Large Language Models at Google, Twitter forgoes image-cropping algorithms, and more!

Last Week in AI #117

May 24, 2021

Mini Briefs

Google's plan to make search more sentient

At last week's Google I/O, the company's annual developer conference that announces new products and initiatives, Google unveiled two new applications of large language models that could significantly impact search and dialogue. One technology is called Multitask Unified Model (MUM), which improves search by better taking into account the context of a search query. In a demo, Google showed how MUM can answer a query that pairs a question like "can I use these to hike Mt. Fuji" with a picture of a hiking boots. The other technology is called LaMDA, which is a large language model specifically trained for open-ended dialogue. In its demo, LaMDA conversed as if it were the dwarf planet Pluto and a paper airplane, answering questions like "what's it like being thrown in the air?"

Both technologies are under development and Google did not specify when they'd be integrated into their products. Deployment of large language models in real products have many potentials for harm and has been the center of many controversies this last year. We point readers to these articles for further reading:

Sharing learnings about our image cropping algorithm

In October last year Twitter users raised concerns with the company's seemingly biased image cropping algorithm. When a user uploads an image that is too big to display on the Twitter feed, a machine learning algorithm was used to automatically choose an area of the image to crop. Some examples showed that for images containing faces of both black and white people, the algorithm tended to crop and show the white face more often. Responding to these concerns, Twitter conducted an internal study and published its results in this blog post.

Among other findings, Twitter's internal experiments showed that when two faces are present in an image, the cropping algorithm favored choosing women 8% more than men and white faces 4% more than black. Due to these discreprencies, the company concludes that "not everything on Twitter is a good candidate for an algorithm, and in this case, how to crop an image is a decision best made by people." Moving forward, Twitter will allow users to manually crop images for display, instead of relying on the algorithm.

Podcast

Check out our weekly podcast covering these stories!
Website | RSS | iTunes | Spotify | YouTube

News

Advances & Business

The Surging Opportunity for Robotic Automation - "Ken Goldberg and The House Fund discuss the opportunities and challenges in state-of-the-art robotics for Manufacturing and Logistics."
Latest data shows Apple has increased self-driving test cars but decreased number of drivers - "Apple Car rumors have been slow for the past couple of months. Today, a story from MacReports says the company has increased the number of self-driving cars but decreased by almost half the number of drivers licensed to drive these cars."
IBM's Project CodeNet will test how far you can push AI to write software - "IBM's AI research division has released a 14-million-sample dataset to develop machine learning models that can help in programming tasks."
GPT-3's free alternative GPT-Neo is something to be excited about - "Elevate your enterprise data technology and strategy at Transform 2021. The advent of Transformers in 2017 completely changed the world of neural networks. Ever since, the core concept of Transformers has been remixed, repackaged, and rebundled in several models."
Facebook Wav2vec-U learns to recognize speech from unlabeled data - "Elevate your enterprise data technology and strategy at Transform 2021. Facebook today announced that it trained an AI model to build speech recognition systems that don't require transcribed data."
These Ex-Journalists Are Using AI to Catch Online Defamation - "Like many stories about people trying to help fix the internet, this one begins in the aftermath of 2016.
Deepfake dubs could help translate film and TV without losing an actor's original performance - "What exactly is lost in translation when TV shows and films are subbed or dubbed into a new language? It's a hard question to answer, but for the team at AI startup Flawless, it may be one we don't have to think about in the future."
Despite challenges, Salesforce says chatbot adoption is accelerating - "Elevate your enterprise data technology and strategy at Transform 2021. Chatbot usage has exploded during the pandemic as organizations look to bridge emerging gaps in customer service and onboarding. In 2020, the chatbot market was valued at $17.17 billion and is projected to reach $102."
Kodiak Robotics Expands Autonomous Truck Operations in Texas - "Autonomous truck developer Kodiak Robotics this summer is expanding its routes, adding a Dallas-to-San Antonio run."
NIST Proposes Method for Evaluating User Trust in Artificial Intelligence Systems - "NIST's new publication proposes a list of nine factors that contribute to a human's potential trust in an AI system. A person may weigh the nine factors differently depending on both the task itself and the risk involved in trusting the AI's decision."
Quizzing Michelangelo AI - "Testing the Museo dell-Opera del Duomo app"
Google Unit DeepMind Tried and Failed to Win AI Autonomy From Parent - "Alphabet cuts off yearslong push by founders of the artificial-intelligence company to secure more independence"
Google details new AI accelerator chips - "Elevate your enterprise data technology and strategy at Transform 2021. At Google I/O 2021, Google today formally announced its fourth-generation tensor processing units (TPUs), which the company claims can complete AI and machine learning training workloads in close-to-record wall clock time."

Concerns & Hype

Confused Waymo robotaxi shows challenges of scaling AVs - "The Waymo robotaxi needs to make a right turn onto a multi-lane main road, but the far right lane was closed off by orange construction cones. And, boy, did they confuse the vehicle."
DMV probing whether Tesla violates state regulations with self-driving claims - "Tesla is "under review" by the California Department of Motor Vehicles to determine whether the company misleads customers by advertising its "full self-driving capability" option, the agency told The Times."
The Chief AI Ethics Officer: A Champion or a PR Stunt? - "We have reached a point where the far-reaching impacts of AI's ability to identify, prioritize and predict can be felt in virtually every industry."
Ways To Stop AI From Recognizing Your Face In Selfies - "We upload so many personal photos on the internet, so we might have questions like who else would have access to them, what would they do with them - and which machine-learning algorithms would be trained with this data?"
60 Minutes ran a segment on racial bias in facial recognition - it left out the Black women who pioneered the research - "Sunday night's episode delved into the cases of two Black men in Detroit, Michigan who were arrested after facial recognition falsely identified them as suspects in criminal cases."
Google's New Dermatology App Wasn't Designed for People With Darker Skin - "The company trained the system to recognize different skin conditions. But like Google itself, the app's data has a diversity problem."

Analysis & Policy

We could see federal regulation on face recognition as early as next week - "Amazon and others are indefinitely suspending police use of face recognition products, but proposed legislation could make bans bigger or more permanent."
Lawmakers want to force Big Tech to give researchers more data - "Facebook's ad library allows researchers to see the content of ads that run on the platform and information on who those ads reach. But there is one key insight Facebook doesn't offer: information on how those ads were targeted."

That's all for this week! If you are not subscribed and liked this, feel free to subscribe below!

Last Week in AI