Big data, deep learning, artificial intelligence, machine learning. I can bet these buzzwords show up at least once a week in your feed and have been for the past few years. Our radar can’t escape them even if we wanted to. By now everyone has some idea of what those words mean, but can you really explain their meaning if someone asked you? Or better yet, do you know enough that you could apply those concepts to your work? In this article, I’ll go through a brief review of machine learning and then explain some of its newest use cases for business.
What exactly is Machine Learning?
In short, machine learning (ML) is the study of statistical methods and algorithms used by computers in order to perform a task without explicitly being told. The ‘learning’ part means that the computer tries to find patterns in the data it’s provided with. The way it learns is through algorithms we devise.
Machine learning is closely related to artificial intelligence. In fact, it can be seen as a discipline within it. The history of machine learning can help us understand it better so let us go through a quick overview.
A Brief History
As is often the case with new technologies, it is hard to pinpoint exactly when in time the birth of ML took place. Often, people who go as far as assigning a date to its inception are really just telling you the date some underlying concept was introduced.
Here we adopt a more practical view:
- 1950 to 1980: We have the early stages of machine learning. From the creation of the Turing Test (a test to tell whether a machine’s ability to exhibit intelligent behaviour is equivalent to, or indistinguishable from, that of a human.) to the creation of the first neural network framework for computers and basic applications such as improving computers’ performance in the game of checkers and rough pattern recognition.
- 1980 to 2000: We see examples of a computer creating a general rule from an input (training data), the pronunciation of words in an early stage and the specialization of programs represented by the iconic chess match between DeepBlue and Garry Kasparov ending in a defeat for the then world champion.
- 2000 to 2010: Geoffrey Hinton coins the term “deep belief networks,” the prototype term for deep learning algorithms, which take the input and make successive transformations on them until we have the output. ImageNet, an extensive visual database, is created for visual object recognition.
- 2010 to 2019: We hear more and more about big data, data science deep learning, and artificial neural networks. In 2010, deep learning emerges as the next step in machine learning methods. In 2016, Google’s AlphaGo (a deep learning based software) beats the world’s best players of GO, a game considered many times harder than chess.
We have seen rapid development, especially in the last few years. Google, Facebook, and Amazon make investments in AI and ML, in particular with the creation of Google’s X Lab, DeepFace algorithm and Amazon’s own Machine Learning platform.
In 2018, Google introduces AutoML, which enables people with less expertise to benefit from machine learning methods. Google’s AutoML trains your data-sets with custom models. AutoML currently focuses on models for Natural Language Processing, Visual Object Recognition, and Translation.
Similarly, MindsDB, in our mission to democratize machine learning, provides an AI service that does the work that a machine learning expert would normally do, training the data with different models & delivering results to users based on their queries.
General applications of Machine Learning
As Ethem Alpaydin put in his book Introduction to Machine Learning:
“… we all became producers of data. Every time we buy a product, every time we rent a movie, visit a web page, write a blog, or post on social media, even when we just walk or drive around, we are generating data.”
And machine learning is all about learning from data. The algorithms machine learning experts use usually fall under the following categories:
- Associations. The software finds associations between two actions and can assign a probability based on how often those actions occur. Example: the software may find out that a customer that buys a book by some author X also typically buys a book by some author Y. So the people who buy X and NOT Y are potential customers for Y and we may find an association rule such as 70% of people who buy X also buy Y.
- Classification. In this category, machine learning systems fit a model to some already available data in order to make predictions. Example: to classify customers as low-risk or high-risk clients, we may gather all the information we have on them and set a particular rule to tell whether or not a customer falls into one of the classes. Then new customers will be labelled as low-risk or high-risk, based on this past data.
- Supervised & Unsupervised Learning. In supervised learning, the aim is to provide a rule that takes the input data and provides the correct output. The supervisor is the person who tells whether the output is correct or not. In unsupervised learning, the aim is to try to detect patterns and regularities in the input data only, without a supervisor to tell you whether there are correct values. For example, a company may want to group customers who are similar, based on the data they keep on them such as demographic, financial and/or past purchases, etc. Then we have a natural customer segmentation and we can learn about the similarities of groups of customers without looking for something in particular. If we want something particular then we apply the supervised learning and set our criterion for what returned value is correct.
- Reinforcement Learning. The focus here is on the sequence of actions required to achieve a goal. We define a reward and possible actions. Then the program learns the correct sequence of actions to achieve the goal or reward. After more trial runs it can learn to do it as quickly as possible.
Uses of ML: How is Machine Learning Being Used in Industry?
“The technologies and techniques of AI and ML are still so new that the main adopters of the techniques are the large software companies able to hire and to invest in the necessary expertise”
Despite machine learning applications being in their early stages, in recent years machine learning adoption has begun to rapidly accelerate as more organizations see the benefits that this technology can bring to their business.
For instance, O’Reilly recently surveyed more than eleven thousand people who worked with AI, Data Science, and related fields. It reports that about half of the respondents said they are ‘just looking’ into the technology and more than one-third have been working with machine learning models for at least the past 2 years. That means about two-thirds of the respondents are already in touch with the technology at some level.
Machine learning is being used in a variety of sectors and use cases are showing up in a wide range of areas. The specific use cases are diverse; they range from adjusting paywall pricing for different readers based on the probability of readers subscribing to reducing scrap rates in semiconductor manufacturing. Here are some of the major ways machine learning is helping organizations:
- Supply Chain Management & Inventory Management: IBM’s “Watson” system was able to detect damaged shipping containers based on visual pattern recognition. Still, in Supply Management, machine learning has also been used to forecast the demand for new products, and in helping identifying factors that might affect this demand. Machine learning is also helping to reduce the costs of inventory management while simultaneously adjusting inventory levels and increasing inventory turnovers.
- Fraud Detection: One can use a combination of supervised learning to learn about past frauds and learn from them — and unsupervised learning in order to find different patterns in the data that might have slipped or anomalies people might have missed. For example, MasterCard uses machine learning to track purchase data, transaction size, location, and other variables to assess whether a transaction is a fraud.
- IoT: Devices and processes themselves generate data. Industrial companies produce large amounts of data in their daily operations. Machine learning can then be used to infer some useful information. For example, by analyzing different processes in a factory, we can use this data to prevent accidents or deal with production difficulties through ML methods.
- Predictive Maintenance: For this one, the rail industry takes the lead. Almost half of the companies in this sector use some sort of predictive analysis supported by machine learning. Infrabel, a Belgian company in this sector, has a single tool to predictive maintenance which combines IoT, monitoring tools and machine learning models to provide alerts when something is not quite right.
- Personalization & Customer Churn Prevention: Personalization shows a customer different offers, providing a personalized experience and therefore increasing chances that the customer will convert. An example of this, Adobe uses machine learning algorithms to provide a personalized user experience with their optimization engine Adobe Target, but no unique experience will surpass a good experience, and one of the most popular metrics used to measure whether clients are satisfied is churn rates. However, there are many more based on how often the customer replies to marketing emails, the time since they last login, are they a daily active user? etc. Then we can train the model to identify customers who might be leaving the service or product.
While the history of machine learning is quite recent even when compared to traditional computing, its adoption has accelerated over the last several years. It’s becoming more and more clear that machine learning methods are helpful to many types of organizations in answering different kinds of questions they might want to ask and answer using data. As technology develops, the future of corporate machine learning lies in is ability to overcome some of the issues that, as of now, still prevent the widespread adoption of machine learning solutions, namely explainability and access to people beyond machine learning engineers.
Explainability refers to the degree to which a given outcome can be explained. In other words, why does this model work and what its limitations are? MindsDB actually already solves this by identifying values where it has enough information to make predictions and, what is perhaps even more important, where it doesn’t have information to make predictions. From there it learns from these values to understand why things didn’t work.
The second point, the democratization of machine learning, has already seen efforts by Google and MindsDB to provide solutions, or at least facilitate, the use of machine learning by a larger group of developers. In particular, MindsDB allows people outside the niche of machine learning experts to use the tool and benefit from the predictions neural networks can make.
Finally, with more access to the tools machine learning provides and a better understanding of how the algorithms work, machine learning experts (and others who use machine learning) can serve as better consultants and managers can make better decisions, which is the ultimate goal of every business.