Machine learning

The concept of machine learning (ML) was first coined by Arthur L. Samuel in 1959 when he programmed a computer to defeat an experienced human player in the game of checkers.  

Today, ML has become a cornerstone of artificial intelligence (AI) research. It’s connected to various daily functions, including algorithmic recommendations, image recognition, and self-driving cars.

What is machine learning?

Fundamentally, ML is precisely what it sounds like—a machine's ability to imitate how humans learn while improving in speed and accuracy over time. Machine learning is algorithmic, meaning a computer is only given data, not commands. 

As a result, ML allows computers to make decisions based on data without being explicitly programmed what to do.

For instance, the software we use for language translations uses natural language processing (NLP), a type of machine learning. Similarly, when Netflix or Spotify recommend a new movie or song to check out, this is also a feature of ML. 

In this sense, machine learning is a general term that describes different functions and programming features that allow computers to process data to "learn" and incrementally improve their functionality. 

Machine learning vs. neural networks vs. deep learning

Machine learning, neural networks, and deep learning are all part of the larger artificial intelligence field. As a result, it's common for these terms to be used interchangeably. 

However, some features differentiate them, with machine learning as the umbrella term above the others. 

Simply put, a neural network is an algorithmic program functioning as a subset of machine learning. By extension, deep learning is a feature of neural networks that allows computers to emulate the decision-making process modeled after the human brain. 

How does machine learning work?

The "learning" in machine learning is best understood when broken down into three processes:

Decision-making

At this stage, ML algorithms help computers classify and analyze data to identify common patterns. These patterns help the computer "learn" what they mean and "decide" what to do. 

For instance, AlphaZero is a neural network developed by Deep Mind, a subsidiary of Google,  programmed to use machine learning principles to identify patterns and learn how to play chess. 

The only data programmers provided AlphaZero were the rules of the game. Afterward, they left the computer to play thousands of games against itself, analyzing the data to slowly—or quickly, by human standards—understand and excel at chess. 

Error-function

On the heels of decision-making is the error functioning stage. This is when computers evaluate the quality of their predictions and analysis. For example, suppose an algorithm provides three new music recommendations to a user on Spotify. Out of the three, the user begins listening to two and ignores the third.

This is an opportunity for a machine to "learn" via error-functioning, incorporating the new data to enhance the quality and accuracy of future recommendations. 

Model optimization

Finally, there's model optimization, when ML's quality, efficiency, speed, and accuracy are all improved through fine-tuning. As more data is analyzed and additional patterns become apparent, various fine-tuning algorithms help evaluate and optimize the process across different metrics. 

This process is repeated as the machine makes incremental adjustments until it reaches a particular accuracy threshold.

The three methods of machine learning

The different methods of machine learning relate directly to one factor: supervision. 

1. Supervised ML

Machine learning is supervised when the sets of data programmers provide for training are pre-labeled and classified. Human programmers cross-validate the results as a machine analyzes data and adjusts the weights in its algorithm. As a result, this is the most accurate, albeit slowest, machine learning method. 

2. Unsupervised ML

Unsupervised ML is what most people think of when they hear machine learning. With an unsupervised method, vast amounts of unlabeled data are clustered and analyzed for patterns without human intervention. 

3. Semi-supervised ML

The happy medium between the two is semi-supervised machine learning. During training, this method uses smaller sets of labeled data to "guide" the machine's learning process as it extracts and classifies data from a more extensive, unclassified set. 

Machine learning applications

There are various applications for machine learning, and advancements in artificial intelligence research are bound to create even more. Currently, machine learning is fundamental to:

  • Human training
  • Fraud detection
  • Self-driving cars
  • Image recognition 
  • Speech recognition
  • Language translation
  • Healthcare diagnostics
  • Social media monitoring
  • Recommendation systems
  • Natural language processing