Discover more from Last Week in AI
Last Week in AI #138: Singapore trials surveillance robots, Deepfake voice used to rob a bank, Google studies the limits of large-scale pretraining
Singapore's patrol robots send live video feeds to the police, Google researchers study nuances of scaling large neural network models, Deepfake voice over the phone used to steal money, and more!
A new robot named Xavier has been developed to patrol the streets of Singapore. It is outfitted with sensors for autonomous navigation, and has a 360 degree camera feed to record what’s happening near it. Officers can receive the video feed and are able to monitor and control multiple robots simultaneously. Its aim is to detect undesirable behavior, such as smoking in prohibited areas, illegal hawking, improperly parked bicycles, and more. In response to such situations it can trigger real-time alerts and display appropriate messages deter such behaviours.
With advances from models like GPT-3 and the even more recent 530B-parameter Megatron model from Microsoft/NVIDIA, it's natural to wonder just how much more performance we can get by scaling up model and training data size. Researchers have already studied scaling laws for natural language models, but a recent paper from Google Research provides further detail. The researchers systematically conducted more than 4,800 experiments on Vision Transformer, MLP-Mixer and ResNet architectures with 10 million to 10 billion parameters and evaluated them on over 20 downstream image recognition tasks. Among other things, they found that with increasing performance on upstream tasks, downstream tasks reached different saturation points. The researchers concluded that scaling does not lead to a "one size fits all" model and that researchers should make architectural choices to improve performance on a range of downstream tasks, rather than focusing on just one. Furthermore, upstream and downstream performance might sometimes be in tension. For researchers and engineers, the answer to developing capable models seems not to be as simple as recent advances might suggest.
A recent investigation has found that deepfake audio technology was used to mislead a bank manager from the United Arab Emirates to distribute funds to a theif's account. The bank manager reported receiving a phone call from someone who sounded liked the director of a company whom the bank manager personally knows, and he transferred the money as requested. Deepfake audios use neural networks, often trained on a dataset of the voice of a particular person, to generate new voice that imitates that person. It is notable that in this case, running deepfake audio during a phone call requires real-time neural network speech generation, something that has only been realized recently.
📚 Last Week in AI is growing!
With 2500+ subscribers, we want to bring you more, and ask for your help by becoming a paid subscriber.
While the newsletter will remain free and unchanged, paid subscribers will have access to the following perks:
Weekly op-eds on the latest developments in AI
Early access to news briefs
Submit AI topic requests for in-depth research and coverage
Are Patches All You Need? New Study Proposes Patches Are Behind Vision Transformers’ Strong Performance - "Vision transformer architectures (ViTs) have achieved compelling performance across many computer vision tasks, often outperforming classical convolutional architectures."
AI Analysis of 100,000 Climate Studies Reveals How Massive The Crisis Already Is - "Some problems are so big, you can't really see them. Climate change is the perfect example. The basics are simple: the climate is heating up due to fossil fuel use. But the nitty gritty is so vast and complicated that our understanding of it is always evolving."
Facebook Loves Self-Supervised Learning. Period. - "Facebook’s chief AI scientist Yann LeCun’s influence seems to have rubbed off on the team, taking a path less travelled – a journey towards self-supervision."
A New Link to an Old Model Could Crack the Mystery of Deep Learning - "To help them explain the shocking success of deep neural networks, researchers are turning to older but better-understood models of machine learning. In the machine learning world, the sizes of artificial neural networks — and their outsize successes — are creating conceptual conundrums."
Duke computer scientist wins ‘Nobel Prize’ worth $1M for artificial intelligence work - "DURHAM – Whether preventing explosions on electrical grids, spotting patterns among past crimes, or optimizing resources in the care of critically ill patients, Duke University computer scientist Cynthia Rudin wants artificial intelligence (AI) to show its work."
This robot looks for lost items - "A busy commuter is ready to walk out the door, only to realize they’ve misplaced their keys and must search through piles of stuff to find them. Rapidly sifting through clutter, they wish they could figure out which pile was hiding the keys."
AI Predicts Accident Hot-Spots From Satellite Imagery and GPS Data - "Researchers from MIT and the Qatar Center for Artificial Intelligence have developed a machine learning system that analyzes high-resolution satellite imagery, GPS coordinates and historical crash data in order to map potential accident-prone sections in road networks, successfully predicting accide"
Ottawa project uses artificial intelligence to get COVID tests to those who are hard to reach - "The molecular COVID-19 self-test consists of a self-nasal swab and a test tube containing an agent. Thirty minutes after placing the swab in the test tube, it flashes positive or negative."
Waymo's self-driving taxi struggles with left turns and puddles. But it's still winning over some Arizona riders - "Waymo says its self-driving vans have given tens of thousands of rides since the company launched a public robotaxi service in Chandler, Arizona a year ago."
AI chipmaker Hailo raises $136M as it doubles down new opportunities for AI modules in edge devices - "Amid a global semiconductor shortage, an upstart in the world of AI chips is announcing a big round of funding to meet a boom in demand for its technology."
How much do Waymo Via’s autonomous trucks benefit from driverless cars? - "More than any autonomous trucking startup, Waymo Via claims that a dozen years of self-driving car development from the original Google self-driving car program affords unparalleled lessons for creating autonomous trucks. The case is one that only one other autonomous trucking startup can make."
Aurora’s autonomous trucks and taxis will be available to customers via subscription - "Aurora, the autonomous vehicle company founded by the former lead engineer for Google’s self-driving car project, announced that its autonomous trucks and taxis will be available to customers via subscriptions."
Building better startups with responsible AI - "Founders tend to think responsible AI practices are challenging to implement and may slow the progress of their business. They often jump to mature examples like Salesforce’s Office of Ethical and Humane Use and think that the only way to avoid creating a harmful product is building a big team."
New robots patrolling for 'anti-social behaviour' causing unease in Singapore streets - "There are new sheriffs in town in Singapore, and they are unnerving many who live there."
Are Under-Curated Hyperscale AI Datasets Worse Than The Internet Itself? - "Researchers from Ireland, the UK and the US have warned that the growth in hyperscale AI training datasets threaten to propagate the worst aspects of their internet sources, contending that a recently-released academic dataset features ‘troublesome and explicit images and text pairs of rape, porno"
Israel's "AI Kill" against Iran has no long-term strategic benefit, but profound moral and legal ramifications - "Israel's alleged use of AI in its assassination program against Iranian nuclear scientists may have enamoured its admirers, but offers no long-term benefit and plenty of moral and legal danger, writes Richard Silverstein. "
Dead-End SF Street Plagued With Confused Waymo Cars Trying To Turn Around ‘Every 5 Minutes’ - "A normally quiet neighborhood in San Francisco is buzzing about a sudden explosion of traffic. Neighbors say their Richmond District dead-end street has suddenly become crowded with Waymo vehicles. The visitors Jennifer King is talking about don’t just come at night."
War dogs - "Actuator: Vineyards, guns and money"
Peter Norvig: Today’s Most Pressing Questions in AI Are Human-Centered - "The AI expert, who joins Stanford HAI as a Distinguished Education Fellow, discusses building inclusive education and broadening access to students."
Copyright © 2021 Skynet Today, All rights reserved.