Microsoft's latest image-captioning AI can describe photos better than humans. Microsoft said it's used the algorithm to update its Seeing AI assistant app for the visually impaired and will later incorporate it into Word, PowerPoint, and Outlook to create alt-text and other tasks.

More:

• The AI automatically added captions to images more accurately than humans in certain limited tests. It achieved the highest scores on the image-captioning benchmark known as “nocaps," an industry-leading scoreboard.
• The benchmark achievement could lead to improvements in “alt text” for online images, which screen readers use to describe images to blind and sight-impaired individuals. This could help them navigate software and the internet better.
• Microsoft said the AI is 2x better than the image captioning model it’s had since 2015. For example, its previous image captioning system described a cat image as “a close up of a cat," while the new iteration labeled it as “a gray cat with its eyes closed.”
• Microsoft called it a "milestone" in its efforts to make its services and products "inclusive and accessible to all users.” It's now offering the updated captioning model to developers via Azure's Cognitive Services.
• Eric Boyd, CVP of Azure AI, called image captioning "one of the hardest problems in AI,” noting that the AI has to understand objects in a scene as well as how they interact and how to describe them.
• Four years ago, Google said its AI could caption images at 94% accuracy, nearly as well as humans.

Google will unveil new AI-powered improvements to its search engine during its "Search On" event tomorrow, Oct. 15. Google says attendees will hear the ways it's "harnessing the power of AI to help people understand the world around them.”

More:

• Google didn't reveal details about what the AI improvements will be. Reports speculate it will also announce new updates for Google Lens, Images, Assistant, and Maps.
• Google normally unveils new search features during its annual I/O developer event, which was canceled this year due to the pandemic. The "Search On" event kicks off at 3 p.m. ET tomorrow on YouTube and the event page on Google's website.
• Google also recently said its Pixel 5 and Pixel 4a phones will have "Hold for Me," an AI-powered service that will monitor call center phone calls that are on hold and send a notification to users when a representative picks up.
• A new update to the Google Photos app has two new AI-based photo editing tools. The first is one-tap edits, which use machine learning to provide suggestions to specific photos. Portrait Light, which is on the new Pixel 5 and Pixel 4a, allows users to change the lighting of their photos after they're captured (see gif above).
• A machine learning algorithm experiment known as Google Tone Transfer allows users to record music in their heads, which an algorithm converts into a digital signal to create actual instrumental music.

Report: How the U.S. can advance AI

The National Security Commission on Artificial Intelligence has released its latest report and third-quarter recommendations on how the U.S. can stay ahead of China and other nations in AI.

The 66 recommendations, which were sent to Congress and President Trump, run the gamut from how to train and recruit AI talent to strengthening international AI partnerships. The commission — chaired by former Google CEO Eric Schmidt and former defense deputy secretary Robert Work — stressed that the U.S. is in an unprecedented "innovation competition" and needs bolder and more aggressive actions to develop and incorporate AI technologies.

We've highlighted some of the major recommendations below:

• Strengthen the "triangular alliance" for AI R&D — referring to academia, government, and industry — through new funding mechanisms, expanding the role of industry in the defense department's AI R&D for next-generation capabilities, and establishing a modern digital ecosystem that would help address national security issues.
• Apply AI to national security missions by leveraging the defense department's AI-enabled analytics and integrating its "technology scouting"...

Facebook and Carnegie Mellon University partnered on an AI-powered project to find new “electrocatalysts," which help to store energy generated by solar and wind power for future use. The Open Catalyst Project could lead to cheaper and more efficient energy storage, resulting in higher use of renewable energy.

More:

• Electrocatalysts are typically expensive and are used to convert excess solar and wind power into other fuels, such as ethanol or hydrogen, making them easier to store.
• The Open Catalyst Project would use AI to more quickly predict atomic interactions and discover new catalysts, possibly in seconds rather than days. They hope to uncover less expensive catalysts.
• Facebook and Carnegie Mellon believe they could test millions or even billions of potential catalysts per year using AI, rather than the tens of thousands tested now.
• The two also released AI software models and a dataset for other scientists to join in the effort.
• The project is a chance to "advance AI in a way that will have a significant real-world impact" as energy needs grow and climate change efforts become more urgent, said Facebook's Research Scientist Larry Zitnick.

An AI-based analysis by Expert.ai predicts that the U.S. presidential race may be closer than many polls suggest. The "sentiment analysis" placed former Vice President Joe Biden ahead of President Trump by 50.2% to 47.3%, a margin of about 3 percentage points.

More:

• Maryland-based Expert.ai used an AI model to analyze ~500k Twitter posts about the candidates. The natural language processing system gauged the amount of positive or negative emotions expressed in posts. The AI then categorized them into 84 different emotional labels.
• Trump accounted for nearly 60% of all the posts analyzed. Biden ranked higher in positive emotions such as "success" and "hope," while Trump scored higher on negative emotions such as "fear" and "hatred."
• The intensity of emotions can play a role in decision-making at the polls, Expert.ai CEO Walt Mayo said. Analyzing social media posts in real-time can provide insights into what voters are actually discussing and what their attitudes are, he added.
• In 2016, Expert.ai used its political forecasting model to predict that the U.K. would vote in favor of Brexit.
• A 2019 analysis found that Twitter users tend to be younger and more likely Democrat than the average U.S. adult.

QUICK HITS

• Facebook AI Research released Opacus, a library for applying differential privacy techniques for training deep-learning models using PyTorch.
• Navina, an AI-based platform for primary care, raised $7M in a funding round led by Grove Ventures.
• As Apple device adoption grows in the enterprise, organizations are looking for ways to seamlessly manage and secure devices while extending the world-class Apple experience users expect.*

*This is sponsored content.

Tweet of the Day: Intel India and CBSD achieved a world record for the most users to take an online AI lesson in 24 hours.