The Latest in AI and Machine Learning Developments

Keeping abreast of the rapidly evolving field of Artificial Intelligence is a challenge. This column aims to provide updates on significant AI and machine learning advancements globally, and to explain their potential impact on technology, startups, and society.

AI and the Arts: Piano Performance Recognition

Researchers are continually exploring creative applications of machine learning. A team at the University of Washington investigated whether a computer vision system could identify the music being played on a piano simply by observing the keys and the player’s hands from above.

The Audeo system, developed by Eli Shlizerman, Kun Su, and Xiulong Liu, analyzes video of piano playing. It initially creates a piano-roll-like sequence of key presses. Subsequently, it incorporates nuances like the duration and intensity of each press, refining the output for a MIDI synthesizer.

Image Credits: Shlizerman, et. al

Shlizerman noted that creating music that sounds authentically performed was previously considered unattainable. An algorithm must identify relevant cues within video frames and “imagine” the sound occurring between them, requiring both precision and creativity. The surprisingly good results were a welcome outcome.

Preserving History: Computational Unfolding of Ancient Letters

Another fascinating application lies in the realm of historical preservation. An MIT team tackled the challenge of digitally unfolding fragile, 17th-century letters intricately folded and sealed to prevent damage upon opening.

Their method involved X-raying the letters and employing a sophisticated algorithm to decipher the resulting imagery and virtually unfold them. “We weren’t sure it would be possible,” stated Erik Demaine of MIT, highlighting the algorithm’s ability to separate extremely thin layers of paper with minimal spacing.

Diagram showing X-ray views of a letter and how it is analyzed to virtually unfold it. Image Credits: MIT

This technique could be valuable for analyzing various delicate documents that are difficult to unravel using conventional X-ray methods. Further details are available in the Nature Communications paper.

Mapping EV Infrastructure Needs with NLP

Negative reviews of electric vehicle charging stations can reveal valuable insights. Thousands of online reviews represent a potential resource for municipalities planning to expand EV infrastructure.

Omar Asensio from Georgia Tech trained a natural language processing model to analyze these reviews, identifying common outage locations, cost comparisons, and other relevant factors. This approach addresses concerns about equitable access to EV infrastructure.

Asensio emphasized the importance of considering social equity and distributional issues when investing in EV infrastructure, suggesting that direct feedback from affected individuals is a crucial source of information.

Drone Safety: Maintaining Control with Limited Sensors

Unexpected service interruptions can lead to drone failures. Enhancing built-in safety mechanisms is paramount, particularly those independent of control signals or GPS. Researchers at the University of Zurich demonstrated that a damaged drone, equipped only with a camera and a functioning CPU, can maintain a significant degree of control.

A quadcopter experiencing a rotor failure can spin uncontrollably. However, the Swiss team, led by Sihao Sun, showed that an onboard camera can quickly analyze the surroundings during a spin, estimating the drone’s position based on the blurring scenery. More information is available from IEEE Spectrum.

AI-Assisted Medical Imaging: Improving Diagnostic Accuracy

The rapid generation of images in medical imaging often exceeds the capacity of doctors to thoroughly scrutinize them. AI is proving to be a valuable tool in this area, particularly in analyzing echocardiograms – ultrasound images of the heart.

A team at Geisinger Research developed an AI system to help doctors sort through thousands of images, aiding in diagnosis and prognosis. The system improved diagnostic accuracy by 13%, as reported in Nature Biomedical Engineering.

The extensive dataset used for training – nearly 50 million images – promises further advancements. The key discovery is the potential of using unstructured imagery databases to create AI systems that assist in decision-making.

Enhancing Transparency in Genomic Sequence Analysis

Verifying the accuracy of AI systems is easier when humans can readily understand the data being analyzed. However, analyzing complex data like DNA sequences presents a challenge, as humans may lack the expertise to confidently monitor the system’s performance.

Peter Koo and Matt Ploenzke at Cold Spring Harbor Laboratory explored methods to make machine learning systems analyzing genomic sequences more transparent. They focused on strongly training one layer of the convolutional neural network with familiar patterns, providing a reference point for subsequent analysis. These improvements in interpretability do not appear to compromise overall model effectiveness.

Addressing Bias in AI Moderation: The Case of Chess Terminology

AI systems are not immune to errors and biases. Researchers at CMU recently discovered that natural language processing systems, including those used by YouTube, were mistakenly flagging discussions about chess as inappropriate due to misinterpreting chess terminology.

Image Credits: Ahmad Hairi Mohamed/EyeEm (opens in a new window) / Getty Images

The use of terms like “white attacking black” can be misinterpreted by AI lacking contextual understanding. This highlights the importance of ensuring AI-powered moderation processes are accurate, explainable, and allow for open discussion.

Perception and Abstraction: The Art of Image Deletion

Understanding how AI perceives images is crucial. One approach involves progressively deleting portions of an image to determine the minimum information required for recognition. This concept was explored in an art project by Korean artists Shinseungback Kimyonghun.

Image Credits: Shinseungback Kimyonghun

The project demonstrates that, in some cases, the remaining image is surprisingly clear to the human eye, even after significant portions have been removed. This underscores the fundamental differences between machine and human perception.

Early Stage is the premier ‘how-to’ event for startup entrepreneurs and investors. You’ll hear first-hand how some of the most successful founders and VCs build their businesses, raise money and manage their portfolios. We’ll cover every aspect of company-building: Fundraising, recruiting, sales, product market fit, PR, marketing and brand building. Each session also has audience participation built-in – there’s ample time included for audience questions and discussion.