Staying Current with AI and Machine Learning Developments

The volume of news surrounding Artificial Intelligence is constantly expanding, making it challenging to remain fully informed. This column aims to provide a digestible overview of the most noteworthy AI and machine learning progress globally, and to contextualize its potential impact on technology, startups, and society as a whole.

The Global AI Landscape: A Comparative Analysis

Prior to delving into recent research, a study conducted by the Information Technology & Innovation Foundation (ITIF) offers a valuable assessment of the relative standings of the United States, the European Union, and China in the field of AI. The term "race" is used cautiously, as the ultimate goals and timelines remain undefined.

Currently, the U.S. maintains a leading position, primarily fueled by substantial private investment from major technology companies and venture capital firms. While China is rapidly increasing its financial commitment and research output, it still trails significantly.

A key factor impacting China’s position is its dependence on U.S.-manufactured silicon and infrastructure components.

EU's Position and Potential Investment Opportunities

The EU is progressing at a more modest pace, with comparatively smaller advancements, particularly in securing funding for AI-driven startups. While inflated valuations in the U.S. contribute to this disparity, a clear trend emerges.

This situation may present a unique opportunity for investors seeking to engage with promising startups that require less initial capital.

For those interested in a detailed examination of the data, the complete ITIF report (PDF) is available for review.

US AI Scholar Production and Partnerships

The recent establishment of a new AI research center at the University of Southern California, supported by Amazon funding, exemplifies the collaborative partnerships that bolster U.S. leadership in AI scholar development. Had this center been known during the ITIF study, it likely would have been highlighted as a positive example.

Exploring Novel Applications of Machine Learning

Moving away from direct profit and practical uses, machine learning is being applied in intriguing ways within fields that prioritize unique human skills and expertise.

Different mode styles are represented by each color. Image Credits: EPFL

Researchers at the Digital and Cognitive Musicology lab at EPFL in Switzerland investigated how the use of musical modes has evolved throughout classical music history – encompassing major, minor, and other variations, or even a complete absence of mode.

Analyzing Classical Music Modes

To objectively classify a vast collection of musical pieces spanning centuries and numerous composers, they developed an unsupervised machine learning system capable of listening to and categorizing these pieces based on their mode. Details regarding the data and methodologies employed are accessible on GitHub.

“Prior research indicated a greater diversity of modes during the Renaissance period. However, distinguishing between modes becomes increasingly challenging in eras following the Classical period. Our goal was to define these distinctions with greater precision,” stated a researcher in a press release from the university.

The model successfully identified several distinct clusters of mode usage, achieving this without any prior knowledge of music theory or historical context. Regardless of interpretation, this represents a novel and compelling approach to analyzing the classical music repertoire.

Human-Like Chess AI

Another enduring cultural element, chess, experienced a pivotal moment with the renowned match between Garry Kasparov and Deep Blue.

Currently, the focus has shifted from developing chess AIs with superior strength – as that goal has already been achieved – to creating AIs that emulate human playing styles. What specifically distinguished Kasparov as a player? As he detailed in his account of the Deep Blue match, it wasn't simply computational power.

Researchers at Cornell believe that replicating human skill presents a more complex and engaging challenge than simply exceeding it.

Their approach involved building an AI that comprehends the game from the perspective of a player with a specific skill level. A key aspect of this involved training the AI to concentrate on individual moves rather than solely focusing on achieving overall victory.

Understanding the moves made by players of varying skill levels, and the reasoning behind those moves, could lead to a more realistic opponent – and a more effective tool for instruction.

The resulting AI, named Maia, is reportedly available for play on lichess.org, though locating it has proven difficult.

The Expanding Role of Artificial Intelligence in Medical Diagnosis and Ecological Monitoring

The field of medical imaging is remarkably expansive, consistently showcasing new applications of machine learning. Identifying schizophrenia, however, presents a significantly greater challenge than detecting conditions like tumors or cataracts. Researchers at the University of Alberta believe they have developed a promising approach.

A computer-vision model, trained using MRI scans from patients diagnosed with schizophrenia, was applied to MRI imagery of their close relatives. Given the familial tendency for increased risk, the model successfully identified 14 individuals who exhibited the highest scores on a self-reported schizotypal personality trait scale.

This outcome is both encouraging and pragmatic. The expectation isn't for an AI to independently diagnose schizophrenia based solely on brain scans. However, discernible patterns within brain tissue appear to correlate with various illnesses, even if the underlying reasons remain unclear. AIs excel at pattern detection, and this model’s accuracy in aligning with other risk indicators suggests its potential as a valuable tool for medical professionals.

While identifying elephants might appear simpler, it isn’t without its difficulties, particularly from space. High-resolution orbital imagery holds substantial value for conservationists and ecologists, but its sheer volume necessitates automation for effective analysis.

Traditionally, this type of analysis has relied on the efforts of graduate students, research assistants, and considerable time investment. However, as previously observed, machine learning demonstrates proficiency in distinguishing animals from their surroundings, as exemplified by its success with dugongs.

Image Credits: Oxford University

Tracking elephants presents challenges due to their extensive range and diverse habitats. A team from Oxford University, collaborating with partners at Bath and ESA, proposes that, with current high-resolution imagery, elephant locations could be updated daily. This isn’t a replacement for human observation, but rather a method to significantly reduce the workload.

Identifying trees from an aerial perspective is another seemingly straightforward task with significant implications. Current forest management practices often involve manual site visits and ground-level surveys, or at best, utilize small aircraft or drones.

Image Credits: Skoltech

The Skolkovo Institute of Science and Technology in Russia has demonstrated the ability to identify dominant tree species within a given area using satellite data. Augmenting this data with height information, potentially from radar sources, further enhances accuracy. The resulting maps will be invaluable to forest ecologists, particularly when tracking changes over time. While mixed forests currently pose a challenge, this software already represents a significant advancement.

Enhancements to Icon Accessibility

This consideration arose during the development of previous features. Google is currently developing a computer vision system designed to identify and categorize a wide array of frequently used icons, including those for functions like forward, backward, options, and list views.

The level of accessibility provided on mobile devices varies considerably. Certain applications and services meticulously label all icons and buttons, while others offer minimal or insufficient labeling. If the operating system of the phone could supplement this information, it would be a significant benefit for individuals with visual impairments.

In some instances, the phone can utilize data extracted from the application's code. However, the ability to simply identify an icon – for example, recognizing it as an undo button – would be incredibly valuable.

Image Credits: Google

Google is actively pursuing this capability with IconNet, a specialized and efficient convolutional neural network. This network is designed to be compact and capable of recognizing over 70 icons within milliseconds, while minimizing computational demands.

This development appears promising. Therefore, it is perhaps unsurprising that Apple introduced a comparable feature with the release of iOS 14 in December.

Despite not always being first to market, the ultimate beneficiaries are the end-users. Increased competition within the field of artificial intelligence is leading to expanded accessibility options for everyone.