What to expect in the field of AI and machine learning in 2021?
Neural networks continued to develop this year: they learned to identify COVID-19 by coughing, found use in advertising, sang with Eminem’s voice. We talk about what to expect in the field of AI and machine learning in 2021.
Pathologists use AI to more accurately diagnose cancer. Data on various types of cancer are used to create a predictive model. For example, PathAI technology is used for this.
Creating new drugs is a time—consuming process, because you need to conduct a lot of tests to find a useful formula. This is helped by AI. Atomwise is one example of a technology that allows you to detect new molecules. It is used in the creation of new medicines for 27 diseases in collaboration with Harvard and Stanford Universities and pharmaceutical companies.
Optimized processing of patient data
The number of patients around the world is growing daily. Automated systems are required to process data on their diseases. AI allows medical institutions to optimize this treatment. For example, OLIVE is a platform for automating healthcare tasks.
Money is becoming digital. The total amount of digital transactions is $4.4 million, and could reach more than 8 million by 2024.
AI-processed data on these transactions could help improve the financial industry in 2021. For example, Dataminr collects information from various text sources and presents the user with a graph of important events. Neural networks allow it to analyze text data.
Machine learning algorithms can be used for automated trading. Theoretically, having information about prices, volumes, dates, public sentiment (or weather), it is possible to create a system that will beat the market (but it is not so easy). The algorithm can be trained and adapted to changes in real time to make predictions more accurately. For example, Kayrros is a company that analyzes data to help market participants invest.
Digital payments have certain risks. In 2018, $24.26 billion was lost due to fraud. Machine learning is perfect for dealing with them. The British company AimBrain, with the help of machine learning, prevents account theft and detects fraudsters’ accounts.
The model can use training data to label each operation (suspicious or not). Then, using accuracy and completeness metrics, we can adjust the model to our risk profile by analyzing the costs of false-positive and false-negative forecasts.
Banks use machine learning to serve customers, predict risks, prevent risks and invest. Let’s say you can offer personalized offers based on the user’s financial behavior. Thus, if a client is looking for a house, he can make a special offer. Envestnet is a financial data collection and analysis company that provides financial management services.
GAN (generative-adversarial networks)
Generative-adversarial networks is an approach to generative modeling using deep AI machine learning techniques such as CNN (Convolutional Neural Network).
GAN uses the model to generate new data that is similar to the ones it was trained on (for example, images).
Using GAN, you can get datasets of images, faces, cartoon characters, translate images into text and back, create 3D objects, etc. There are many applications of GAN, but they can bring not only benefits. One of the latest uses of GAN applications is deepfakes.
Terro Carras, in his article “Progressive development of GAN to improve quality, stability and diversity” demonstrates generated realistic images of human faces. The model is trained on celebrities, so she creates faces similar to existing ones.
But don’t be fooled by the wonders of data processing. We will soon face the consequences of such deepfakes. Nowadays, any person’s reputation can be destroyed with the help of technologies from public repositories.
Reinforcement learning (reinforcement learning)
RL is an area of AI machine learning in which it is studied how the test system (agent) interacts with some environment to obtain maximum reward. The reward is the response of this environment.
Reinforcement learning is one of the three machine learning paradigms, along with learning with a teacher and learning without a teacher.
RL is how we learn on a daily basis. Imagine that you took a puppy one month old. To train him, you need to use a reward system. If the dog listens, you give him a treat. This is how you use positive reinforcement to train your dog.
Programmers from OpenAI, a company founded by Elon Musk, showed how agents play hide-and-seek.
They were not given explicit instructions on how to play. After millions of simulations, agents have learned to interact with the environment:
- the one who hides has learned to cost small forts and barricades;
- the one who is looking has started using ramps to climb the walls and find those who are hiding.
AR and VR (Augmented/Virtual reality)
Augmented reality is a bridge between virtual and physical reality. Visual data collected by AR applications can be used for image recognition. Augmented reality and AI are different, but complementary technologies. Using both, you can create something beautiful.
The model is built from camera frames that can be used to classify the location. Here’s a great article about it.
The frames are transmitted to the model, which is able to estimate the size and location of objects on the scene. This data can be used to create a frame around the object. For example, AnyVision can help identify a person or object, even in a crowd.
Definition of posture
There are two types of pose definition:
- 2D — calculates the coordinates of each joint (x, y) on the RBG image;
- 3D — calculates coordinates (x, y, z) for each joint in the RGB image.
Posture detection is an important part of motion recognition, animation, game creation, etc.
Here’s a good blog about it.
Adapted translation of the article Machine Learning Trends to Watch Out in 2020 and 2021