Technology (Page 32)

You have a lot of scientists, mathematicians, and engineers to thank if your teeth dropped while you watched the most recent AI-generated videos, your bank balance was saved from crooks by a fraud detection system, or your time was made a much easier because you were allowed to determine a text message while on the go.

But two names, Princeton University scientist John Hopfield and University of Toronto system professor Geoffrey Hinton, stand out for fundamental contributions to the profound learning technology that makes those experience possible.

For their ground-breaking work in the field of artificial neural networks, the two experts received the Nobel Prize in Physics on October 8, 2024. Although natural neural networks are the modeled for artificial neural networks, both researchers ‘ work relied on statistical science, which is why the prize in science was awarded.

How a nerve computes

The study of natural cells in living neurons is where artificial neural networks come from. A straightforward concept of the neuron’s functioning was developed by neurophysiologist Warren McCulloch and mathematician Walter Pitts in 1943. A synapse is connected to its surrounding cells in the McCulloch-Pitts design, and they can send signals to them. Therefore, it can incorporate those signs to give signals to additional neurons.

But there is a spin: It does weigh signals coming from different companions separately. Consider whether you are trying to decide whether to purchase a brand-new smartphone. You talk to your buddies and ask them for their suggestions.

Collect all companion tips and choose to go along with what the majority of them say is a straightforward method. For example, you ask three companions, Alice, Bob and Charlie, and they say yay, yay and no, respectively. Because you have two yays and one no, you decide to purchase the telephone.

However, you may believe some friends more because they have in-depth knowledge of technological devices. So you might decide to give more weight to their suggestions. For instance, if Charlie is very experienced, you might qualify his nay three times before deciding to not purchase the phone.

If you’re unfortunate to have a friend who fully despises you in terms of technical gadgets, you may even give them a bad name. Their phew is therefore counted as both a phew and a yay.

When you’ve made your own choice about whether the new phone is a good choice, another friends may ask you for your advice. Also, in artificial and natural neural networks, neurons may index signals from their relatives and give a signal to other neurons.

This potential leads to a vital variation: Is there a cycle in the system? For instance, if I ask Alice, Bob and Charlie today, and tomorrow Alice asks me for my advice, then there is a period: from Alice to me, and from me again to Alice.

If the connections between neurons do n’t have a cycle, then computer scientists refer to it as a feedforward neural network. A proposes network’s cells may be arranged in layers.

The sources are the first part. The second level sends its signals to the second level, and so on. The network’s outcomes are reflected in the final level.

However, if the system contains a pattern, computer experts refer to it as a recurrent neural network, and the plans of cells can be more challenging than those in feedforward neural networks.

Hopfield network

Biology served as the initial source of artificial neural networks ‘ inspiration, but soon other fields began to influence their development. These included logic, mathematics and physics.

The Hopfield network, or Hopfield network, was a particular type of recurrent neural network that the physicist John Hopfield studied. In particular, he studied their dynamics: What happens to the network over time?

Similar dynamics are crucial when social networks transmit information. Everyone is aware of the rise in memes and the creation of echo chambers in online social networks. These are all collective phenomena that ultimately result from straightforward information exchanges between network users.

Hopfield was the first to investigate the dynamics of recurrent neural networks by using physics-based models, particularly those created for studying magnetism. He further demonstrated that such neural networks can have a memory function using their dynamics.

Boltzmann machines and backpropagation

During the 1980s, Geoffrey Hinton, computational neurobiologist Terrence Sejnowski and others extended Hopfield’s ideas to create a new class of models called Boltzmann machines, named for the 19th-century physicist Ludwig Boltzmann.

As the name suggests, Boltzmann’s statistical physics is the inspiration for the design of these models. Boltzmann machines could generate new patterns, planting the seeds of the modern generative AI revolution, in contrast to Hopfield networks, which could store patterns and correct errors in patterns like a spellchecker does.

Hinton was also part of another breakthrough that happened in the 1980s: backpropagation. You must somehow select the appropriate weights for the connections between artificial neurons if you want them to perform interesting tasks.

Backpropagation is a crucial algorithm that enables the selection of weights based on the network’s performance on a training dataset. However, training complex artificial neural networks continued to be challenging.

Hinton and his coworkers figured out how to train multilayer networks using Boltzmann machines in the 2000s by first pretraining the network layer by layer, then applying a different fine-tuning algorithm to the pre-trained network in order to further adjust the weights. Deep networks were given the name of multiple layers, and the deep learning revolution had already begun.

AI pays it back to physics

The physics Nobel Prize demonstrates how physics’ ideas contributed to the development of deep learning.

Deep learning has now begun to pay its respects to physics, allowing quick and accurate simulations of everything from molecules and materials to the climate of the planet.

The prize committee’s decision to award Hopfield and Hinton the Nobel Prize in physics demonstrates its belief in humanity’s ability to use these discoveries to advance human development and create a sustainable world.

Ambuj Tewari is professor of statistics, University of Michigan

The Conversation has republished this article under a Creative Commons license. Read the original article.

Continue Reading

The US Department of Commerce has announced an empty competitors to show “how AI can help in developing new green silicon materials and processes that meet industry requirements and can be designed and adopted within five decades.”

Under Secretary of Commerce for Standards and Technology Laurie Locascio calls this” a special chance to make the United States a world leader in effective, safe, high-volume, and dynamic silicon manufacturing”. Locascio serves as the director of the National Institute of Standards and Technology.

Up to US$ 100 million will be awarded by the CHIPS Research and Development Office ( CHIPS R&amp, D) to winners who “develop university-led, industry-informed, collaborations about artificial intelligence-powered autonomous experimentation ( AI/AE ) relevant to sustainable semiconductor manufacturing”.

The US CHIPS and Science Act, which US President Joe Biden signed into law in August 2022, established CHIPS R&amp, D. The Department of Commerce is given$ 50 billion for initiatives that aim to revive and strengthen US semiconductor production and R&amp, D.

Of that sum,$ 39 billion went to the CHIPS Program Office for investment in infrastructure and technology in the United States, including high-profile companies being built by Taiwan’s TSMC and America’s Intel. $ 11 billion was allocated to CHIPS R&amp, D for projects such as this one.

The Commerce Department notes that” for the US semiconductor business to prosper in the long-term, it must be able to develop innovative and economically dynamic systems to safely make materials and production chips in a way that protects the environment and local communities.”

That seems clear. After all, the largest American manufacturer of semiconductor production equipment named itself Applied Materials, whereas silicon companies spend billions annually on developing more advanced integrated circuits, using electricity and water more effectively in the production process, and reducing industrial waste and greenhouse gas emissions.

However, Gina Raimondo, the secretary of commerce, feels more necessity. ” Best now”, she says, “new silicon materials generally take years to become production-ready and are very resource-intensive.

We need to use AI to rapidly develop green material processes if we want to quickly expand America’s semiconductor manufacturing base in a way that is long-term sustainable in the face of growing threats from the climate crisis.

Raimondo even feels a sense of vision, saying,” With this new software, the Biden-Harris management will harness the huge potential of our workers and innovators to create a more stable and enduring home semiconductor industry.

What are these presentations all around, aside from making grandiose claims for a sum of money that seems like a drop in the bucket in comparison to the billions of dollars spent annually on R&amp, D? ( In the second quarter of this year alone, Intel’s R&amp, D budget was$ 4.2 billion. )

The answer is that AI/AE, which combines system understanding and automatic facilities, is” ushering in a paradigm change in materials science, “according to Taro Hitosugi, Ryota Shimizu and Naoya Ishizuki of the Tokyo Institute of Technology”. These systems make decisions and carry out all exploratory steps without the need for human intervention by using computer systems and robots.

” Given the possible mixtures of elements,” they continue”, there is an almost infinite number of new materials … Thus, optimizing high-dimensional synthesis parameters in a vast search area is necessary for materials production … In a way, the world of materials is a border for investigation, much like place or the deep water.”

AI/AE should enable a vast acceleration of the process of materials discovery and synthesis, not only in the semiconductor industry but across the spectrum of applied science, from electronics, energy, aerospace and defense to biology, chemistry and pharmaceuticals.

Writing in Nature Synthesis, Eugenia Kumacheva of the University of Toronto and Milad Abolhasani of North Carolina State University write:

Through the integration of machine learning, lab automation, and robotics, the recent growth of data science and automated experimentation techniques has led to the development of self-driving labs ( SDLs ).

An SDL is a machine-learning-assisted modular experimental platform that iteratively operates a series of experiments selected by the machine-learning algorithm to achieve a user-defined objective. Through quick exploration of the chemical space, these intelligent robotic assistants aid researchers in accelerating the pace of fundamental and applied research.

The main benefit of SDLs is the “research acceleration” to produce new knowledge that can lead to the development of novel compounds or manufacturing processes for the best-performing materials 10 to 1000 times more quickly than with one-at-a-time variable exploration or combinatorial experiments.

In other words, AI and robots can perform tasks much more effectively than trial and error that are scientifically sound. According to researchers led by Professor Alán Aspuru-Guzik of the University of Toronto’s Department of Chemistry,

The Aspuru-Guzik group’s goal is to reduce the amount of time and money needed to develop new functional materials or improve existing ones by a factor of ten, namely from ten million dollars and ten years of development to one million dollars and one year. This will eventually change the way we conduct scientific research.

Aspuru-Guzik is also a professor of computer science, a member of the University of Toronto’s strategic initiative Acceleration Consortium, which brings together researchers from industry, government, and academia.

This may be the model for the US Commerce Department’s semiconductor materials initiative. In addition, the department’s AI/AE competition bears a strong resemblance to the SDL Grand Challenge proposed by the Washington-based Center for Strategic and International Studies ( CSIS ) think tank in its January 2024 report entitled” Self-Driving Labs: AI and Robotics Accelerating Materials Innovation.”

The CSIS report asks whether the United States is giving enough policy attention and resources to ensure the advantage in SDLs, stating that” the development and adoption of alternative and new materials is central to US leadership in emerging technologies.

At that time, according to CSIS, US spending on SDLs was less than$ 50 million and” not done in a directed, programmatic manner, “while Canada had awarded$ 200 million to the Acceleration Consortium at the University of Toronto.

In this context, the US Commerce Department’s$ 100 million award will be a belated but meaningful step forward. Its five-year time frame matches the semiconductor industry’s roadmap to 1nm process technology.

CSIS also pointed out that the University of Liverpool, Lawrence Berkeley National Lab, Argonne National Lab, and Carnegie Mellon University were creating SDLs, noting that” University of Liverpool researchers in 2020 used a mobile platform robot arm to create and search for catalysts across 10 design parameters, ultimately conducting 688 experiments over eight days completely autonomously and identifying chemical formulations that were 6 times better than the baseline. ” &nbsp,

Imec, the Inter-university Microelectronics Centre headquartered in Belgium that conducts advanced R&amp, D with and for the semiconductor industry, is using AI to identify new materials. For example, scientists affiliated with imec write:

Semiconductors are becoming more challenging to manufacture as a result of decreasing dimensions and increasing complexity. In particular, the allowed deposition temperature becomes lower. Amorphous materials, which do not require annealing steps, are therefore becoming more interesting.

However, modeling crystalline materials is much more challenging than modeling crystalline ones from first principles. Especially to screen for new materials, a fully&nbsp, ab initio approach is hence too expensive. We address this issue by combining high throughput first principles calculations with artificial intelligence ( AI ).

The Johns Hopkins University Applied Physics Laboratory ( APL) is using AI to accelerate the development of new materials capable of withstand the harsh conditions that characterize deep-sea exploration, space exploration, hypersonic vehicles, and other applications that are related to national security.

Morgan Trexler, program manager for Science of Extreme and Multifunctional Materials&nbsp, at APL says,

There are more operations in austere environments as the US faces pressing national security challenges, and those operations require revolutionary new materials. We ca n’t wait for decades to find the materials that will satisfy those demands. By infusing AI approaches throughout the discovery process, we can more quickly and intentionally identify materials for complex, specific applications.

Keith Caruso, chief scientist at APL’s Research and Exploratory Development Department, adds that” The approach to building on existing materials will only ever yield limited improvements. To create groundbreaking materials, we need to make a fundamental leap.”

The RIKEN National Research and Development Agency in Japan is utilizing high-performance computing and AI for genomic medicine and drug discovery.

Japanese analytical instrument maker Shimadzu Corporation, which works with Kobe University, is targeting” a platform for autonomous scientific discoveries by robots and AI” as its vision of future laboratories for the development of new materials, pharmaceuticals and biotechnology, including” smart cells” with altered genes.

According to Science China Press,” the idea of large materials models as deep-learning computational models for materials design has attracted great interest.”

Researchers at Tsinghua University are pursuing the creation of “models” that can handle a range of material structures across the periodic table’s various components.

A robotic chemist with an AI background and a team of Chinese scientists created a catalyst to produce oxygen from Martian meteorites almost a year ago, according to China Daily.

The catalyst can consistently produce oxygen without apparent deterioration, according to a stress test at minus 37 degrees Celsius, which suggests it can operate in the harsh conditions on Mars.

Although it is unclear what the Chinese are doing with the development of autonomous materials for the semiconductor and other industries, it is possible.

Follow this writer on&nbsp, X: @ScottFo83517667

Continue Reading