The state of AI in 2020: discoveries you should know about
Discover some of the most exciting breakthroughs in the field of Artificial Intelligence for the year 2020
OpenAI’s new language generator GPT-3
OpenAI first described GPT-3 in a research paper published in May 2020. It is the largest and most powerful language model ever created which can generate human-like texts on demand. Its forerunner GPT-2, was already able to generate different styles of convincing text when prompted with an opening sentence. But GPT-3 is a big step forward. The model has 175 billion parameters compared to GPT-2’s 1.5 billion. These parameters are the values that a neural network tries to optimize during training.
Arram Sabeti published a blog post showing short songs, stories, press releases, and technical manuals that he generated using artificial intelligence. Pretty impressive! Others have found that GPT-3 can generate any kind of text, including computer code, guitar tabs, and pastiches of particular writers. But despite its new advancements, the system still has some way to go. GPT-3 still has some limitations. The co-founder of OpenAI Sam Altman tweeted “...AI is going to change the world, but GPT-3 is just a very early glimpse. We have a lot still to figure out.”
Facebook’s new AI language model can translate between 100 languages without relying on English
For years, AI researchers have been working on building a single universal model that can understand all languages across different tasks. Now, Facebook has developed its AI model called M2M-100 that can translate between any pair of 100 languages. Of the possible 4,450 language combinations, it translates 1,100 of them directly. This is different from the previous multilingual models, which heavily rely on English as an intermediate language. For example, a translation of Chinese to French language typically passes from Chinese to English, and then English to French, which increases the chance of errors. The AI model was trained on 7.5 billion sentence pairs. In order to be able to train on a dataset that large, the researchers relied heavily on automated data curation. This milestone is a result of years of Facebook’s AI work in machine translation. Breaking language barriers through machine translation (MT) is a great way to bring people together, provide authoritative information for example on COVID-19, and keep them safe from harmful content and hate speech.
AI can detect COVID-19 from the sound of your cough
AI has been very helpful in fighting the COVID-19 pandemic, by improving the screening, tracking, and predicting how COVID-19 will develop in current and future patients. So far, AI has been beneficial for early detection and diagnosis of COVID-19 infections, for development of drugs and vaccines, and for the reduction of workload for healthcare workers.
Perhaps the most fascinating example is the use of artificial intelligence for early detection of COVID-19, more specifically using AI to detect asymptomatic COVID-19 infections through cellphone-recorded coughs. MIT researchers have found out that people who are asymptomatic may differ from healthy individuals in the way that they cough. The difference in cough is not noticeable with the human ear, but turns out COVID-19 coughs can be picked up by artificial intelligence. “The effective implementation of this group diagnostic tool could diminish the spread of the pandemic if everyone uses it before going to a classroom, a factory, or a restaurant.” says Brian Subirana, a research scientist at MIT’s Auto-ID Laboratory.
Scientists have used AI to discover promising drug-like compounds
In September, a team of researchers at Hong Kong-based Insilico Medicine and the University of Toronto used Artificial Intelligence to create 30,000 designs for molecules that target a protein linked with fibrosis (tissue scarring). It took them 21 days to create the molecule designs, and 46 days to identify a potential new drug. They selected six to synthesize and test, while one was particularly active and proved promising in animal tests.
The AI model works in a way such that it examines previous research and patents for molecules known to work against the drug target, and prioritize new structures that could be synthesized in the lab.
Kantify is also active in applying AI to drug discovery. We have developed a novel AI algorithm that can perform Virtual High-Throughput Screening and help identify more promising drug candidates. Discover more in our use case on Virtual High-Throughput Screening with AI.
AlphaFold can accelerate drug discovery by predicting how proteins fold
DeepMind is also among the companies that is working on accelerating drug discovery with AI by predicting how proteins fold. The company has developed AlphaFold - an AI algorithm to help predict the structure of proteins from their genetic sequence. AlphaFold can model the complex folding patterns of long chains of amino acids, based on their chemical interactions, and form a 3D shape of a protein. Being able to predict a protein’s structure, can allow scientists to synthesize new protein-based drugs to treat new diseases, or to break down pollutants in the environment.
Tech giants and academic researchers have been working on a new algorithm to shrink the overall size of AI algorithms. Tiny AI is especially beneficial for AI algorithms that require large computational power and datasets to work. The new advances are just starting to become available to consumers, particularly through tech giants like Google, Apple, and Amazon. Google announced that it can now run Google Assistant on users’ phones, and Siri’s speech recognition capabilities are now run locally, without sending requests to a remote server.
The benefits of tiny AI can be enormous, particularly in the field of personalized medicine. Improvements in data usage and algorithms can lead to accessible personalized healthcare through our phones, small wearable devices, and more. For example, mobile-based medical image analysis, using wearables to predict an occurrence of atrial fibrillation episode.
Tiny AI may also improve the reaction times for self-driving cars, improve data privacy, decrease environmental impact, and more.