AI & Robotics News and Discussions

[Yuli Ban]

[Yuli Ban]

The days and sometimes weeks it took to train AIs only a few years ago was a big reason behind the launch of billions of dollars-worth of new computing startups over the last few years—including Cerebras Systems, Graphcore, Habana Labs, and SambaNova Systems. In addition, Google, Intel, Nvidia and other established companies made their own similar amounts of internal investment (and sometimes acquisition). With the newest edition of the MLPerf training benchmark results, there’s clear evidence that the money was worth it.

The gains to AI training performance since MLPerf benchmarks began “managed to dramatically outstrip Moore’s Law,” says David Kanter, executive director of the MLPerf parent organization MLCommons. The increase in transistor density would account for a little more than doubling of performance between the early version of the MLPerf benchmarks and those from June 2021. But improvements to software as well as processor and computer architecture produced a 6.8-11-fold speedup for the best benchmark results. In the newest tests, called version 1.1, the best results improved by up to 2.3 times over those from June.

According to Nvidia the performance of systems using A100 GPUs has increased more than 5-fold in the last 18 months and 20-fold since the first MLPerf benchmarks three years ago.

[Nero]

The days and sometimes weeks it took to train AIs only a few years ago was a big reason behind the launch of billions of dollars-worth of new computing startups over the last few years—including Cerebras Systems, Graphcore, Habana Labs, and SambaNova Systems. In addition, Google, Intel, Nvidia and other established companies made their own similar amounts of internal investment (and sometimes acquisition). With the newest edition of the MLPerf training benchmark results, there’s clear evidence that the money was worth it.

The gains to AI training performance since MLPerf benchmarks began “managed to dramatically outstrip Moore’s Law,” says David Kanter, executive director of the MLPerf parent organization MLCommons. The increase in transistor density would account for a little more than doubling of performance between the early version of the MLPerf benchmarks and those from June 2021. But improvements to software as well as processor and computer architecture produced a 6.8-11-fold speedup for the best benchmark results. In the newest tests, called version 1.1, the best results improved by up to 2.3 times over those from June.

According to Nvidia the performance of systems using A100 GPUs has increased more than 5-fold in the last 18 months and 20-fold since the first MLPerf benchmarks three years ago.

A 20-fold increase in 3 years? Can you imagine if we see the same sort of increase by 2024? We already know that people can make the false argument that GPT-3 is true AI - we will be truly beyond the pail within just a year at this rate. Truly we live in interesting times.

[Ozzie guy]

The days and sometimes weeks it took to train AIs only a few years ago was a big reason behind the launch of billions of dollars-worth of new computing startups over the last few years—including Cerebras Systems, Graphcore, Habana Labs, and SambaNova Systems. In addition, Google, Intel, Nvidia and other established companies made their own similar amounts of internal investment (and sometimes acquisition). With the newest edition of the MLPerf training benchmark results, there’s clear evidence that the money was worth it.

The gains to AI training performance since MLPerf benchmarks began “managed to dramatically outstrip Moore’s Law,” says David Kanter, executive director of the MLPerf parent organization MLCommons. The increase in transistor density would account for a little more than doubling of performance between the early version of the MLPerf benchmarks and those from June 2021. But improvements to software as well as processor and computer architecture produced a 6.8-11-fold speedup for the best benchmark results. In the newest tests, called version 1.1, the best results improved by up to 2.3 times over those from June.

According to Nvidia the performance of systems using A100 GPUs has increased more than 5-fold in the last 18 months and 20-fold since the first MLPerf benchmarks three years ago.

Am I misunderstanding something or is AI literally getting 20 times better every 3 years !?

If so victory is coming fast.

1X20=20 20X20=400 400X20=8000 by the logic of 20 times better every 3 years we will have an AI 8000 times better than GPT3 on 11 Jun 2029.

[Yuli Ban]

We can add suggesting and proving mathematical theorems to the long list of what artificial intelligence is capable of: Mathematicians and AI experts have teamed up to demonstrate how machine learning can open up new avenues to explore in the field.

While mathematicians have been using computers to discover patterns for decades, the increasing power of machine learning means that these networks can work through huge swathes of data and identify patterns that haven't been spotted before.

In a newly published study, a research team used artificial intelligence systems developed by DeepMind, the same company that has been deploying AI to solve tricky biology problems and improve the accuracy of weather forecasts, to unknot some long-standing math problems.

"Problems in mathematics are widely regarded as some of the most intellectually challenging problems out there," says mathematician Geordie Williamson from the University of Sydney in Australia.

[caltrek]

Engineered Arts, a UK-based designer and manufacturer of humanoid robots, recently showed off one of its most lifelike creations in a video posted on YouTube. The robot, called Ameca, is shown making a series of incredibly human-like facial expressions.

https://www.theverge.com/2021/12/5/2281 ... xpressions

[Yuli Ban]

The idiom “actions speak louder than words” first appeared in print almost 300 years ago. A new study echoes this view, arguing that combining self-supervised and offline reinforcement learning (RL) could lead to a new class of algorithms that understand the world through actions and enable scalable representation learning,

Machine learning (ML) systems have achieved outstanding performance in domains ranging from computer vision to speech recognition and natural language processing, yet still struggle to match the flexibility and generality of human reasoning. This has led ML researchers to search for the “missing ingredient” that might boost these systems’ ability to understand, reason and generalize.

In the paper Understanding the World Through Action, UC Berkeley assistant professor in the department of electrical engineering and computer sciences Sergey Levine suggests that a general, principled, and powerful framework for utilizing unlabelled data could be derived from RL to enable ML systems leveraging large datasets to better understand the real world.

[funkervogt]

The "UnifiedQA" and "Gopher" chatbots are better at answering certain categories of questions than GPT-3, and are nearing the same levels of accuracy as human subject matter experts.
https://deepmind.com/blog/article/langu ... g-at-scale

I predict the Turing Test will be passed this decade.

[Nero]

I do believe that it will not require anymore than three years to reach beyond the limits of the Turing Test. This time in 2018 GPT-2 was the most advanced AI chatbot, that has been surpassed several orders of magnitude in the years since there is no way that by 2024 or 2025 even if the growth somehow slowed we would not be looking at similar advances.

Again you have to credit Mr Yuli-Ban for most of this, he saw this coming and was able to predict this well in advance of today.

[Yuli Ban]

Researchers at Kobe University and Osaka University have successfully developed artificial intelligence technology that can extract hidden equations of motion from regular observational data and create a model that is faithful to the laws of physics.

This technology could enable researchers to discover the hidden equations of motion behind phenomena for which the laws were considered unexplainable

[Yuli Ban]

On December 8, Peng Cheng Laboratory (PCL) and Baidu held a press conference to announce the release of PCL-BAIDU Wenxin (version Ernie 3.0 Titan), the world’s first knowledge-enhanced 100-billion-scale pretrained language model and largest Chinese-language monolithic model. PCL-BAIDU Wenxin achieves state-of-the-art results on more than 60 natural language processing (NLP) tasks and significantly advances more than 30 benchmarks in zero-shot and few-shot learning.

PCL-BAIDU Wenxin is based on Peng Cheng’s industry-leading Pengcheng Cloud Brain II computing power system and Baidu’s PaddlePaddle deep learning platform, and its 260 billion parameters exceed GPT-3’s total by 50 percent. PCL-BAIDU Wenxin aims to solve common model bottlenecks such as poor generalization ability, strong reliance on expensive manually labelled data and high application cost; and simplify the development of artificial intelligence (AI) applications.

At the presser, PCL Director and Chinese Academy of Engineering academician Wen Gao said the large model is crucial for continuing technological development and innovation and will enable more industries to benefit from the power of AI.

[Yuli Ban]

Not a day passes without a fascinating snippet on the ethical challenges created by “black box” artificial intelligence systems. These use machine learning to figure out patterns within data and make decisions – often without a human giving them any moral basis for how to do it.

Classics of the genre are the credit cards accused of awarding bigger loans to men than women, based simply on which gender got the best credit terms in the past. Or the recruitment AIs that discovered the most accurate tool for candidate selection was to find CVs containing the phrase “field hockey” or the first name “Jared”.

More seriously, former Google CEO Eric Schmidt recently combined with Henry Kissinger to publish The Age of AI: And Our Human Future, a book warning of the dangers of machine-learning AI systems so fast that they could react to hypersonic missiles by firing nuclear weapons before any human got into the decision-making process. In fact, autonomous AI-powered weapons systems are already on sale and may in fact have been used.

Somewhere in the machine, ethics are clearly a good idea.

[Yuli Ban]

Large language models (e.g., GPT-3) have many significant capabilities, such as performing few-shot learning across a wide array of tasks, including reading comprehension and question answering with very few or no training examples. While these models can perform better by simply using more parameters, training and serving these large models can be very computationally intensive. Is it possible to train and use these models more efficiently?

In pursuit of that question, today we introduce the Generalist Language Model (GLaM), a trillion weight model that can be trained and served efficiently (in terms of computation and energy use) thanks to sparsity, and achieves competitive performance on multiple few-shot learning tasks. GLaM’s performance compares favorably to a dense language model, GPT-3 (175B) with significantly improved learning efficiency across 29 public NLP benchmarks in seven categories, spanning language completion, open-domain question answering, and natural language inference tasks.

[raklian]

A pre-, pre-, pre-taste for the Singularity to come.

[weatheriscool]

Machine learning speeds up vehicle routing

by Becky Ham, Massachusetts Institute of Technology
https://techxplore.com/news/2021-12-mac ... uting.html

Waiting for a holiday package to be delivered? There's a tricky math problem that needs to be solved before the delivery truck pulls up to your door, and MIT researchers have a strategy that could speed up the solution.

The approach applies to vehicle routing problems such as last-mile delivery, where the goal is to deliver goods from a central depot to multiple cities while keeping travel costs down. While there are algorithms designed to solve this problem for a few hundred cities, these solutions become too slow when applied to a larger set of cities.

To remedy this, Cathy Wu, the Gilbert W. Winslow Career Development Assistant Professor in Civil and Environmental Engineering and the Institute for Data, Systems, and Society, and her students have come up with a machine-learning strategy that accelerates some of the strongest algorithmic solvers by 10 to 100 times.

The solver algorithms work by breaking up the problem of delivery into smaller subproblems to solve—say, 200 subproblems for routing vehicles between 2,000 cities. Wu and her colleagues augment this process with a new machine-learning algorithm that identifies the most useful subproblems to solve, instead of solving all the subproblems, to increase the quality of the solution while using orders of magnitude less compute.

Their approach, which they call "learning-to-delegate," can be used across a variety of solvers and a variety of similar problems, including scheduling and pathfinding for warehouse robots, the researchers say.

[Yuli Ban]

Eureka Robotics, a tech spin-off from Nanyang Technological University, Singapore (NTU Singapore), has developed a technology, called Dynamis, that makes industrial robots nimbler and almost as sensitive as human hands, able to manipulate tiny glass lenses, electronics components, or engine gears that are just millimeters in size without damaging them.

This proprietary force feedback technology developed by NTU scientists was previously demonstrated by the Ikea Bot which assembled an Ikea chair in just 20 minutes. The breakthrough was first published in Science in 2018 and went viral on the internet when it could match the dexterity of human hands in assembling furniture.

NTU Associate Professor Pham Quang Cuong, Co-founder of Eureka Robotics, said they have since upgraded the software technology, which will be made available for a large number of industrial robots worldwide by Denso Wave, a market leader in industrial robots, which is part of the Toyota Group.

Clients purchasing the latest robots sold by Denso Wave will have an option to include this new technology as part of the force controller, which reads the force detected by a force sensor on the robot's wrist and applies force accordingly: Apply too little force and the items may not be assembled correctly while applying too much force could damage the items.

[caltrek]

Trial Identifies Five-drug Combo for ‘Ultra high risk’ Bone Marrow Cancer
December 12, 2021

https://www.eurekalert.org/news-releases/browse

Introduction:

(Institute of Cancer Research via EurekAlert!) A combination of five existing drugs keeps cancer at bay for longer in patients with a highly aggressive type of bone marrow cancer, a major new trial reveals.

The five-drug cocktail, along with a stem cell transplant, allowed people with ultra-high risk multiple myeloma to live longer before their disease progressed than those who received the standard of care.

The ‘MUK Nine OPTIMUM’ trial used a highly innovative methodology to open the door to the first tailored approach for people with the highest risk forms of myeloma – whose outcomes from treatment are currently poor.

It is also the latest study to demonstrate the benefit of combining drugs with different mechanisms of action to combat cancer’s ability to evolve and become resistant to treatment.

Further extract:

OPTIMUM is the first ‘digital comparator’ trial for multiple myeloma

caltrek's comment: Correct me if I am wrong, but I presume AI, or at least computers, would be used as part of the "digital"
comparison process.

[Yuli Ban]

[Yuli Ban]

Bringing a new robot to market is exciting: new capability, new hardware, new services. The problem is when you get to software, where everything feels harder and takes longer than you think it should. Like Tesla’s full self-driving, which has all the hardware and intelligence it needs — with the possible exception of LIDAR — but is perpetually just ... about ... to ... arrive ... and even so, was recently savaged by Consumer Reports as buggy and ineffective.

Hardware is necessary, but software provides the animating intelligence that allows it to do useful, efficient, and safe work.

That’s why Nader Elm, CEO of autonomous drone company Exyn Technologies, compared robots today to the iPhone before the App Store: the hardware’s there, but the software layer is immature.

Amazon’s hard at work on that.

And the early results are impressive, with easily 1000X cheaper training that is helping robot makers bring products to market 300X faster ... and enabling entire new capabilities.

[weatheriscool]

Moving toward the first flying humanoid robot
https://techxplore.com/news/2021-12-humanoid-robot.html
by Ingrid Fadelli , Tech Xplore

Researchers at the Italian Institute of Technology (IIT) have recently been exploring a fascinating idea, that of creating humanoid robots that can fly. To efficiently control the movements of flying robots, objects or vehicles, however, researchers require systems that can reliably estimate the intensity of the thrust produced by propellers, which allow them to move through the air.

As thrust forces are difficult to measure directly, they are usually estimated based on data collected by onboard sensors. The team at IIT recently introduced a new framework that can estimate thrust intensities of flying multibody systems that are not equipped with thrust-measuring sensors. This framework, presented in a paper published in IEEE Robotics and Automation Letters, could ultimately help them to realize their envisioned flying humanoid robot.

"Our early ideas of making a flying humanoid robot came up around 2016," Daniele Pucci, head of the Artificial and Mechanical Intelligence lab that carried out the study, told TechXplore. "The main purpose was to conceive robots that could operate in disaster-like scenarios, where there are survivors to rescue inside partially destroyed buildings, and these buildings are difficult to reach because of potential floods and fire around them."