I’m back on Medium. What happened with AI during 2020 up to now?
Breakthroughs that shaped AI in 2020.
Hello, reader. I am Oleksii, and I have been involved in the world of AI and Data Science for over 4 years. I am just an entrepreneur with many ideas and a big desire to spread knowledge and make things simpler.
A long time ago in a galaxy far, far away… I was already here, leading my blog on Medium, where I covered all of my insights and thoughts on AI, ML, Data Science. But one day, I discovered my blog disappeared from the surface of Medium, and mainly Towards Data Science, where all of my articles appeared very often.
With more than 4,2k subscribers and nearly 50 articles, my blog sank into oblivion for unknown reasons. But, I decided to start everything over again. Just like always, you can expect here my thoughts on the latest innovations in AI, non-boring guidance, complex math behind everything in plain language, and more.
So, without further ado, let’s jump right in!
What shaped AI in 2020?
Let’s start with the fact that 2020 was undoubtedly a complicated year since it was marked by such a black swan event as the Covid-19 pandemic. It was hardly impossible for AI to not overlap with it. The crisis required new solutions, and actually, AI managed to provide them causing a serious technology shift towards new demands dictated by a coronavirus.
Both small companies and tech giants were forced to move their focus on AI. I think it is worthwhile to mention several projects that I find valuable and useful:
- Microsoft spent more than $25 million to launch AI for Health.
- SenseTime has developed software for contactless temperature screening.
- Alibaba presented its AI-driven system that detects coronavirus in seconds with 96% accuracy. For these purposes, computer tomography is used, which records all the parameters of the human body and performs rapid data analysis.
- Ardi AI platform presented by Graphen jointly with Columbia university can determine the canonical form of localization of each gene of the virus.
Despite amendments made by the Covid-19, АІ didn’t go astray from its own enhancements. During 2020, we got improvements on unsupervised and self-supervised learning in the field of deep learning. Besides, there was an advance in low-code automated machine learning platforms (like AutoML, AutoML-Zero), making machine learning algorithms even smarter.
So, let’s take a closer look at some of these and other breakthroughs in more detail. Below I have gathered the most prominent innovations from my point of view. I had only one criterion when making this list — the practical value (and for some cases like #6 important concerns that contradict human rights).
#1 GPT-3 from OpenAI
An innovation that makes a lot of noise over the previous year is undoubtedly GPT-3. The third version of Generative Pre-trained Transformer crafted by OpenAI can produce human-like text better than anything else. Its power is that it was trained with 175 bn parameters — the biggest number of parameters compared to all previous autoregressive language models that use deep learning.
GPT-3 can certainly be called a huge leap forward to supercharge Natural Language Processing. All this because of a new convenient paradigm — “pre-training + fine-tuning.” In its nutshell, it is an autocomplete program that can act as a translator, a programmer, a poet, or a famous author, and it can do it with its user (you) providing fewer than 10 training examples. Just type, it will predict what comes next. Amazing, isn’t it?
But, I think soon we will be on the threshold of an even greater breakthrough, namely GPT-4. As more people experiment with GPT-3, we’ll find it falls a little short of most practical usefulness sooner or later. I think GPT-4 will fill all the gaps of its predecessor and so will be a much more powerful tool for auto-text-generation tasks.
Sources & Research papers to go deeper:
#2 MLOps
MLOps (machine learning operations) or AIOps is another breakthrough that I think pushed AI to the next level in 2020. And although this concept is still new, every day it becomes even more in demand. All this because putting MLOps in practice helps to avoid common pitfalls and problems that Data Scientists face every day.
Like its DevOps approach, MLops can cover a variety of tasks. More precisely, it extends the CRISP-DM methodology with an Agile approach and technical tools to automate operations with data, ML models, code, and environment. These tools include, for example, Cloudera Data Science Workbench.
MLOps automates development, integration, testing and turns deployment processes into a single and effective pipeline. Behind this magic-automation is a set of best practices on operationalizing and managing lab-trained machine learning models.
Sources & Research papers to go deeper:
- MLOps Core
- Kaz Sato — ML Ops Best Practices on Google Cloud (Cloud Next ‘19)
- Kubeflow — The Machine Learning Toolkit for Kubernetes
#3 AutoML — AI creating AI?
Well, 2020 can surely be called a year of improvement of automated machine learning, which is AutoML. Projects like Explainable AI have shown great results through deriving as much insight as possible from data minimizing human intervention. All this because AutoML enables developers to ‘visually investigate model behavior’ in machine learning models and uncover the complexity of machine learning models and gain better insights.
What’s the technology behind AutoML? It is an algorithm that autonomously builds the best machine learning model for a given problem. It enables custom ML models at scale, generating billions of predictions without needing an army of data scientists. That’s why many say it will lead to the death of data scientists which I think is total nonsense and just loud praises. I can explain why.
AutoML is just one piece of the puzzle for implementing AI, which can’t handle all automation processes. For instance, it doesn’t AutoSelect a business problem to solve, it doesn’t AutoSelect indicative data, it doesn’t AutoAlign stakeholders, it doesn’t provide AutoEthics in the face of potential bias, it doesn’t provide AutoIntegration with the rest of your product, and it doesn’t provide AutoMarketing after the fact.
Sources & Research papers to go deeper:
- AutoML-Zero: Evolving Code that Learns
- AutoML-Zero: Evolving Machine Learning Algorithms From Scratch
- Addressing Drawbacks Of AutoML With AutoML-Zero
#4 AI solved one of the greatest challenges in biology
Another loud event for AI in 2020 is new development presented by DeepMind’s AlphaFold. More precisely, AlphaFold managed to solve one of the greatest challenges in medicine — the Protein Folding Problem. The essence of the problem is that it was impossible to determine what structure will make each of the over 400,000 proteins. But, finally, AlphaFold managed to do it.
Why is this breakthrough so important for medicine? This helps to understand all the problems associated with the processes of protein folding and unfolding. And this, in turn, leads to better treatment of diseases and producing more effective medicines and vaccines.
Here are two examples of protein targets in the free modeling category. AlphaFold predicts highly accurate structures measured against experimental results:
Sources & Research papers to go deeper:
#5 AI can feel pain and self-heal
The first thing I want to start with is that now robots can mimic human neurological functions. I find this innovation truly impressive because it is like a starting point towards making robots experience human feelings (or at least simulate them), give them human-like traits and so maybe it forms the next generation of robots to interact effectively with humans.
Singapore-based scientists from Nanyang Technological University initiate the first attempt to force robots to feel something. More specifically, they have developed AI mini-brain for robots and equipped them with sensor nodes on their skin so they can feel when some damage happens to them and then respond to harmful effects with self-healing procedures thanks to self-healing ion gel material.
No need for human intervention. AI algorithms work that way to process the ‘pain’ by calculating the force of impact or pressure. This power enables the robot to understand where exactly it has sustained damage and if the damage is slight, begin repairing itself without human intervention.
Sources & Research papers to go deeper:
#6 AI Automates Mass Surveillance
One of the most questionable advances this year is the growing use of AI in mass monitoring. For instance, companies such as Clearview and Amazon revealed advanced facial recognition technologies for mass monitoring to identify and track individuals, which can mean solving more crimes. It is obviously a significant breakthrough from the technical side, but when it goes to privacy and the possible erosion of civil liberties, it falls short behind. Let me explain why.
The industry is largely uncontrolled, which means that facial recognition technology can also be misused and lead to innocent people’s arrest based on incorrect facial recognition matches. Another concern, especially in the United States, is that the watch lists that police use to check images against can be enormous — and can include people without their knowledge. I think we should certainly be aware of such things and I hope that soon we will develop some projects or solutions that will regulate all concerns.
Sources & Research papers to go deeper:
#7 AI Solves Quantum Chemistry
One more achievement is a successful attempt to solve Schrodinger’s Equation. A team of scientists at Freie Universität Berlin a method for probing the complex world of quantum chemistry. The value of this new method is that it makes it possible to find an exact solution for arbitrary molecules that can be efficiently computed.
Sources & Research papers to go deeper:
- Deep-neural-network solution of the electronic Schrödinger equation
- Artificial intelligence solves Schrödinger’s equation
Bottom Line: Let’s talk science!
Image source: Unsplash
Well, 2020 was certainly a very unusual and difficult year. But even despite all the difficulties, AI progress doesn’t halt to make machine learning algorithms smarter.
From medicine to surveillance, every area comes up with change. To sum up all the innovations, some of the common themes were AutoML, MLOps, Open GPT-3, DeepMind’s AlphaFold, AI mini-brain that feels pain, and more. The only thing that upsets me is of course Clearview’s controversial facial recognition, but I hope they will solve all the issues with human rights soon.
I hope that key developments rewinded in this article will help you to keep an eye on technological trends. And yes, I believe that the rest of 2021 will be no less good for AI. What do you think? Leave your suggestions or questions, and we will discuss everything.
P.S.
If you somehow reach my new profile again (or for the first time), please support my efforts and follow me to read more dope articles later. In case you enjoy some of my previous articles and you want to use them again, visit my Linkedin (hopefully, everything is totally safe there).
By the way, here are some of my previous guides (for total beginners) which come in handy for many:
- Deep Learning Algorithms — The Complete Guide
- A Gentle Intro to Probability and Statistics for Data Science
- Who Is a Data Engineer & How to Become a Data Engineer?
- A Beginner’s Guide to Python for Data Science
- A Beginner’s Guide To Data Science
Welcome to join Magic people, AI people community.
Thank you for reading!
