AI & Robotics News and Discussions

[wjfox]

Researchers in Hong Kong create 'soft robot' made of magnetic slime

Researchers at The Chinese University of Hong Kong have created a "soft robot" made of slime containing magnetic particles, which can be manipulated using external magnets.

The magnetic particles are toxic, but have theoretically been made safe to enter the human body after being covered in a layer of silicone compound - although further safety testing will be needed in the future.

The team in Hong Kong hope the slime will one day be used to collect objects which have been accidentally swallowed.

https://www.bbc.co.uk/news/av/science-environment-60961184

[raklian]

[weatheriscool]

A new approach that could improve how robots interact in conversational groups
https://techxplore.com/news/2022-04-app ... roups.html
by Ingrid Fadelli , Tech Xplore

To effectively interact with humans in crowded social settings, such as malls, hospitals, and other public spaces, robots should be able to actively participate in both group and one-to-one interactions. Most existing robots, however, have been found to perform much better when communicating with individual users than with groups of conversing humans.

Hooman Hedayati and Daniel Szafir, two researchers at University of North Carolina at Chapel Hill, have recently developed a new data-driven technique that could improve how robots communicate with groups of humans. This method, presented in a paper presented at the 2022 ACM/IEEE International Conference on Human-Robot Interaction (HRI '22), allows robots to predict the positions of humans in conversational groups, so that they do not mistakenly ignore a person when their sensors are fully or partly obstructed.

"Being in a conversational group is easy for humans but challenging for robots," Hooman Hedayati, one of the researchers who carried out the study, told TechXplore. "Imagine that you are talking with a group of friends, and whenever one of your friends blinks, she stops talking and asks if you are still there. This potentially annoying scenario is roughly what can happen when a robot is in conversational groups."

One of the reasons why many robots occasionally misbehave while participating in a group conversation is that their actions heavily rely on data collected by their sensors (i.e., cameras, depth sensors, etc.). Sensors, however, are prone to errors, and can sometimes be obstructed by sudden movements and obstacles in the robot's surroundings.

[Yuli Ban]

In recent years, large neural networks trained for language understanding and generation have achieved impressive results across a wide range of tasks. GPT-3 first showed that large language models (LLMs) can be used for few-shot learning and can achieve impressive results without large-scale task-specific data collection or model parameter updating. More recent LLMs, such as GLaM, LaMDA, Gopher, and Megatron-Turing NLG, achieved state-of-the-art few-shot results on many tasks by scaling model size, using sparsely activated modules, and training on larger datasets from more diverse sources. Yet much work remains in understanding the capabilities that emerge with few-shot learning as we push the limits of model scale.

Last year Google Research announced our vision for Pathways, a single model that could generalize across domains and tasks while being highly efficient. An important milestone toward realizing this vision was to develop the new Pathways system to orchestrate distributed computation for accelerators. In “PaLM: Scaling Language Modeling with Pathways”, we introduce the Pathways Language Model (PaLM), a 540-billion parameter, dense decoder-only Transformer model trained with the Pathways system, which enabled us to efficiently train a single model across multiple TPU v4 Pods. We evaluated PaLM on hundreds of language understanding and generation tasks, and found that it achieves state-of-the-art few-shot performance across most tasks, by significant margins in many cases.

I like this gif

[Yuli Ban]

[agi]

deleted

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

All-DayErrDay

It’s crazy just how dominant this is. Sometimes I think about if OAI just released an improved version of their API with a new model that wasn’t supposed to be a complete makeover GPT 4. They could already do a decent job without having to train with a ridiculous amount of compute right now. I say this considering that I think they would want to avoid making the price of inference sky rocket. Look at the 62B model that has roughly the same compute as GPT 3, but performs significantly better than gopher with less compute and isn’t taking advantage of the data optimization we found out about recently. In fact going by table 20 it’s pretty much on par with Chinchilla. All of this improvement would definitely bring tangible benefits to peoples API experience.

The full model on its own is just so dominant and SOTA on almost everything out there. This continues to show the potential of ML. We just need more compute honestly. To just have it mentioned, someone with 20,000 H100s could train a model 10x this size in a reasonable amount of time (per my calculations - ignoring complications of scale). This probably took a few months so that’s what I’m comparing timeframe wise.

“PaLM is only the first step in our vision towards establishing Pathways as the future of ML scaling at Google and beyond. To that end, we chose to demonstrate this scaling capability on a well-studied, well-established recipe: a dense, decoder-only, full-attention Transformer model, which is trained to perform autoregressive language modeling. However, our wider goal is to explore a diverse array of novel architectural choices and training schemes, and combine the most promising systems with the extreme scaling capabilities of Pathways.”

“Input: Michael is at that really famous museum in France looking at its most famous painting. However, the artist who made this painting just makes Michael think of his favorite cartoon character from his childhood. What was the country of origin of the thing that the cartoon character usually holds in his hand? Model Output: The most famous painting in the Louvre is the Mona Lisa. The artist who made the Mona Lisa is Leonardo da Vinci. Leonardo da Vinci is also the name of the main character in the cartoon Teenage Mutant Ninja Turtles. Leonardo da Vinci is from Italy. The thing that Leonardo da Vinci usually holds in his hand is a katana. The country of origin of the katana is Japan. The answer is "Japan".”

[Yuli Ban]

starspawn0

There are several steps separating this from what you might call "AGI" -- e.g. long-term memory (not just x number of tokens, where x is small); a sense of time (instead of just a sequence of tokens where each could be separated by an indefinite duration; this could be solved by training on voice and/or video, both of which contain a sense of time); and need for multi-modal, which might help make it learn representations quicker than just using text.

One capability that would be interesting to see in these models is self-awareness. They may already have some degree of it. What would that look like? If you ask it to answer a question, and then ask it to explain what its weakness was in doing so, and it comes back with, "I wasn't very confident with that answer. I guessed correctly." -- and then you check the neural net's internal dynamics, and find that, indeed, the probability distribution for answers was spread out in such a way as to suggest "uncertainty". If all that happens, then it would suggest that the model has some degree of self-awareness (and you'd want to check lots more examples of how its self-analyses compare with its own internal dynamics).

If models attain some level of self-awareness, then they would know what they are weak at, and suggest how to improve. They could tell you what you should be doing to make them smarter.

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

DALL-E 2 is here
OpenAI's DALL·E 2

[caltrek]

Scientists Develop AI Camera To Take Full-Color Photos In Complete Darkness
by Dr Katie Spalding
April 7, 2022

https://www.iflscience.com/technology/s ... -darkness/

Introduction:

(IFL Science) Humans are, let’s face it, kind of the dunces of the animal kingdom. We can’t sniff stuff as well as dogs or bees; we can’t hear as well as bats; even our primary sense, sight, pales in comparison to animals that can see ultraviolet or infrared. In fact, the only advantage we have really isn’t a sense at all, but our big old brains.

Sometimes, though, that’s all you need as a species. We can’t see in the dark, but we can invent infrared cameras to do it for us – and now, researchers from the University of California, Irvine, have developed a way to make those images even closer to the real thing.

“Some night vision systems use infrared light that is not perceptible to humans and the images rendered are transposed to a digital display presenting a monochromatic image in the visible spectrum,” explains a paper describing the technology, published this week in the journal PLOS ONE.

“We sought to develop an imaging algorithm powered by optimized deep learning architectures whereby infrared spectral illumination of a scene could be used to predict a visible spectrum rendering of the scene as if it were perceived by a human with visible spectrum light,” the paper continues. “This would make it possible to digitally render a visible spectrum scene to humans when they are otherwise in complete ‘darkness’ and only illuminated with infrared light.”

So: a camera that can reconstruct color images from infrared light? Well, actually, no – not quite. The important bit isn’t the camera, but the algorithm the team used to reconstruct the images. They created a special type of AI known as a neural network – a kind of deep learning algorithm designed to simulate how human brains learn – which they then trained to spot correlations between how images look under infrared and under the visible spectrum.

[weatheriscool]

A robot that can put a surgical gown on a supine mannequin
https://techxplore.com/news/2022-04-rob ... equin.html
by Bob Yirka , Tech Xplore

A pair of researchers working in the Personal Robotics Laboratory at Imperial College London has taught a robot to put a surgical gown on a supine mannequin. In their paper published in the journal Science Robotics, Fan Zhang and Yiannis Demiris described the approach they used to teach the robot to partially dress the mannequin. Júlia Borràs, with Institut de Robòtica i Informàtica Industrial, CSIC-UPC, has published a Focus piece in the same journal issue outlining the difficulties in getting robots to handle soft material and the work done by the researchers on this new effort.

As researchers and engineers continue to improve the state of robotics, one area has garnered a lot of attention—using robots to assist with health care. In this instance, the focus was on assisting patients in a hospital setting who have lost the use of their limbs. In such cases, dressing and undressing falls to healthcare workers. Teaching a robot to dress patients has proven to be challenging due to the nature of the soft materials used to make clothes. They change in a near infinite number of ways, making it difficult to teach a robot how to deal with them. To overcome this problem in a clearly defined setting, Zhang and Demiris used a new approach.

[Yuli Ban]

We've got yet ANOTHER AI breakthrough


This week has been outrageous, historic even

[Yuli Ban]

Elon Musk gave a timeline to production for the first time for the Tesla Optimus project, a humanoid robot capable of doing general tasks.

The CEO believes the company can bring the ambitious project to production as soon as next year. It’s an ultra-ambitious timeline even for him.

When Tesla announced the “Tesla Bot” project at its A Day last year, Elon Musk presented it as something the company could do by leveraging existing work and parts from the development of self-driving technology, and if they don’t do it, someone else will.

At the time, it certainly didn’t sound like a priority for Tesla and many saw it mainly as a recruitment tool as the automaker is trying to boost its AI team to deliver its full self-driving system.

A few months later, the project’s priority level went up fast.

Musk announced that Tesla is now prioritizing product development of Tesla Bot, which he now calls Optimus, in 2022 over some of its upcoming vehicles.

The CEO appeared a lot more excited about the project and its potential to impact labor shortage and eventually the broader economy.

When first announcing the project, Tesla was aiming to have a prototype of the humanoid robot ready by the end of 2022, but there was no talk of a production timeline just yet.

[Yuli Ban]

[raklian]

So Ukraine has some pretty smart people, eh? See the Ukrainian flag on top of a building behind the dude explaining what he thought about PaLM at the beginning.

[Yuli Ban]