What is machine learning? Explaining in simple terms

What is machine learning?

Machine learning (ML) is one of the methods of artificial intelligence that solves the problem not in a direct way, but by searching for patterns in the data after training the algorithm on many examples.

Such algorithms can answer the question whether a fruit in a photo is a banana or an apple, identify people crossing the road in front of an unmanned vehicle, recognize spam in incoming emails, and generate subtitles for YouTube videos.

The key difference from traditional programming is that the developer does not write strict code to instruct the system to distinguish between a banana and an apple. Instead, he creates a model that learns to distinguish between fruits on a large amount of data. In this case, on a huge number of images of bananas and apples.

How is machine learning different from artificial intelligence?

Machine learning is one of the methods of artificial intelligence (AI).

What are the types of machine learning?

There are several types of machine learning. To date, the most popular are:

• training with a teacher;
• learning without a teacher;
• reinforcement learning.

What is unsupervised learning?

Unsupervised learning task algorithms try to identify similarities in input data and categorize them. As a rule, training of such models takes place without human intervention.

For example, Airbnb’s short-term rental algorithms cluster houses available for rent by neighborhood, and Google News aggregator creates collections of articles on similar topics every day.

Unsupervised learning algorithms are not designed to highlight certain types of data. They are simply looking for information that can be grouped by similarity, or to highlight anomalies.

What is reinforcement learning?

This method implies that AI agents will learn to interact with a certain environment on their own.

The easiest way to understand the essence of reinforcement learning is to think about a person playing a computer game for the first time, learning the rules in the process. As a result, by looking at the relationship between button presses, the result on the screen and the score, the performance of the gamer will increase from level to level.

In 2013, the DeepMind Lab developed a deep reinforcement learning algorithm that surpassed humans in a wide range of classic video games. The system takes pixels from each game, determines various information about its state (for example, the distance between objects on the screen), and considers how the control affects what is happening on the screen and correlates with the score scored.

During many cycles of the game, the system builds a model of what actions will maximize the score and receive rewards.

Another striking example of reinforcement learning is the AlphaGo algorithm developed by the same DeepMind. In 2016 the program beat professional go player Lee Sedol with a score of 4:1. At the same time, the algorithm did not calculate all possible options in advance.

Scientists have found out long before that the number of available combinations in this game is greater, than atoms in the observable universe. Instead, AlphaGo assessed the situation in the context of events and adjusted to changing conditions.

How are machine learning results evaluated?

After training, the model is evaluated using data that was not used during training.

About 60% of the dataset is typically used to develop an algorithm. Another 20% of the dataset is selected to validate predictions and adjust additional parameters that optimize the output of the model. This fine-tuning is intended to improve the model’s prediction accuracy when new data is introduced.

The remaining 20% ​​of the set is used to test the output of the trained and tuned model to check the accuracy of the predictions when presented with new information.

What is driving the popularity of machine learning?

Even though machine learning is not a new technique, interest in this field has exploded in recent years.

This was facilitated by a series of breakthrough innovations that saw the MOD set new accuracy records in areas such as natural speech processing and computer vision. The success was made possible due to two factors: a huge amount of data for training and the availability of huge parallel computing power with the help of modern GPUs.

In addition, entire cloud clusters for machine learning have appeared. Today, any user can use the services of companies like Amazon, Google and Microsoft to develop their own models.

As ML grows in popularity, tech giants are creating specialized hardware designed to run and train machine learning models. For example, Google is developing specialized tensor processors (TPUs) that speed up the learning process of algorithms.

In 2021, the company introduced the fourth version of the chip to improve Google’s cloud infrastructure. According to the developers, a cluster of 4096 TPUv4 will be able to provide the performance of more than one exaflops.

ML tasks are increasingly being performed on consumer-grade phones and computers, not just in cloud data centers. In 2017 Apple introduced iPhone X with A11 Bionic processor equipped with a special chip for calculating machine learning tasks. Every year the company improves the processor, making it possible to deploy demanding algorithms on mobile devices. Google is also trying its best to support the ML race on mobile devices. In the summer of 2021, the company introduced Android ML platform and added TensorFlow Lite to Play Services. According to the developers, thanks to this, the processing of algorithms on the device will provide lower latency, more efficient battery use, and features that do not require a network connection.

What is machine learning used for?

Today, machine learning systems are used everywhere and are the cornerstone of the modern Internet.

Each search query in Google launches several ML models at once: text recognition, personalization of results, and so on. The spam detection system in Gmail works the same way, identifying fraudulent messages.

Recommender systems in online stores can predict what product you’ll want to buy next or what movie you’ll like on Netflix.

One of the most prominent examples of the use of machine learning in everyday life are virtual assistants like Apple’s Siri, Amazon’s Alexa or Google Assistant. Each of them relies heavily on ML for voice recognition and natural language understanding, and also needs a huge information base to answer queries.

The systems are used in a variety of industries, including:

• computer vision for unmanned vehicles, drones and robot couriers;
• natural speech processing for chatbots and virtual assistants;
• face recognition;
• detection of tumors on x-rays;
• the ability to preventively maintain infrastructure by analyzing IoT sensor data.

And this is far from an exhaustive list.

Are machine learning models objective?

The quality and amount of data used to train systems affects the tasks for which they are suitable. Recently, there has been growing concern in the scientific community about how machine learning systems codify human biases and social inequalities reflected in training data.

2016 National Science Foundation Fellow from the University of Washington Linguistics Department Rachel Tatman discoveredthat Google’s speech recognition works better with male voices than female voices when automatically adding subtitles to YouTube videos. She attributed the study’s findings to “unbalanced training sets” dominated by talking males.

Facial recognition systems face difficulties in identifying women and people with dark skin tones. Questions about the ethics of using such potentially biased systems for policing have led major tech companies to temporarily suspend their sale to law enforcement.

In June 2020 Amazon banned US law enforcement to use facial recognition software in the midst of protests against police brutality. A year later, the company extended the moratorium for an indefinite period.

In 2018, Amazon also refused from a machine learning-based hiring tool that identified male candidates as preferred.

As machine learning systems move into new areas, such as helping diagnose diseases, the potential for systems to shift towards providing better services or treating certain groups of people more equitably becomes an increasing concern.

To date, research continues on ways to reduce bias in self-learning systems.