This week’s list of data news highlights covers May 13-19, 2017, and includes articles about a new AI tool that can reduce bullying in videogames and an initiative to use machine learning to aid in the fight against ISIS.
1. AI Learns to Make New Kinds of Sound
Google Magenta, a team of researchers developing AI that can make art, has developed a system called NSynth that uses neural networks to analyze mathematical characteristics of notes produced by different instruments, allowing users to mix and match these characteristics to produce entirely new kinds of sounds that would be impossible to produce from standard synthesizers. The team trained NSynth on an annotated dataset of 300,000 notes from 1,000 different instruments, and it has also made this dataset freely available.
2. Centralizing Federal Student Data
A bipartisan group of U.S. senators have introduced a bill called the College Transparency Act of 2017 to direct the National Center for Education Statistics to develop a secure system for linking databases of student records maintained by multiple federal agencies. The bill would overturn a ban of such a system established by the 2008 reauthorization of the Higher Education Act, which has limited the government’s ability to evaluate graduation and employment outcomes across the country despite the fact that universities have to report this data already.
3. Making it Easier for AI to Have Conversations
Facebook has developed an open source platform called ParlAI (pronounced “parlay”) designed to make it easier for AI developers to train natural language processing systems by combining different training approaches. ParlAI includes built-in training and benchmarking datasets developed by Microsoft, Stanford University, and Facebook, enables developers to integrate popular machine-learning libraries, and integrates with Amazon’s Mechanical Turk crowdsourcing service to allow developers to easily recruit humans to help train their systems.
4. Building a Common Database for Understanding the Human Genome
A team from Baylor College, Texas Children’s Hospital, and Harvard University have developed a tool called Model Organism Aggregated Resource for Rare Variant Exploration (MARRVEL) that combines six different research databases of human genomes and aggregates this data in a common format to make it easier for medical researchers to access and study useful data. Different databases use different formats to store genetic information, making it difficult for researchers to find relevant data across multiple systems. MARRVEL functions like a search engine for these different databases and is capable of providing summaries of relevant data to reduce the time and effort it takes for a researcher to access genetic data.
5. Machine Learning Joins the Fight Against ISIS
The Algorithmic Warfare Cross Functional Team (AWCFT) at the United States Department of Defense has launched a new initiative called Project Maven to use machine learning to help extract intelligence from aerial imagery aid efforts to fight ISIS. Project Maven will focus on reducing the time human analysts spend on administrative aspects of intelligence analysis, such as manually transcribing data. AWCFT will develop or procure this machine learning system within 90 days, and then will work on procuring the necessary hardware to implement the system.
6. Talking to AI Without Making Noise
Researchers at the University of Bristol working on a project called Project Telepathy have developed a machine learning system capable of figuring out what a person is saying by analyzing subtle movements in his or her facial muscles. The system analyzes data from a series of electrodes on a subject’s face, and a specialized speaker mounted on a subject broadcasts this information using ultrasonic waves that can only be received by another subject with the same kind of speaker, meaning when one subject silently mouths a word, a second subject receives an audio broadcast of that word. In initial tests, the system can recognize ten words with 80 percent accuracy.
7. Sensing Sounds for the Deaf
Fujitsu has developed a wearable device called Ontenna that uses audio sensors to detect noises and vibrates to provide nuanced haptic feedback for wearers, potentially allowing deaf people to be more aware of and responsive to sounds. The sensor clips onto a wearer’s clothes or hair and will automatically produce one of 256 different kinds of haptic feedback to best mimic a sound’s rhythm and volume.
8. Helping Humans Coordinate with Unpredictable AI
Researchers at Yale University have conducted a study showing that an AI system working with humans could improve how effectively humans coordinate by introducing a certain level of randomness in its decision-making. The researchers had groups of people and AI-powered bots play a game that required all players to change the color of a node so that it was different from that of adjacent players without communicating. The researchers found that when bots were instructed to pick colors randomly 10 percent of the time, rather than make decisions based on other observable colors, the group as a whole completed the game within the five-minute time limit 85 percent of the time, whereas an all-human group, or groups with bots with no or high degrees of randomness, only successfully completed the game 67 percent of the time.
9. Predicting When Patients Will Fall
El Camino Hospital in California partnered with analytics firm Qventus to develop a predictive analytics system that uses machine learning to identify patients most likely to fall, which can cause injuries and health complications. The system analyzes data from patient’s electronic health records and data from the system hospital beds use to call nurses to develop a method for assigning fall risk scores to patients, which the hospital used to take preventative action, such as moving a patient closer to a nurse’s station or putting certain patients under constant supervision. With this system, El Camino was able to reduce the number of falls by 39 percent in six months.
10. Preventing Cyber Bullying with AI
A startup called Spirit AI has developed an AI system called Ally that can monitor interactions between players in online video games, detect signs of abuse, and flag it for game developers. In addition, Ally can control an avatar in the game that uses natural language processing to check in with bullied players. Ally can detect both verbal abuse and nonverbal harassment, such as a player filing false reports about another player.
Image: Ville Hyvönen.