What is machine learning?

Machine learning is an increasingly common computer technology that allows algorithms to use large data sets to analyze, classify, and predict. Machine learning is not as complex and powerful as related technologies, but it has a variety of uses and is used by many large companies around the world.

Machine learning aims to help computers learn in a way similar to how the human brain learns. Machine learning uses large data sets and algorithms (models) to analyze and classify data or make predictions. The more a machine learning model is used and the more data it processes, the better it performs its task. Models can improve themselves or be updated by humans.

Unlike similar techniques such as deep learning, machine learning does not use neural networks. While machine learning is related to developments such as artificial intelligence, it is neither as advanced nor as powerful as these technologies.

Machine learning has existed in various forms since the 1960s and is increasingly used. About 70% of financial services companies use some form of machine learning to some extent.

Machine learning starts with two elements: an algorithm and a data set. The algorithm instructs the ML model what to do (analyze images, detect patterns, make predictions). Data sets may or may not be classified or labeled to aid the algorithm. The algorithm then processes the data to produce an output.

The more data an algorithm processes, the more accurate it should become. Typically, models improve based on built-in logic, but humans can also update the algorithm or make other changes to improve the quality of the output.

This is what "learning" means. Humans learn basic concepts or skills and then improve through repetition and extrapolation. This is also the goal of machine learning. Traditional computer programs are designed to perform given functions, but these functions are relatively limited and can only be changed when the programmer changes them. With machine learning, the model is designed to change itself based on more data and experience with the task.

For example, an image detection algorithm might analyze pictures that contain redheads. The first time you use the model, its output will be less accurate than the second time, and the third time it will be more accurate. This improvement occurred as the model developed better techniques for distinguishing humans from trees or cattle and distinguishing redheads from blondes.

While there are subtypes within each category, the four main types of machine learning are:

  • Supervised machine learning: It uses labeled structured data and minimal human intervention to find patterns sought by the model creator. Supervised learning works best for relatively simple tasks that can be automated, have rules that are easy to define and understand, and for which there is a large amount of data available.
  • Unsupervised machine learning: Unlike supervised machine learning, there is no labeled or structured data available. In contrast, unsupervised machine learning models detect patterns and draw conclusions based on the data, including things the model creator was not looking for. It is used to detect patterns and classify (such as grouping customers based on behavior) and then take action based on those patterns.
  • Semi-supervised machine learning: combines the above two types by first training the model using labeled data and then letting the model process unlabeled data. Semi-supervised machine learning can be helpful when there is not enough labeled data or it is impractical to generate it.
  • Reinforcement: This type of machine learning is based on rewards or positive feedback and is best suited for systems where right/wrong answers can be easily defined along with the best action given. Machine learning models for playing games like chess are often reinforcement models.

Applications of machine learning that many people have experienced include the following common uses:

  • Recommendation algorithms: The pattern detection and classification capabilities of ML models are at the core of algorithms that recommend content or products.
  • Speech Recognition: ML is used in text-to-speech software and natural language processing applications.
  • Chatbots: Chatbots used for sales and customer service, especially those that are relatively simple decision trees, are often based on machine learning.
  • Computer Vision: Giving computers the ability to “see” and understand images—whether it’s something as complex as a self-driving car or something as simple as analyzing a photo—relies on machine learning.
  • Fraud detection and anti-spam: Pattern detection helps financial services companies flag potentially fraudulent transactions and allows your email account to remove spam from your inbox.

While machine learning is powerful and widely used, it has also been criticized for:

  • Privacy: Because machine learning models require large amounts of data, machine learning has the potential to handle sensitive personal data. In some cases, the use of this data may not be authorized or fully understood by the individual. Therefore, machine learning may use or disclose sensitive information.
  • Lack of transparency: Due to their ability to learn, it can be very difficult (and sometimes impossible) to understand every step a model takes to reach a conclusion or recommendation. This restriction makes it very difficult to regulate to ensure they do not violate laws, such as fair housing or lending laws.
  • Bias and Discrimination: Biases (conscious or unconscious) in the model creator or in the data used to train the model will affect how the model learns and its output. Therefore, unless special attention is paid to correcting biases, machine learning models may inadvertently reinforce biases. A good example is that facial recognition systems identify certain skin tones more accurately than others based on the data used to train them and the race of the model creator. This could lead to significant problems with law enforcement, for example.
  • Overfitting is an error in the output of a machine learning program. This usually means that the output is too similar (fits well) to the training data, which indicates that the algorithm is not inferring or extrapolating as much as it should and therefore is not producing usable results.

  • Cross-validation is a method of testing machine learning models; developers often use it to combat overfitting. One version of cross-validation involves splitting the original data set into smaller chunks. Some are blocked, while others run through the model. You can then compare the "control" group to the "test" group to see how well the algorithm performs.