Dr. T's AI brief
Daniel Tauritz
Deep Learning Networks Prefer the Human Voice—Just Like Us
Columbia Engineering
Holly Evarts
April 6, 2021
Columbia University's Hod Lipson and Boyuan Chen demonstrated that artificial intelligence systems programmed with sound files of human language can outperform those coded with numerical data labels. The engineers created two neural networks and trained them to recognize 10 different types of objects in a set of 50,000 photos. One system was trained with binary inputs, while the other was fed a data table containing photos of animal or objects with corresponding audio files of a human voice speaking the names of those animals or objects. The Columbia researchers found that when presented with an image, the binary-programmed network answered with 1s and 0s, while the other network vocalized the name of the imaged object. When tested with ambiguous images, the voice-trained network was found to be 50% accurate, while the numerically trained network was only 20% accurate.
Screening for Skin Disease on Your Laptop
University of Houston Cullen College of Engineering
Laurie Fickman
April 6, 2021
University of Houston (UH) researchers have developed a deep neural network architecture that facilitates early diagnosis of systemic sclerosis (SSc) by immediately differentiating between images of healthy and diseased skin. The network was trained using the parameters of the MobileNetV2 mobile vision application, pretrained on the 1.4-million-image ImageNet dataset. The UH team added layers to the UNet, a modified convolutional neural network (CNN) architecture, then devised a mobile training module. Results indicated the proposed architecture outperformed CNNs for SSc image classification. UH's Yasmin Akay said, "After fine-tuning, our results showed the proposed network reached 100% accuracy on the training image set, 96.8% accuracy on the validation image set, and 95.2% on the testing image set."
AI Tool Can Help Detect Melanoma
MIT News
Megan Lewis
April 2, 2021
Researchers at the Massachusetts Institute of Technology (MIT) have designed an artificial intelligence system that analyzes wide-field images of patients' skin in order to detect melanoma more efficiently. The process applies deep convolutional neural networks (DCNNs) to optimize the identification and classification of suspicious pigmented lesions (SPLs) in wide-field images. The MIT researchers trained the system on 20,388 wide-field images from 133 patients at Spain's Hospital Gregorio Marañón, and on publicly available images. Dermatologists visually classified lesions in the images for comparison, and the system achieved more than 90.3% sensitivity in differentiating SPLs from nonsuspicious lesions, skin, and complex backgrounds.
ML Approach Speeds Up Search for Molecular Conformers
Aalto University (Finland)
April 1, 2021
Researchers at Finland's Aalto University developed a molecular conformer search procedure that integrates an active learning Bayesian optimization algorithm with quantum chemistry techniques to accelerate the process. Searching for molecular conformers previously required the relaxation of thousands of structures, entailing a significant commitment of time and computational resources even when applied to small molecules. The Aalto team's algorithm samples the structures with low energies or high energy uncertainties, to minimize the required data points. The researchers tested the machine learning procedure on four amino acids, and found low-energy conformers in good correspondence with experimental measurements and reference calculations while using less than 10% of the computational cost of the current fastest method.
Researchers Develop 'Explainable' Algorithm
University of Toronto (Canada)
Matthew Tierney
March 31, 2021
An "explainable" artificial intelligence (XAI) algorithm developed by researchers at Canada's University of Toronto (U of T) and LG AI Research was designed to find and fix defects in display screens. XAI addresses issues with the "black box" approach of machine learning strategies, in which the artificial intelligence makes decisions entirely on its own. With XAI's "glass box" approach, XAI algorithms are run simultaneously with traditional algorithms to audit the validity and level of their learning performance, perform debugging, and identify training efficiencies. U of T's Mahesh Sudhakar said LG "had an existing [machine learning] model that identified defective parts in LG products with displays, and our task was to improve the accuracy of the high-resolution heat maps of possible defects while maintaining an acceptable run time." The new XAI algorithm, Semantic Input Sampling for Explanation (SISE), outperformed comparable approaches on industry benchmarks.
AI Tool 85% Accurate at Recognizing, Classifying Wind Turbine Blade Defects |
AI-Based Tool Detects Bipolar Disorder at Earlier Stages
Folio (University of Alberta, Canada)
March 30, 2021
A machine learning (ML) model developed by researchers at Canada's University of Alberta (UAlberta) and Chinese colleagues can help to identify subtle cognitive deficits that signify early-stage or first-episode bipolar disorder. The team trained the model by comparing patients with chronic bipolar disorder to healthy controls, then showed that the model could differentiate first-episode bipolar disorder patients from controls with 76% accuracy. The researchers think a cognitive test that uses ML analysis is a far less expensive and time-consuming technique for diagnosing bipolar disorder than brain imaging, and it can also monitor symptoms over time.
AI Method for Generating Proteins Will Speed Up Drug Development
Chalmers University of Technology (Sweden)
March 30, 2021
Researchers at Sweden's Chalmers University of Technology have developed artificial intelligence (AI) that can synthesize novel, functionally active proteins. Chalmers' Aleksej Zelezniak said the method can proceed from design to working protein in just a few weeks, much more quickly than current protein-engineering techniques. The ProteinGAN approach involves feeding the AI a large dataset of well-studied proteins, which it analyzes and attempts to generate new proteins; concurrently, another part of the AI tries to determine if the synthetic proteins are natural or not. Said Chalmers' Martin Engqvist, "Accelerating the rate at which we engineer proteins is very important for driving down development costs for enzyme catalysts. This is the key for realizing environmentally sustainable industrial processes and consumer products, and our AI model, as well as future models, will enable that."
MIT Study Finds 'Systematic' Labeling Errors in Popular AI Benchmark Datasets
VentureBeat
Kyle Wiggers
March 28, 2021
An analysis by Massachusetts Institute of Technology (MIT) researchers demonstrated the susceptibility of popular open source artificial intelligence benchmark datasets to labeling errors. The team investigated 10 test sets from datasets, including the ImageNet database, to find an average of 3.4% errors across all datasets. The MIT investigators calculated that the Google-maintained QuickDraw database of 50 million drawings had the most errors in its test set, at 10.12% of all labels. The researchers said these mislabelings make the benchmark results from the test sets unstable. The authors wrote, "Traditionally, machine learning practitioners choose which model to deploy based on test accuracy—our findings advise caution here, proposing that judging models over correctly labeled test sets may be more useful, especially for noisy real-world datasets."
Deep Learning Tool Predicts Hosts Based on Early SARS-CoV-2 Samples
News-Medical Life Sciences
Angela Betsaida B. Laguipo
March 24, 2021
Researchers at China's Peking University designed a deep learning tool that can predict hosts for new viruses, including SARS-CoV-2. DeepHoF (Deep learning-based Host Finder) forecasts host probability scores for five host types: germ, plant, invertebrate, non-human vertebrate, and humans. DeepHoF was based on BiPath Convolutional Neural Network, and can automatically infer genomic features from inputted viral sequences. The Peking researchers conducted a deep analysis of the host likelihood profile calculated by DeepHoF, using the earliest samples of SARS-CoV-2 isolates to determine that minks, dogs, bats, and cats were potential hosts—with minks particularly significant. The new model also can help predict potential hosts of viruses that may induce another outbreak or pandemic.
UCLA Researchers Develop Noninvasive AI Method to Inspect Live Cells, Gain Critical Data
UCLA Samueli School of Engineering
March 24, 2021
At the University of California, Los Angeles (UCLA) Samueli School of Engineering, researchers have developed a noninvasive artificial intelligence (AI)-based technique to analyze live biological cells. Using a deep learning model, the cells are viewed and a snapshot captured under a brightfield microscope. The UCLA team trained the model to deduce and identify antibody-labeled fluorescent cellular images, and to note subtle distinctions in size and shape in order to predict protein levels. This analysis yields data on protein concentrations and location without destroying the sample, and the AI tool can predict as many proteins as the model has been trained to identify. UCLA's Neil Lin said this method could be of use to academic and industrial cell biology laboratories, and "it could be especially important in cell therapies, where the cells themselves are valuable."
Navy Official Says AI Critical To Digital Progress
Inside Defense 
(3/23) reports Navy deputy assistant secretary for research development, test, and engineering Joan Johnson said Tuesday that the Navy, along with other services, are establishing digital readiness goals to compete in distributed maritime operations with countries such as China. Johnson said AI is a “critical component” to moving the service’s engineering domain from being “digitally aware” to “digital practitioners.”
Synthetic Data “Could Save AI”
VentureBeat (3/20, Grossman) reports that AI “is facing several critical challenges,” including the need for “huge amounts of data” that isn’t biased. According to the piece, “We’re seeing a new industry emerge that promises to be a saving grace: synthetic data,” referring to “data is artificial computer-generated data that can stand-in for data obtained from the real world.”
Energy Industry Turns To AI To Deal With Data
The Houston Chronicle (3/19, Magill) reported the energy industry is turning to technology to help energy companies “tap oil and gas wells more efficiently, pipeline companies identify leaks, and power grids integrate wind and solar energy with batteries and traditional generators.” As they collect increasing amounts of data, companies are using AI to make sense of all that “data collected from exploration activities, drilling activities and other operations that deliver energy to the global economy.” Shell has been a leader in the adoption of AI, using it to help a variety of projects including “its Quest carbon capture and storage facility in Alberta...to ensure that the carbon dioxide the plant injects into the ground stays in the ground.” Shell GM of Data Science Dan Jeavons said, “Digital in general and AI in particular are very much front and center of Shell’s thinking.”