This article has been reviewed according to ScienceX's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

close

In a review published in Engineering, scientists explore the burgeoning field of machine learning (ML) and its applications in chemistry. Titled "Machine Learning for Chemistry: Basics and Applications," this comprehensive review aims to bridge the gap between chemists and modern ML algorithms, providing insights into the potential of ML in revolutionizing chemical research.

Over the past decade, ML and artificial intelligence (AI) have made remarkable strides, bringing us closer to the realization of intelligent machines. The advent of deep learning methods and enhanced data storage capabilities has played a pivotal role in this progress. ML has already demonstrated success in domains such as image and speech recognition, and now it is gaining significant attention in the field of chemistry, which is characterized by complex data and diverse organic molecules.

However, chemists often face challenges in adopting ML applications due to a lack of familiarity with modern ML algorithms. Chemistry datasets typically exhibit a bias towards successful experiments, while a balanced perspective necessitates the inclusion of both successful and failed experiments. Furthermore, incomplete documentation of synthetic conditions in literature poses additional challenges.

Computational chemistry, where datasets can be reliably constructed from quantum mechanics calculations, has embraced ML applications more readily. Nonetheless, chemists need a basic understanding of ML to harness the potential of data recording and ML-guided experiments.

This review serves as an introductory guide to popular chemistry databases, two-dimensional (2D) and three-dimensional (3D) features used in ML models, and popular ML algorithms. It delves into three specific chemistry fields where ML has made significant progress: retrosynthesis in organic chemistry, ML-potential-based atomic simulation, and ML for heterogeneous catalysis.

These applications have either accelerated research or provided innovative solutions to complex problems. The review concludes with a discussion of future challenges in the field.

The rapid advancement of computing facilities and the development of new ML algorithms indicate that even more exciting ML applications are on the horizon, promising to reshape the landscape of chemical research in the ML era. While the future is difficult to predict in such a fast-evolving field, it is undeniable that the development of ML models will lead to enhanced accessibility, generality, accuracy, intelligence, and ultimately, higher productivity.

The integration of ML models with the Internet offers a promising avenue for sharing ML predictions worldwide.

However, the transferability of ML models in chemistry poses a common challenge due to the diverse element types and complex materials involved. Predictions often remain limited to local datasets, resulting in decreased accuracy beyond the dataset.

To address this issue, new techniques such as the global neural network (G-NN) potential and improved ML models with more fitting parameters are being explored. While ML competitions in data science have produced exceptional algorithms, there is a need for more open ML contests in chemistry to nurture young talent.

Excitingly, end-to-end learning, which generates final output from raw input rather than designed descriptors, holds promise for more intelligent ML applications. AlphaFold2, for example, utilizes the one-dimensional (1D) structure of a protein to predict its 3D structure. Similarly, in the field of heterogeneous catalysis, an end-to-end AI model has successfully resolved reaction pathways. These advanced ML models can also contribute to the development of intelligent experimental robots for high-throughput experiments.

As the field of ML continues to evolve rapidly, it is crucial for chemists and researchers to stay informed about its applications in chemistry. This review serves as a valuable resource, providing a comprehensive overview of the basics of ML and its potential in various chemistry domains. With the integration of ML models and the collective efforts of the scientific community, the future of chemical research holds immense promise.

More information: Yun-Fei Shi et al, Machine Learning for Chemistry: Basics and Applications, Engineering (2023). DOI: 10.1016/j.eng.2023.04.013

See more here:
Machine learning for chemistry: Basics and applications - Phys.org

Apple may not be as flashy as other companies in adopting artificial intelligence features. Still, the already has a lot of smarts scattered throughout iOS.

Apple does not go out of its way to specifically name-drop "artificial intelligence" or AI meaningfully, but the company isn't avoiding the technology. Machine learning has become Apple's catch-all for its AI initiatives.

Apple uses artificial intelligence and machine learning in iOS in several noticeable ways. Here is a quick breakdown of where you'll find it.

It has been several years since Apple started using machine learning in iOS and other platforms. The first real use case was Apple's software keyboard on the iPhone.

Apple utilized predictive machine learning to understand which letter a user was hitting, which boosted accuracy. The algorithm also aimed to predict what word the user would type next.

Machine learning, or ML, is a system that can learn and adapt without explicit instructions. It is often used to identify patterns in data and provide specific results.

This technology has become a popular subfield of artificial intelligence. Apple has also been incorporating these features for several years.

In 2023, Apple is using machine learning in just about every nook and cranny of iOS. It is present in how users search for photos, interact with Siri, see suggestions for events, and much, much more.

On-device machine learning systems benefit the end user regarding data security and privacy. This allows Apple to keep important information on the device rather than relying on the cloud.

To help boost machine learning and all of the other key automated processes in iPhones, Apple made the Neural Engine. It launched with the iPhone's A11 Bionic processor to help with some camera functions, as well as Face ID.

Siri isn't technically artificial intelligence, but it does rely on AI systems to function. Siri taps into the on-device Deep Neural Network, or DNN, and machine learning to parse queries and offer responses.

Siri can handle various voice- and text-based queries, ranging from simple questions to controlling built-in apps. Users can ask Siri to play music, set a timer, check the weather, and much more.

Apple introduced the TrueDepth camera and Face ID with the launch of the iPhone X. The hardware system can project 30,000 infrared dots to create a depth map of the user's face. The dot projection is paired with a 2D infrared scan as well.

That information is stored on-device, and the iPhone uses machine learning and the DNN to parse every single scan of the user's face when they unlock their device.

This goes beyond iOS, as the stock Photos app is available on macOS and iPadOS as well. This app uses several machine learning algorithms to help with key built-in features, including photo and video curation.

Apple's Photos app using machine learning

Facial recognition in images is possible thanks to machine learning. The People album allows searching for identified people and curating images.

An on-device knowledge graph powered by machine learning can learn a person's frequently visited places, associated people, events, and more. It can use this gathered data to automatically create curated collections of photos and videos called "Memories."

Apple works to improve the camera experience for iPhone users regularly. Part of that goal is met with software and machine learning.

Apple's Deep Fusion optimizes for detail and low noise in photos

The Neural Engine boosts the camera's capabilities with features like Deep Fusion. It launched with the iPhone 11 and is present in newer iPhones.

Deep Fusion is a type of neural image processing. When taking a photo, the camera captures a total of nine shots. There are two sets of four shots taken just before the shutter button is pressed, followed by one longer exposure shot when the button is pressed.

The machine learning process, powered by the Neural Engine, will kick in and find the best possible shots. The result leans more towards sharpness and color accuracy.

Portrait mode also utilizes machine learning. While high-end iPhone models rely on hardware elements to help separate the user from the background, the iPhone SE of 2020 relied solely on machine learning to get a proper portrait blur effect.

Machine learning algorithms help customers automate their general tasks as well. ML makes it possible to get smart suggestions regarding potential events the user might be interested in.

For instance, if someone sends an iMessage that includes a date, or even just the suggestion of doing something, then iOS can offer up an event to add to the Calendar app. All it takes is a few taps to add the event to the app to make it easy to remember.

There are more machine learning-based features coming to iOS 17:

One of Apple's first use cases with machine learning was the keyboard and autocorrect, and it's getting improved with iOS 17. Apple announced in 2023 that the stock keyboard will now utilize a "transformer language model," significantly boosting word prediction.

The transformer language model is a machine learning system that improves predictive accuracy as the user types. The software keyboard also learns frequently typed words, including swear words.

Apple introduced a brand-new Journal app when it announced iOS 17 at WWDC 2023. This new app will allow users to reflect on past events and journal as much as they want in a proprietary app.

Apple's stock Journal app

Apple is using machine learning to help inspire users as they add entries. These suggestions can be pulled from various resources, including the Photos app, recent activity, recent workouts, people, places, and more.

This feature is expected to arrive with the launch of iOS 17.1.

Apple will improve dictation and language translation with machine learning as well.

Machine learning is also present in watchOS with features that help track sleep, hand washing, heart health, and more.

As mentioned above, Apple has been using machine learning for years. Which means the company has technically been using artificial intelligence for years.

People who think Apple is lagging behind Google and Microsoft are only considering chatGPT and other similar systems. The forefront of public perception regarding AI in 2023 is occupied by Microsoft's AI-powered Bing and Google's Bard.

Apple is going to continue to rely on machine learning for the foreseeable future. It will find new ways to implement the system and boost user features in the future.

It is also rumored Apple is developing its own chatGPT-like experience, which could boost Siri in a big way at some point in the future. In February 2023, Apple held a summit focusing entirely on artificial intelligence, a clear sign it's not moving away from the technology.

Apple can rely on systems it's introducing with iOS 17, like the transformer language model for autocorrect, expanding functionality beyond the keyboard. Siri is just one avenue where Apple's continued work with machine learning can have user-facing value.

Apple's work in artificial intelligence is likely leading to the Apple Car. Whether or not the company actually releases a vehicle, the autonomous system designed for automobiles will need a brain.

See the original post here:
How Apple is already using machine learning and AI in iOS - AppleInsider

If you have a curiosity about how fancy graphics cards actually work, and why they are so well-suited to AI-type applications, then take a few minutes to read[Tim Dettmers] explain why this is so. Its not a terribly long read, but while it does get technical there are also car analogies, so theres something for everyone!

He starts off by saying that most people know that GPUs are scarily efficient at matrix multiplication and convolution, but what really makes them most useful is their ability to work with large amounts of memory very efficiently.

Essentially, a CPU is a latency-optimized device while GPUs are bandwidth-optimized devices. If a CPU is a race car, a GPU is a cargo truck. The main job in deep learning is to fetch and move cargo (memory, actually) around. Both devices can do this job, but in different ways. A race car moves quickly, but cant carry much. A truck is slower, but far better at moving a lot at once.

To extend the analogy, a GPU isnt actually just a truck; it is more like a fleet of trucks working in parallel. When applied correctly, this can effectively hide latency in much the same way as an assembly line. It takes a while for the first truck to arrive, but once it does, theres an unbroken line of loaded trucks waiting to be unloaded. No matter how quickly and efficiently one unloads each truck, the next one is right there, waiting. Of course, GPUs dont just shuttle memory around, they can do work on it as well.

The usual configuration for deep learning applications is a desktop computer with one or more high-end graphics cards wedged into it, but there are other (and smaller) ways to enjoy some of the same computational advantages without eating a ton of power and gaining a bunch of unused extra HDMI and DisplayPort jacks as a side effect. NVIDIAs line of Jetson development boards incorporates the right technology in an integrated way. While it might lack the raw horsepower (and power bill) of a desktop machine laden with GPUs, theyre no slouch for their size.

Read more:
Here's Why GPUs Are Deep Learning's Best Friend - Hackaday

This article has been reviewed according to ScienceX's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

fact-checked

peer-reviewed publication

trusted source

proofread

close

In the past two decades, the carbon-nitrogen bond forming reaction, known as the Buchwald-Hartwig reaction, has become one of the most widely used tools in organic synthesis, particularly in the pharmaceutical industry given the prevalence of nitrogen in natural products and pharmaceuticals.

This powerful reaction has revolutionized the way nitrogen-containing compounds are made in academic and industrial laboratories, but it requires lengthy, time-consuming experimentation to determine the best conditions for a highly effective reaction.

Now, Illinois researchers in collaboration with chemists at Hoffman La-Roche, a pharmaceutical company in Switzerland, have developed a machine learning tool that predicts in a matter of minutes the best conditions for a high-yielding reaction with no lengthy experimentation.

In a recently published article in Science, Illinois chemistry professor Scott Denmark and Ian Rinehart, a recent Ph.D. graduate in the Denmark lab, describe how they developed, trained, and tested their machine learning model to drastically accelerate the identification of substrate-adaptive conditions for this palladiumcatalyzed carbon-nitrogen bond forming reaction.

Denmark said this reaction is a very general transformation so there is much structural diversity among reactant pairings and a lot of "levers to pull" to make it work.

"And that's what we have figured out," Denmark said.

close

User guides and cheat sheets have evolved in the nearly 30 years since this reaction was discovered, and they can provide some direction, Rinehart explained, but experimentation is often necessary. Basically, a trial-and-error process in a lab.

"It's a problem that everyone in the pharmaceutical industry recognized was ripe for intervention by informatics methods," Denmark said. "Lots of people have tried to use the US Patent and Trademark Office or Chemical Abstracts or other huge databases to try to model to make predictive tools for this one very important reaction. But they haven't been able to do very well because the information in the literature is just not very reliable."

The design and construction of their machine learning tool required the generation of an experimental dataset that explores a diverse network of reactant pairings across a set of reaction conditions. A large scope of CN couplings was actively learned by neural network models by using a systematic process to design experiments.

The challenge for a project like this, Denmark said, was the amount of potential data to collect and the thousands and thousands of experiments required to build a database of information for modeling.

"One of Ian's biggest contributions was figuring out the workflow to decide what experiments to do to get a valid predictive model with about 3,500 experiments and still be able to make predictions without an enormous database," Denmark said.

They also experimentally validated the predictions from the machine learning tool.

"We tested them and found with pretty good statistics that the conditions were producing compounds when we expected," Denmark said.

The researchers report that their models showed good performance in experimental validation: Ten products were isolated in more than 85% yield from a range of couplings with out-of-sample reactants designed to challenge the models.

Rinehart said they taught machine learning models to have a kind of chemical intuition like what an expert has.

"So, we have now run or talked about so many of these couplings that we have a good intuition about what's going to happen, but someone who hadn't run hundreds or thousands of these might not have a good first guess. We have taught a model at a much more granular level [than user guides] to have an intuition. It's not perfect. But that's kind of the point. It doesn't have to be. It just has to get you to the answer faster," Rinehart said.

And the coolest part, Rinehart explained, is that intuition gets honed over time as more people use the machine learning tool. The developed workflow continually improves the prediction capability of the tool as the corpus of data grows.

"It's an exciting time as data science merges with chemistry," Denmark said. "And this is the perfect marriage. A lot of people recognized this, but no one has done it, at least not in a meaningful way that is experimentally validated."

The Denmark group is creating a cloud-based version of the workflow to enable scientists around the world to use this tool which will continuously add data to improve the model as more structurally diverse substrates are tested and different catalysts and conditions are added to the database.

Rinehart said the code is public and on an open-source license, so anyone can download and use it. Also, he is currently working on a more user-friendly interface that will allow someone to draw the two molecules they want to react, copy and paste them into the program, and get predictions in minutes instead of hours, depending on the complexity of the molecules.

"I think it's really exciting to do something like that," Rinehart said. "We don't often publish a paper and put out a tool in the public domain that people can use in the field. People in academic labs like ours could use this tool and get an answer faster in their own research."

More information: N. Ian Rinehart et al, A machine-learning tool to predict substrate-adaptive conditions for Pd-catalyzed CN couplings, Science (2023). DOI: 10.1126/science.adg2114

Journal information: Science

Go here to see the original:
Machine learning tool simplifies one of the most widely used reactions in the pharmaceutical industry - Phys.org

In a recent event by the University of California, Berkeley, focused on Open Source vs. Closed Source: Will Open Source Win? Challenges & Future in Open Source LLM, experts in the field of machine learning shared their insights into the role of open source technology in shaping the future of this dynamic industry.

In the world of AI development, the acronym LLM typically stands for Language Model, more specifically, Large Language Models. These sophisticated AI models are designed with the purpose of understanding and generating human language.

Through rigorous training using vast amounts of data, they acquire the remarkable ability to tackle diverse tasks like natural language understanding, text generation, and translation, among others. An exemplary illustration of such language models is GPT-3 (Generative Pre-trained Transformer 3).

Recently, the use of AI has become a huge topic for discussion around the world. An integral part of the conversation is whether open source will be the future of machine learning as it relates to AI or a closed system.

Ion Stoica, Professor of Computer Science at the University of California, Berkeley, outlined three key reasons why open-source technology will play a pivotal role in the future of machine learning.

Firstly, he said that the current limitation in the availability of high-quality data for training machine learning models is a challenge for further development.

However, with the use of open-source systems, more quality data becomes accessible, and the cost of training models will decrease, making larger models more effective.

Secondly, Ion pointed out that machine-learning technology is becoming increasingly strategic for many countries.

Unlike search technology which still requires human intervention, machine learning models can make autonomous decisions, rendering them highly valuable for specific applications.

He also added that there is a need for fine-tuning machine learning models for particular tasks rather than creating general-purpose models.

He believes that experts should focus on developing models that excel in specific use cases. And an open-source model will make this even easier to implement.

Another speaker at the event, Nazneen Rajani, Research Lead at Hugging Face, said that open-source technology is essential for crafting smaller, more specialized language models tailored for specific use cases.

She revealed that most companies and consumers do not require large, general intelligence models; instead, they need models that excel at specific tasks.

The Researcher also expressed excitement about Metas entry into the open-source arena, anticipating increased funding and resources for open-source projects, paving the way for further innovation and development.

In support of the first two speakers, Tatsunori Hashimoto, an Assistant Professor at Stanford University, proposed that language models could become a public good and serve as a foundational layer for intelligent agents.

He cited initiatives like the UKs Brit-GPT, government-run language models available to everyone, as examples. Once these models are open and accessible, they can form the basis of an open-source innovation ecosystem.

Tatsunori also noted that the future of open source depends on who provides the base layer and how much innovation is generated atop it.

More here:
Open source in machine learning: experts weigh in on the future - CryptoTvplus