in

Artificial intelligence, Protein, Human, Biology State of AI 2022 – My Highlights

This years report came out around two weeks ago and you can find the link to it in the video description I really enjoyed reading it and therefore wanted to share some of my highlights in this video. In case you havent seen a report, its a Google Docs presentation, divided into four categories, which are research, industry, politics and safety. Of course, there are many great things that have been going on and I can highly recommend reading it for the following Ive picked out. My overall highlights from the categories: research, industry and a little bit of politics. My first highlight is generally the progress in the field of life scientists mainly in biology and chemistry, Im, not an expert in this field, but will give you a short introduction to better understand the following. Most of you probably remember, DNA from school, its a double helix made up of four chemical bases and encodes. The information of human life. Now genes are specific regions or units of the DNA that carry information of certain characteristics like eye color. The human genome has around 20 to 25 000 of these genes. All human cells carry the whole DNA information and a small subset of it is used to build cell specific functions. Most of the genes carry information to build proteins which end up in human cells and execute a certain functionality, for example enzymes or hormones, and this is also called gene expression. All of the previous mainly belongs to the field of biology.

Now lets also connect it with chemistry. The treatment of human diseases is dominated by small molecule drugs, which are the core component in chemistry. These components can affect the functionality of human cells in various ways, for example, they serve as enzyme Inhibitors and because of that drug design often aims to Target proteins in cells. This interaction is typically called protein. Ligand interaction so thats a very high level overview in layman terms, but hopefully helps to create some context for the following contents. With this little background, knowledge on biology and chemistry lets have a look at four interesting AI advances. In 2022, deepminds researchers working on Alpha Fall 2 have published the predicted structures for over 200 million proteins for the ones who are not familiar with it. Its an AI system that is able to predict the 3D structure of a protein just from its 1D amino acid sequence, and once you know the structure of proteins, you can also better understand their function and how they work. So, overall, this is a big step towards the better understanding of biology and diseases and helps to design suitable treatments. There was also a lot of progress in applying language models, or rather large language models in short, llms to protein sequences, for example, Pro Gen 2 showed state of the art performance on all sorts of protein related tasks. They essentially talk about model scaling and show that increasing model sizes produce better results. Thats a phenomenon thats also mentioned in several places in the state of AI report, increasing language model sizes tend to perform better end of March.

This year, open sale was published, which uses machine learning to improve the understanding of how proteins are organized within human cells. As a first step, a data set of more than 1 000 protein locations was open sourced and if this is extended to the full Human Genome, its possible to explore the entire Gene function and interaction in human cells, the researchers have also created an interactive website. In case you want to have a closer look. Finally, there are also some insights from 2022 that signalized, that not all of the advances will benefit humans. Last year, the drug Discovery platform megasin was published, which helps to find better drugs. In the paper, they propose different generative, deep learning models like lstms scans and baes, which are used to generate desired molecules. A core component of this model is a special score function, which includes, among others, attributes like drug likeliness toxicity and how good it binds to specific receptors. Of course, this model is optimized in such a way that it generates drug likely molecules which are not toxic. Then researchers were asked by the Swiss government what would happen if their model was programmed? The other way around the results were frightening, as the model produced 40 000 highly toxic molecules, among which also the nerve agent VX, an extremely toxic synthetic compound used in chemical warfare. Further interesting results are published in a paper called dual use of AI power, drug discovery, so overall, very interesting and important advances in the field of healthcare and life sciences and now lets move on to my next highlight, even though the fusion models were already part of Last years report they again made it into the list because of the amazing applications and successes throughout the year.

There was Delhi 2 in April, which fascinated with state of the art text to image generation results and then Googles, Imogen came out in May this year and finally, stable diffusion, free, open source, pre trained model for text to image generation was released in August this year. In addition to that, the successes of the fusion models were also shown on other data modalities like molecules or audio, but it didnt end here just last month. We could also witness amazing results in text to video generation, such as meta AIS, make a video, but also many others, overall, plenty of reasons for putting diffusion models on my list of highlights and Im, especially excited to see many more future applications a rather funny part Of the reports, industry section was the slide with a title. Attention is all you need to build your AI startup. It turned out that almost all researchers from the attention is all you need. Paper are now part of an AI startup with pretty good fundings, so it seems like the paper did not only disrupt AI research but also the careers of the researchers. Another thing I found quite interesting is the early warning system that predicted high risk covet variants. The vaccine leader bionetech and the AI company insta deep collaborated to identify these variants as early as possible. The result was that all 16 variants were predicted more than one and a half months prior to officially receiving the designation by The Who.

The underlying model is a large pre trained protein language model. Many of you have probably already played around with medical imaging data sets, for example, predicting skin cancer. Using a CNN Ive always wondered: are these systems actually used in practice, and if how automated are they meaning do we still have humans that check every prediction? The startup occipit received the first autonomous certification for their diagnostic system. Its task is to detect the normal chest x rays, which means Radiologists, dont have to look at them anymore. The report also states that there is a general shortage of Radiologists and increasing volume of Imaging. Therefore, such a system helps to assess more x rays. In shorter time, and especially if forward, only the relevant ones to Radiologists, the reported sensitivity of the AI system is at around 99.8 percent, meaning that there are very few false negatives at first. I was surprised that this system was approved, even though it doesnt hit 100 percent, but then I thought Radiologists probably also wont hit 100 and therefore the predictive quality is acceptable. Overall, the system was able to filter out almost 40 percent of x rays, reliably with almost no compromise on patient safety. I also enjoyed reading about AI startups and scale ups, and the report gave me a good overview on the current unicorns in the startup scene. Popular names are data breaks which provide an AI platform or waymo researching on autonomous driving. Overall, there is insane growth in AI companies, and it seems that there are many valuable, real world applications of AI.

The last point on my list is AI and defense technology. There are several new startups that received heavy funding by big tech companies like Amazon, Microsoft and Google. Defense technology includes military drones, surveillance systems and underwater autonomous vehicles, so overall, very important things to ensure the security. These were my highlights from this years. State of AI report, of course, there are many other interesting topics and I highly recommend having a look at it. Id also like to thank the authors for providing such a beautiful overview. I hope you found this collection as interesting as I did and hope to see you again in the future.

What do you think?

Written by freotech

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *

Loading…

0

Android, Microsoft Corporation, Microsoft Office, Microsoft Word, Microsoft PowerPoint Get Free MS OFFICE Without SUBSCRIPTION!! Genuine For Windows and Mac

Artificial intelligence, Protein, Human, Biology What If Moon Is A Spaceship Of The First Civilizations On Earth? | UFO News, UFO Sightings, Moonfall