Welcome to Monday's Inside AI! In today's issue:
- An AI model that can help doctors determine how far a patient's Parkinson's disease has progressed.
- The Las Vegas Metropolitan Police Department fed a facial recognition system with “non-suitable” suspect images in nearly half of its searches last year.
- The latest funding rounds for AI/ML startups (premium only).
Premium subscribers have access to all content, ad-free. We offer a $10/month plan or the option to be billed $100 annually, and are currently offering a 14-day free trial! Click here to sign up!
Thanks for your support,
|
|
|
The Michael J. Fox Foundation and IBM Research developed an AI model that can help doctors determine how far a patient's Parkinson's disease has progressed. The progression model stems from a collaboration that the organizations started last year, which aims to use machine learning to find treatments and, one day, a cure for the neurodegenerative disorder.
More:
- The modeling method helps clinicians understand how the disease progresses based on a patient's symptoms, which is usually difficult to assess quantitatively.
- The algorithms consider personalized factors, such as the impact of medications, that can affect how Parkinson's disease manifests in each person.
- IBM wants to train the model further using patient data from the Michael J. Fox Foundation, which could lead to more clinical insights. They hope it could be used for patient care management, cohort generation, and clinical trial outcome modeling.
- The organizations released a related paper showing statistical representations of patients' movements and detailed the model at last week's 2020 Machine Learning for Healthcare Conference.
- From Twitter: @RosaLSmothers wrote the research has "fascinating implications for progressive disease treatments."
ZDNET
|
|
The Las Vegas Metropolitan Police Department fed a facial recognition system with “non-suitable” or grainy images in nearly half of its searches last year, some of which led to arrests. Using poor "probe" images in facial recognition is directly correlated with more false identifications of suspects, and false arrests. In Las Vegas, “non-suitable” probe images fed through facial recognition returned positive suspect matches less than 20% of the time.
More:
- Many probe images used in searches are from CCTV footage, which tends to be grainy with subjects looking away from the camera. These "non-suitable” probe images, though common, return fewer and less accurate matches, which could lead to more false arrests, the ACLU says.
- LVMPD did 924 facial recognition searches last year using Vigilant Solutions' facial recognition technology. It used these non-suitable images to identify suspects in at least 73 cases.
- The department says it uses facial recognition only as a lead and not as an outright identification. Vigilant's facial system allows officers to edit and filter photos to improve the photo, and any matches are investigated further by the department, it said.
- Police should not even run an FR search if they have a non-suitable probe photo, argued Clare Garvie, a Georgetown Law Center on Privacy and Technology senior associate who specializes in facial recognition.
Related:
- In May, the ACLU filed a complaint against Detroit police for wrongfully arresting Robert Julian-Borchak Williams based on a faulty facial recognition match, the first known case in the U.S. Williams, who is Black, was detained for a day in January after the face recognition service falsely matched him to a shoplifter.
- A month later, Inside reported that the facial recognition used by Detroit police had falsely linked a second man to a crime he didn't commit. Michael Oliver, who also is Black, was improperly flagged and later charged as a suspect in a phone-damaging incident, despite no resemblance to the perpetrator.
- The false matches renewed concerns about racism in AI facial recognition systems, which are shown to perform noticeably worse on nonwhite faces, a NIST study shows.
MOTHERBOARD BY VICE
|
|
New funding for AI/ML startups:
- Syntiant (Irvine, CA), edge AI voice and sensor solutions: $35M C led by M12, Applied Ventures, participation from Atlantic Bridge Capital, Alpha Edison, Miramar Digital Ventures.
- Voiceitt (Ramat Gan, Israel), commercial speech recognition platform: $10M A from Viking Maccabee Ventures, M12, AMIT Technion, Cahn Capital Corp, Connecticut Innovations and AARP along with Quake Capital, SLJ Family Office, Dreamit Ventures, The Disability Opportunity Fund...
|
|
FedEx installed a robotic system to sort through packages at its Memphis facility. The automated system, which is faster than human sorters, is made up of robots from Japanese manufacturer Yaskawa and software from Plus One Robotics. FedEx hopes to process up to 80% of its packages through automated systems by 2021.
More:
- The robotics arm system sorts and picks up packages and envelopes to place them on a conveyor belt.
- FedEx said it's helped speed up package processing during the pandemic, when more people are buying goods online.
- Human workers who once did the sorting supervise the system. One human supervisor manages eight of the bots.
- A Wall Street Journal video shows the system in action.
ROBOTICS & AUTOMATION
|
|
A new study suggests that benchmarking metrics currently used to gauge many AI/ML models are inadequate. Benchmarks don't often sufficiently reflect a model's true performances, and some, like accuracy, can stress certain performance characteristics while ignoring others, according to the research from the Institute for Artificial Intelligence and Decision Support.
More:
- The researchers analyzed 32,209 benchmark results from 2,298 data sets taken from 3,867 papers published over the last 20 years.
- Accuracy was the top metric found in 38% of the benchmark data sets, followed by precision and F-measure.
- 77% of the benchmark data sets had only one performance metric; 14.4% reported two top-level metrics. None used superior alternatives such as the Fowlkes-Mallows index and Matthews correlation coefficient.
- The researchers pointed out inconsistencies in the way metrics are reported and found that many benchmarks, such as accuracy, tend to be problematic and can cause ambiguities in research.
VENTUREBEAT
|
|
Chinese autonomous vehicle startup Xpeng Motors filed for an IPO on the New York Stock Exchange. The automaker will sell 429,846,136 Class B ordinary shares and an undisclosed number of Class A ordinary shares. It plans to raise $100m through the filing, although that figure may change.
More:
- According to the company’s IPO prospectus, Xpeng considers potential misconduct by its employees as a risk factor. In March 2019, Tesla sued Guangzhi Cao, a former Autopilot engineer, who reportedly stole source code for Xpeng.
- Other risk factors in Xpeng’s prospectus: R&D efforts may not yield expected results, the company admits it faces challenges as a new entrant in the EV space and demand for EVs may be cyclical, among others.
- AJPMorgan, Bank of America and Credit Suisse are serving as underwriters for the IPO.
- The automaker raised $500m in a Series C+ round of funding, which includes investors like Sequoia Capital China, Hillhouse and Aspex, among others. The company said it would use the funds to expand the development of “intelligent vehicle technologies,” as the automaker has touted its autonomous vehicle technology.
A version of this story first appeared in Inside Electric Vehicles.
CNBC
|
|
QUICK HITS
*This is a sponsored post.
|
|
Tweet of the Day: IBM CEO Arvind Krishna pays homage to the late Frances Allen, a renowned computer scientist and the first woman to win a Turing Award.
|
|
|
|
Beth Duckett is a former news and investigative reporter for The Arizona Republic, who has written for USA Today, American Art Collector, and other publications. A graduate of the Walter Cronkite School of Journalism, she won a First Amendment Award and a Pulitzer Prize nomination for her original reporting on problems within Arizona's pension systems.
|
|
Editor
|
Sheena Vasani is a journalist and UC Berkeley, Dev Bootcamp, and Thinkful alumna who writes Inside Dev and Inside NoCode.
|
|