What is machine learning?

What is machine learning?

Machine learning (ML) is a branch of artificial intelligence that addresses tasks not through explicit instructions but by finding patterns in data after training an algorithm on many examples.

Such algorithms can decide whether a fruit in a photo is a banana or an apple, detect pedestrians crossing in front of a self-driving car, filter spam in email inboxes and generate subtitles for YouTube videos.

The key difference from traditional programming is that a developer does not write rigid rules to tell a system how to distinguish a banana from an apple. Instead, they build a model that learns to tell fruits apart from a large volume of data—in this case, vast collections of images of bananas and apples.

How does machine learning differ from artificial intelligence?

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

Alongside it are other approaches for creating AI systems—for example, evolutionary algorithms that model natural selection, and expert systems, in which computers are programmed with rules that mimic a human expert in a specific domain—such as an aircraft’s autopilot.

What types of machine learning exist?

Several types of machine learning are commonly distinguished. Today the most popular are:

• supervised learning; 
• unsupervised learning;
• reinforcement learning.

What is unsupervised learning?

Unsupervised-learning algorithms look for similarities in inputs and split them into categories. As a rule, such models are trained without human intervention.

For example, Airbnb’s algorithms cluster homes available for rent by neighbourhood, and Google News compiles daily bundles of articles on similar topics.

Unsupervised algorithms are not designed to pick out a predetermined class. They simply search for information that can be grouped by similarity or flagged as anomalous.

What is reinforcement learning?

This approach trains AI agents to interact with an environment on their own.

The simplest way to grasp reinforcement learning is to imagine someone playing a video game for the first time, learning the rules as they go. By observing the relationship between button presses, on-screen results and the score, the player’s performance improves level by level.

In 2013 DeepMind developed a deep reinforcement-learning algorithm that outperformed humans across a range of classic video games. The system ingests pixels from each game, infers information about state (for example, distances between on-screen objects) and explores how control inputs affect the game and the score.

Over many play cycles, the system builds a model of which actions maximise the score and yield rewards.

Another prominent example of reinforcement learning is AlphaGo, developed by the same DeepMind. In 2016 the program defeated professional Go player Lee Sedol 4–1. The algorithm did not compute every possible continuation.

Researchers had long shown that the number of legal positions in the game is greater than atoms in the observable Universe. Instead, AlphaGo assessed context and adapted to changing conditions.

How are machine-learning models evaluated?

After training, a model is evaluated on data that were not used during training.

Roughly 60% of a dataset is used to develop the algorithm. A further 20% is set aside to validate predictions and tune auxiliary parameters that optimise the model’s outputs. This fine-tuning aims to increase accuracy when the model sees new data.

The remaining 20% is used to test the trained, tuned model to verify how accurate its predictions are on previously unseen information.

What is driving machine learning’s popularity?

Though not new, machine learning has seen a surge of interest in recent years.

A series of breakthroughs has pushed accuracy records in fields such as natural-language processing and computer vision. Two factors made this possible: vast training datasets and the availability of massive parallel compute via modern graphics processors.

Entire cloud clusters dedicated to ML have emerged. Today anyone can turn to the likes of Amazon, Google and Microsoft to build their own models.

As ML’s popularity has grown, tech giants have built specialised hardware for training and running models. Google, for example, has developed tensor processing units (TPUs) that accelerate training.

In 2021 the company unveiled the fourth generation of the chip to bolster Google’s cloud infrastructure. According to its designers, a cluster of 4,096 TPUv4s can deliver more than one exaflops of performance.

ML workloads are increasingly executed on consumer phones and PCs, not only in cloud data centres. In 2017 Apple introduced the iPhone X with an A11 Bionic processor featuring a dedicated ML chip. Each year the company has improved the processor, enabling demanding algorithms to run on mobile devices. Google likewise has pushed ML on mobile. In summer 2021 the company announced Android’s updateable ML platform and added TensorFlow Lite to Play services. According to the developers, on-device processing reduces latency, uses the battery more efficiently and enables features that do not require a network connection.

What is machine learning used for?

ML systems are now embedded everywhere and are a cornerstone of the modern internet.

Every Google search query triggers several ML models at once: text recognition, personalised ranking and more. Gmail’s spam filters likewise spot fraudulent messages.

Recommendation engines in online shops can predict what you might buy next, or which film you will like on Netflix.

Among the most visible everyday applications are virtual assistants such as Apple’s Siri, Amazon’s Alexa and Google Assistant. Each relies heavily on ML for speech recognition and natural-language understanding, and they require large knowledge bases to answer queries.

ML finds uses across many industries, including:

• computer vision for self-driving cars, drones and delivery robots;
• natural-language processing for chatbots and virtual assistants;
• face recognition; 
• tumour detection in X-ray images; 
• predictive maintenance of infrastructure by analysing data from Internet-of-Things sensors.

And that list is far from exhaustive.

Are machine-learning models objective?

The quality and quantity of training data shape what systems are good at. Researchers have grown increasingly concerned about how ML codifies human biases and social inequalities reflected in training data.

In 2016 Rachel Tatman, a National Science Foundation fellow in the University of Washington’s linguistics department, found that Google’s speech-recognition system worked better on male voices than on female ones when auto-generating YouTube captions. She linked the finding to “imbalanced training sets” dominated by male speakers.

Face-recognition systems struggle to identify women and people with darker skin tones. Questions over the ethics of deploying such potentially biased systems in policing led big tech firms to pause sales to law enforcement.

In June 2020 Amazon banned US police from using its facial-recognition software amid protests against police brutality. A year later the company extended the moratorium indefinitely.

In 2018 Amazon also abandoned a machine-learning-based hiring tool that favoured male candidates.

As ML moves into new areas, such as assisting disease diagnosis, the risk that systems deliver better services—or fairer treatment—to some groups than others is an ever more serious concern.

Research into ways of reducing bias in self-learning systems continues.