasure.

Advancements in Video Understanding at Facebook

Facebook, with its vast video dataset, is pushing the boundaries of video understanding. The company is adapting techniques successful with still images and short clips to analyze longer videos.

Understanding video for Facebook involves connecting it to context and related videos. Multiple videos may depict different actions in different locations, but shared audio—such as a common song—can reveal a unifying theme. Integrating visual and auditory data enhances content discovery and organization.

The audio component also provides access to a more naturalistic and diverse range of language compared to curated language model training datasets. Facebook’s wav2vec 2.0 model learns quickly and outperforms traditional models in understanding real-world audio.

Effective language understanding is crucial for features like real-time captioning in Messenger and VR, as well as improving ad targeting. Facebook is currently developing a large model supporting 25 languages, with further announcements anticipated.

Finally, congratulations to Danish researcher Christian Wulff-Nilsen for his near-optimal solution to the Single-Source Shortest Path problem, demonstrating that fundamental graph navigation techniques can still be improved.