Today we will tell you how scientists have applied the LLM principles to study the DNA language, what methodology MIT researchers propose to combat overconfidence of language models, and where in Russia smart agricultural machinery to be tested. Also in this issue of #InfocusAI, there is a space for news about ultra-thin memory chips for on-device AI and new release dates for smart features from Apple.
AI-focused digest – News from the AI world
Issue 47, July 25 – August 8, 2024
Scientists have developed a new AI model for studying the DNA language
Scientists from Germany have developed a new AI model for studying the human genome, which treats DNA as a language, that analyzes its rules and context in order to extract new knowledge. This new DLM (DNA Language Model) called GROVER can predict DNA sequences and extract contextual information, for example, identify gene promoters (this is a piece of DNA that RNA polymerase recognizes as a launching pad for transcription, i.e. RNA synthesis) and protein binding sites. It can also be used to study epigenetic processes that happen on the DNA and are not associated with any change in its code. To train this model, scientists needed to create a special DNA dictionary, which, as noted, distinguishes GROVER from previous models of this type. They analyzed the genome in several hundred cycles looking for combinations of “letters” (i.e. nucleotides A, G, C, T) that occur most often and fragmenting the DNA into “words”. Scientists believe that studying DNA using language models will help the development of genomics and personalized medicine. For more information about this development, refer to Nature Machine Intelligence or go to the website of the Dresden University of Technology.
MIT Is looking for a more effective way to prevent language models from being overconfident
A well-calibrated artificial intelligence model should be less confident about a wrong prediction and more confident about the right one. However, conventional approaches to AI calibration are not always effective for LLMs, according to researchers from MIT and MIT-IBM Watson AI Lab. They are working on an alternative way to calibrate language models called Thermometer, which is based on temperature scaling. In fact, scientists propose to add a smaller auxiliary model to the main model and train it to predict the “temperature” to calibrate the main model for performing new tasks. To train this smaller model, they use labeled datasets for several representative tasks, and after training, it can be generalized to new tasks in a similar category without the need for additional labeled data. Ideally, of course, scientists want to ensure that after training their thermometer model can work with different types of tasks, but so far it cannot. Nevertheless, according to researchers, the proposed method produces more calibrated uncertainty measures while calling for lower computational costs. To learn more, go to the MIT News website.
Samsung starts to mass produce memory chips for on-device AI
Samsung Electronics has started mass production of the industry’s thinnest low-power dynamic random-access memory (DRAM) chips. VentureBeat writes that this is how the South Korean tech giant seeks to meet the growing demand for on-device AI. We are talking about 12-nanometer class memory chips with a thickness of a fingernail for 12 and 16 GB, which do not take up much space in digital devices. This contributes to better airflow, which is especially important to support high-performance applications with advanced features such as AI, the article explains.
Apple postpones the launch of new AI-based features in its products until October
Last week, the mood of iPhone owners was slightly spoiled by the news that Apple postponed the launch of new AI features until about October. This is reported by Bloomberg in particular, citing unnamed sources, emphasizing that this way the company will have more time to fix bugs. Initially, the mass media wrote that the AI “pumping” of the iPhone and iPad software was going to start in September with the release of iOS 18, iPadOS 18, and macOS Sequoia. Well, we will wait a little longer. In the meantime, here you can find out some technical details about Apple Intelligence.
Testing site for agricultural AI to appear in the Lipetsk region
A site for testing agricultural machinery with artificial intelligence will be opened on the basis of the Yelets College of Innovative Technologies in the Lipetsk region in September this year. It will also be used to train operators of smart machines. This is reported on the official website of the regional government. Maintenance of agricultural machinery with artificial intelligence will become a mandatory discipline for future tractor drivers. This module will be included in the curriculum for first-year students, and additional seminars and workshops will be organized for senior students.
