Discover more from Last Week in AI

Weekly text and audio summaries of the most interesting AI news, as well as editorials commenting on recent events

Over 23,000 subscribers

News

Last Week in AI #162: OpenAI DALL-E 2 generates amazing images, Google's 540 billion parameters language model, Clearview AI branches out beyond police, and more!

OpenAI's DALL-E 2 draws convincing pictures from detailed prompts, Google's new language model shows that scaling up has yet to reach its full potential, Clearview AI to offer "consent-based" products

Jacky Liang

Apr 11, 2022

OpenAI's new DALL.E model turns your words into pieces of art

In January 2021, OpenAI introduced DALL·E, a neural-net based method for generating images from textual descriptions. Now it has announced DALL·E 2, a model which uses a new algorithmic approach to create more realistic and accurate images with 4x greater resolution. The research paper states that images generated by a 3.5 billion parameter version of the model trained a certain way created images favored by human evaluators over those of DALL·E, even when the latter uses extra computation. DALL·E 2 can also make realistic edits to existing images from a natural language caption and create different variations of it.

Pathways Language Model (PaLM): Scaling to 540 Billion Parameters for Breakthrough Performance

Google announced the new Pathways Language Model (PaLM) that achieves significant performance improvements in many language tasks over the state-of-the-art. The Pathways system can train large language models in a highly efficient and distributed way, which allowed the model to scale up and perform well in many challenging tasks, like explaining jokes, identifying causes and effects, and fixing bugs in code snippets. PaLM actually achieves higher scores than the average human performance on many of these tasks. Interestingly, in Google’s experiments, PaLM’s performance as a function of model parameters does not seem to have plateaued, suggesting that further scaling up model sizes can yield additional benefits.

Our take: Improvements in large language models do not seem to be slowing down. The first breakout success of these models, GPT-2, was only released 3 years ago, and since then we’ve seen drastic improvements in performance, scale, and applications. While training these large models still remains inaccessible to individuals or smaller institutions (estimates put the minimum cost of training PaLM at around $9M), their usage, both academic and commercial, will only become more and more prevalent.

Face scanner Clearview AI aims to branch out beyond police

After a BuzzFeed News expose found that facial recognition provider Clearview AI had offered its services to over 2200 customers including Macy's, Walmart, and the NBA, the startup received intense backlash. While Clearview later said it would only offer its technology to law enforcement, it now seems interested in branching out again. However, new customers would not have access to its treasure trove of 20 billion images. The new "consent-based" product, with aspirations of competing with Amazon and Microsoft, would use Clearview's algorithms to verify a person's face for bank transactions or commercial purposes. Clearview has faced plenty of lawsuits in recent months but continues to serve customers, and recently offered its services to the Ukrainian military for free.

Our Take: Given Clearview's reputation, it feels easy to look for fault when they enter the picture. It is probably difficult to separate the question of Clearview's new product from the privacy concerns surrounding their dataset since Clearview's algorithms were trained using those data. I have many questions about their intentions, given Ton-That's business history and the fact that they sold services to Macy's and the NBA at all. Furthermore, the ABC article notes a December financial presentation that proposed uses for Clearview technology such as monitoring gig economy workers. I think we need to continue to carefully monitor how facial recognition systems are used and deployed and prioritize rights and human well-being as we make judgments about the use of those systems.

Other News

Research

The 11 Commandments of Hugging Robots - "How to build a humanoid that gives perfect hugs"

CMU & Google Extend Pretrained Models to Thousands of Underrepresented Languages Without Using Monolingual Data - "most world languages do not fit these criteria, as they present relatively little or even no textual data."

Stanford Researchers Have Developed a Machine Learning-Based Algorithm To Detect Autism in Brain “Fingerprints” - "Patients with autism who are diagnosed early and definitively may benefit from earlier therapies and better results. Stanford researchers have created an algorithm that can tell if someone has autism by analyzing brain images."

AI Betting Agents Weigh In On Reproducibility Of Research Findings - "Penn State researchers have created synthetic “prediction markets” for speculating on the replicability of published studies with unparalleled accuracy."

Baidu Proposes PP-YOLOE: An Evolved Version of YOLO that Achieves SOTA Performance in Object Detection - "Object detection is a fundamental task in computer vision, and YOLO (You Only Look Once) one-stage object detectors have set the performance standard since the YOLOv1 debut in 2015."

Creepy AI reconstructs your portrait based only on your voice - "Turning speech into text has become so common that i’s a part of almost every smartphone. But have you ever thought about turning your speech into a portrait? Researchers have, and they’ve even made it possible."

Solving the challenges of robotic pizza-making - "A new technique could enable a robot to manipulate squishy objects like pizza dough or soft materials like clothing."

Does this artificial intelligence think like a human? - "In machine learning, understanding why a model makes certain decisions is often just as important as whether those decisions are correct."

Stanford Researchers Propose 'Time Control (TC)' : A Language Model that Uses Stochastic Processes to Enhance the Efficiency and Coherence of Long Text Generation - "Human assessors scored the outputs 28.6 percent higher than baseline approaches, indicating that the unique strategy enhances performance on long text production."

AI turns infrared images taken in total darkness into full colour - "The black-and-white images provided by night-vision cameras can be colourised using AI, but it must always be trained on similar images and is unlikely to ever work on unfamiliar general scenes"

Johns Hopkins’ heart-scanning AI predicts cardiac arrests up to 10 years ahead - "Researchers at Johns Hopkins University have developed an artificial intelligence approach they say can help predict if and when a person could die of cardiac arrest based on imaging scans of the heart."

Research combines artificial intelligence and computational science for accurate and efficient simulations of complex systems - "Predicting how climate and the environment will change over time or how air flows over an aircraft are too complex even for the most powerful supercomputers to solve. Scientists rely on models to fill in the gap between what they can simulate and what they need to predict."

New machine learning models can predict adverse outcomes after abdominal hernia surgery with high accuracy - "Procedure-specific risk calculator has potential to encourage changes in patient behaviors before surgery to improve success in abdominal wall reconstruction operations"

Applications

Microsoft and HPE put AI to the test on International Space Station … with gloves - "An image of a spacewalker’s gloves is analyzed for signs of wear using AI"

Adding AI to Museum Exhibits Increases Learning, Keeps Kids Engaged Longer - "They used artificial intelligence (AI) to create a new genre of interactive, hands-on exhibits that includes a virtual assistant to interact with visitors."

AI is explaining itself to humans. And it's paying off - "LinkedIn boosted subscription revenue by 8% after arming its sales team with artificial intelligence software that not only predicts clients at risk of canceling, but also explains how it arrived at its conclusion."

Meet ‘CodeGen’: An AI Model That Turns Simple Natural Language Requests Into Executable Code - "Imagine being able to tell a machine to write an app simply by telling it what the app does. As far-fetched as it may appear, this scenario is already a reality."

Meta AI’s open-source system attempts to right gender bias in Wikipedia biographies - "AI models don’t cover everyone in the world equally"

Can A.I.-Driven Voice Analysis Help Identify Mental Disorders? - "Early tests have been promising, but issues involving bias, privacy and mistrust of “black box” algorithms are possible pitfalls."

AI-powered camera traps could protect Gabon wildlife from poachers - "AI-powered camera traps give Gabon wildlife rangers a new tool in the fight against poaching and biodiversity loss. "

First autonomous X-ray-analyzing AI is cleared in the EU - "An artificial intelligence tool that reads chest X-rays without oversight from a radiologist got regulatory clearance in the European Union last week — a first for a fully autonomous medical imaging AI, the company, called Oxipit, said in a statement."

What AI Can Do for Climate Change, and What Climate Change Can Do for AI - "To tackle the climate crisis, artificial intelligence is becoming more open and democratic"

9 Out of 10 Depressed Twitter Users Spotted by AI Bot - "A newly developed algorithm can spot depression in Twitter users with 88.39 percent accuracy."

Machine Learning Model Measures MLB Players’ Performances - "A team of researchers at the Penn State College of Information Sciences and Technology has developed a machine learning model that can better measure baseball players’ and teams’ short- and long-term performance."

Using machine learning to improve student success in higher education - "Deploying machine learning and advanced analytics thoughtfully and to their full potential may support improvements in student access, success, and the overall student experience."

Business

Built Robotics raises another $64M to make construction equipment autonomous - "If you’re still wrapping your head around the idea of self-driving cars, get ready for something that might seem even wilder: self-driving construction equipment exists, and it’s already out there digging away on job sites."

Health-Tech Startup Viz.ai Raises $100 Million - "The funding round led by Tiger Global Management and Insight Partners raises the six-year-old company’s valuation to $1.2 billion"

This California-based HealthTech Startup is Using Artificial Intelligence To Modernize and Personalize Fertility Care - "Alife Health, a San Francisco-based startup is using artificial intelligence to improve IVF fertility outcomes."

Werner to use self-driving trucks on 600 mile lane deemed ‘unappealing’ for human truck drivers - "Werner Enterprises in partnership with autonomous vehicle tech company Aurora Innovation, Inc. launched a commercial pilot program in Texas."

Elon Musk says production of Tesla's robot could start next year, but A.I. experts have their doubts - "Tesla has yet to reveal a working prototype of the robot, however, and it’s unclear how sophisticated Optimus is at this stage."

Lilt raises $55M to bolster its business-focused AI translation platform - "Lilt, a provider of AI-powered business translation software, today announced that it raised $55 million in a Series C round led by Four Rivers, joined by new investors Sorenson Capital, CLEAR Ventures and Wipro Ventures."

Levi’s Chief AI Officer Talks Raising the Bar (in Everything) with Machine Learning - "Levi’s is automating as much as it can when it comes to pricing and promotions, which Walsh says has enabled them to avoid deep discounting during store closures thanks to data showing consumers were willing to pay the full price based on brand recognition."

Salesforce upgrades marketing, field service, and sales tools with AI - "Today, timed with a conference out of Sydney, Australia, Salesforce announced new analytics-powered features across its Service and Marketing Cloud suites designed to “‘humanize’ engagement between companies and their customers."

New AMD CPU patent reveals 3D-stacked machine learning accelerator design - "AMD is beginning to focus on 3D-stacked machine learning accelerators in the newest patent innovations"

Concerns

AI algorithms could disrupt our ability to think - "In other words, we could already be in the process of outsourcing our thinking to machines and, as a result, losing a portion of our agency. "

Google Says AI Generated Content Is Against Guidelines - "Google’s Search Advocate John Mueller says content automatically generated with AI writing tools is considered spam, according to the search engine’s webmaster guidelines."

As Russia Plots Its Next Move, an AI Listens to the Chatter - "With vast amounts of data becoming available to intelligence analysts, new tools will help them sift and interpret it all—but they will introduce new risks, too."

Analysis

Toward a Broad AI - "Deep learning models were shown to fail at new data, new applications, deployments in the wild, and stress tests"

AI Is Improving Its Artistic Skills, But Who Owns Its Output? - "This post dives into AI's latest attempts to create art, then examines who owns the output."

How AI and Humans Can Best Collaborate at Work - "Who decides who does what? And how can humans learn to trust AI? Research offers some answers."

Policy

California suggests taking aim at AI-powered hiring software - "A newly proposed amendment to California's hiring discrimination laws would make AI-powered employment decision-making software a source of legal liability."

Last Week in AI