This week’s list of data news highlights covers July 14-20, 2018, and includes articles about a company using AI to dicver new drugs and a software toolkit for programming quantum computers.
1. Predicting Schizophrenia Treatment Outcomes
Researchers at the University of Alberta and the University of Texas have developed a machine learning system that can analyze magnetic resonance imaging (MRI) scans of patients’ brains and both diagnose schizophrenia and predict whether a patient with schizophrenia would respond well to a particular treatment with high degrees of accuracy. The system analyzes connections between the brain’s superior temporal cortex and other regions of the brain, and can identify patients with schizophrenia with 78 percent accuracy. The system can also identify whether a schizophrenia patient would benefit from an antipsychotic medication called risperidone with 82 percent accuracy.
A startup called Verge Genomics is using AI and genetic sequencing to accelerate the process of discovering new drugs. Verge focuses on mapping the complex interplay of genes that cause diseases, such as the hundreds of genes linked to Alzheimer’s, and unlike many other drug research initiatives, Verge relies on human data for pre-clinical trials, rather than animal data. Verge sequences RNA from brain samples from humans that have passed away from Alzheimer’s, Parkinson’s, or other diseases, to gather data about gene expression at late stages in a disease’s progression, and then uses algorithms to identify viable drug targets in these genes and designs new drugs accordingly.
3. Using AI to Develop Better Nuclear Reactors
Researchers at the University of Wisconsin and Oak Ridge National Laboratory have developed a machine learning system that can evaluate microscopic radiation damage to materials being considered for use in nuclear reactors. The researchers trained their system on 270 images of materials exhibiting a type of difficult-to-detect radiation damage called dislocation loops, enabling it to spot this kind of damage in new materials with 86 percent accuracy, while human experts can only identify dislocation loops with 80 percent accuracy.
4. Machine Learning Can Match Your Pose
Google has published a machine learning tool called Move Mirror capable of analyzing video of people moving and identifying images with matching poses in real-time. Move Mirror relies on a machine learning model called PoseNet that can track the relative position of 17 different body parts and a database of over 80,000 images of people in a wide variety of poses. Google developed PoseNet, which is freely available as open-source, to encourage the development of augmented reality and animation applications that take advantage of pose tracking and machine learning.
5. Making Robotic Art Competitive
An annual competition called Robotart is pitting artists and AI researchers against each other to accelerate the development of AI systems and robotics capable of creating compelling artwork. Participating teams use AI systems and industrial robotic arms to generate and execute ideas for artwork judged based on three criteria, including originality and aesthetics; technique, such as layering and blending of brushstrokes; and technical contribution to the field. While robotic systems can create artwork that is indistinguishable from human-made art from a technical perspective, it is difficult to program a system to be creative.
6. Making it Easier to Program Quantum Computers
Google has published an open-source software toolkit called Cirq designed to make it easier to develop algorithms for quantum computers. Programming for quantum computers requires highly specialized knowledge due to the differences between traditional computing, which uses standard bits that can represent either 1 or 0, and quantum computing, which relies on qubits that can be in both states at once, as well as influence other qubits that they are not touching. Cirq allows developers to develop algorithms in simulated quantum computing environments. Google has also published OpenFermion-Cirq, a specialized version of the toolkit designed for developing algorithms that simulate molecules.
7. Helping AI Developers Protect Their Intellectual Property
IBM has developed a method for embedding intellectual property protections in artificial neural networks, similar to how photographers embed watermarks in their photos to prevent unauthorized re-use. The method involves coding specific information into a neural network to make it produce a particular response when the algorithm analyzes a specific image, allowing testers to verify the ownership of the code. The virtual watermark is embedded in the neural network in such a way that makes it impossible for a bad actor to simply edit and delete the code, as well as without affecting the size or accuracy of the model.
8. Making it Easier to Transfer Your Data
Facebook, Google, Microsoft, and Twitter have launched a new initiative called the Data Transfer Project, which will allow users to easily port their data from one platform to another without needed to download and re-upload it. Users can transfer photos, calendars, mail, and other data from the companies’ platforms, including Instagram and Flickr, via the Data Transfer Project’s APIs.
9. Catching Counterfeits with AI
Though counterfeiters are beginning to use AI to make more convincing fake products, several companies are using AI to get better at spotting counterfeit goods. Companies such as Entrupy, Red Points, and Cypheme are use AI to analyze materials, colors, packaging, and other factors to identify fakes. IBM has developed an AI tool for smartphones called Crypto Anchor Verifier that allows users to take pictures of products and compares them against manufacturer-provided images of authentic products to compare subtle patterns and determine their authenticity.
10. Studying the Origins of Life
Researchers at the University of Glasgow have developed a robotic system that uses machine learning to run chemical experiments that could lead to new insights about how the first self-replicating molecules came into existence. The researchers trained their system to analyze chemical reactions and spot nucleophiles and electrophiles—chemical reagents that influence the reactivity of different molecules. The system can predict the reactivity of different chemical interactions with 86 percent accuracy, allowing researchers to simulate large amounts of chemical reactions to observe how molecules could become self-replicating.