In the new issue of #InfocusAI we will tell you how AI helped to find an antibiotic combating the drug-resistant Acinetobacter baumannii bacteria, what mix of ML techniques US scientists have come up with to improve performance of ML-apps, and how Japan is going to catch up with the far-advanced generative AI train. You will also learn where and how ML is used in Apple products and why the European Commission wants to make it compulsory for technology companies to label content created with neural networks.
AI-focused digest – News from the AI world
Issue 19, May 18 – June 8, 2023
US scientists create new method for ML training that replicates cognitive process in human
Researchers from MIT and other USA institutions and laboratories have developed a new method for analysing unlabelled audio and visual data that could improve the performance of speech recognition and object detection ML-applications. The contrastive audio-visual masked autoencoder or CAV-MAE, suggested by them, combines two self-supervised learning techniques – contrastive learning and masked data modelling, thereby imitating how humans understand and perceive the world. The study authors clarify that a larger portion of human knowledge is learned in a self-supervised way, without a supervisor, and they want to enable the ML model to have the same ability. Contrastive learning and masked data modelling are used as complementary methods, improving computation results. For example, masked data modelling cannot capture the association between the video and audio pair, but contrastive learning can, and vice versa – masked data modelling compensates for losses in contrastive learning. Read more about the approach and its testing results on the MIT News website.
AI helps to find an antibiotic against superbug Acinetobacter baumannii
Researchers from McMaster University (Canada) and MIT used artificial intelligence to identify an antibiotic that can kill Acinetobacter baumannii, a superbug that can cause serious infections such as pneumonia and meningitis, MIT News reports. Acinetobacter baumannii is classified as a critical threat to human health and is notorious for its resistance to most current antibiotics. Moreover, this microbe can also cause drug-resistance in other bacteria. The researchers used a machine learning technique, which had previously proved effectiveness in similar tasks, for finding a new antibiotic against Acinetobacter baumannii. To obtain training data, Acinetobacter baumannii was exposed to about 7,500 chemical compounds in laboratory conditions and the effects of each compound on bacterial growth were recorded. The ML model was trained on this database to identify chemical features of drugs associated with growth inhibition. Then, the researchers used the model to analyse over 6,000 chemical compounds it had not seen before, and, based on the computational results, chose 240 of these that were best suited for laboratory testing. One compound was found to be extremely effective at killing Acinetobacter baumannii, although it originally was explored as a potential diabetes drug. It’s important to note that the found antibiotic has a narrow spectrum ability, i.e., it doesn’t affect other bacteria. This is a desirable feature because it minimizes the risk of bacteria rapidly spreading resistance against the drug. Read about the details of the method for finding a new drug in the Nature Chemical Biology journal.
A new ChatGPT equivalent to be developed in Japan
The Tokyo Institute of Technology and Fujitsu plan to create a large language model focusing on understanding Japanese by the end of 2023, and make it available and free of charge for domestic companies and universities by 2024. The LLM will be trained on Japanese-language news websites, blogs and online encyclopedia. RIKEN, a research institute and the operator of the Japanese supercomputer Fugaku, and several other universities will be involved in the project. The collaboration is motivated by the desire to reduce dependence on foreign generative technologies that are mostly trained on English text data and are not precise enough in Japanese, and to increase Japan’s expertise, experience and knowledge in AI. This is expected to push the development of Japanese LLM-based intellectual products forward and upward and increase the country’s competitive ability internationally. More on the JapanForward website.
Apple presented new features of their product with ML under the hood
Judging only by artificial intelligence news, it may seem that Apple is standing aside in the global AI race. However, the latest conference of developers from Cupertino showed that this is not quite the case. This week, the company (finely and minimalistic, as always) announced a range of new features in its products with AI under the hood without any hype on the technology itself, instead talking more about the functionality. In particular, an improved autocorrect feature, based on the same technology as ChatGPT, was introduced. The Apple language model will learn from how the user texts to suggest better correction options. ML models are also featured in AirPods Pro that automatically turns off noise cancelling when the user engages in conversation. The most sensational product showcased at the conference was the Apple Digital Persona. The service makes a 3D scan of the user’s face and body and then can recreate what they look like during a videoconference while wearing the Vision Pro augmented reality headset. More about Apple’s policy of AI you can read in this article on CNBC.
EU suggests flagging AI content
The European Union intends to oblige technology companies to flag neural network generated content as a measure to tackle disinformation. The news was reported on Monday by all global media with reference to the European Commission vice president, Vera Jourova. For example, you can read about it in Bloomberg. For now, the initiative is voluntary, but there is a good chance that flagging will become mandatory for all generative services, social media and other platforms distributing audio, visual and textual content. This clause could be included in the Artificial Intelligence Act currently being prepared by the EU. Apart from flagging materials, the EU insists that technology companies take action against the use of generative services to create fake content. This is something the global AI development community will have to seriously address…