ML Explained

Machine learning is a subset of artificial intelligence in which software adapts without being explicitly programmed. Here’s how it works - and what matters for business owners.

Machine learning is more than a popular buzzword - it’s a movement. Search the term on Google and you’ll get millions of results ranging from simple definitions to news about the latest advances in the field.?

But what most business owners or executives really want to know is how ML applies to any given business or, more specifically, what it can do to help make a business more competitive. Interested in learning more about what machine learning is, how it works and how we use it at AltaML to build custom solutions that help businesses achieve their goals? Let this be your guide.

What is machine learning?

Machine learning (ML) can be defined in a number of ways but, at its simplest, it involves creating an algorithm with the ability to adapt without being explicitly programmed. With traditional software, you need to know the business rules. With ML, you take a bunch of data, and run it through the algorithm to identify patterns. Or, at least, from a non-technical perspective, you can think of ML that way, understanding that a precise definition can get far more complex.?

You’ve probably heard of artificial intelligence (AI) and are wondering how ML is related. Let’s clear that up first: ML is a subset of AI, which is the science of giving computers some of the cognitive abilities that have traditionally been assumed to be human.?

Yes, computers have always been better at dealing with large amounts of data than human beings. But machine learning has revolutionized computing because many problems cannot be solved by writing a traditional algorithm. The data set might be too large to know what to look for, or there could be numerous ways to approach a problem. Machine learning allows us to overcome these problems, tap into the potential behind a large data set, and even uncover new opportunities within it

Plus, while most computer programs are composed of algorithms that tell the program exactly what to do, machine learning is designed to perform a task without being told how to do it. Instead, a computer “learns” how to perform a task from historical observations, while, in a way, a human designs the algorithms that are telling that machine how to learn. This is the defining feature of machine learning and one, frankly, that is very hard for many people to grasp. So, let’s take a simple look at how machine learning works.

How does machine learning work?

How can a computer learn? Most ML learning tutorials like this will tell you that there are three common methods:

• Supervised learning
• Unsupervised learning
• Reinforcement learning

But hold up. Unless you’re planning on going into data science. You don’t really need to know that. If you’re interested, you can read about it in more detail in the breakout box below. (If not, feel free to skip it; you don’t need to know how a combustion engine works to drive your car either.)

If you aren’t planning on becoming a data scientist, what you need to understand about machine learning is that it learns by examples. As it turns out, we can learn by examples too. Here are a few to help you understand machine learning better.

Computer Vision

One really simple-to-understand field in machine learning is computer vision, or teaching computers how to understand visual data like images or videos.?

This is another skill humans are adept at; we can recognize objects and people quite easily, even when they look quite different. So, we can recognize that a picture of a cat, a cartoon image of that cat and a cat that’s peeking out from behind a fence are all similar (they’re all cats). A computer has a much harder time with this. So, why do we want them to do it? Well, once a computer can recognize a cat with relative accuracy, it can do so on a scale that a human - or even a whole fleet of humans - absolutely cannot do. Plus, machine learning can do this by learning from examples.?

The training process involves running the algorithm over and over and seeing how well it can correctly identify matches. As the computer identifies defining characteristics of a cat, the importance of these characteristics is evaluated and adjusted to guide the algorithm and improve the accuracy.

Finally, the model is tested. Just because the algorithm is good at categorizing the training data does not mean it can successfully identify new data. The testing data that was set aside is now used. The data is fed to the model and the efficacy of the model is measured by how accurately it can be. With enough data and the right algorithm, a computer can learn to identify images of objects with relative accuracy, even when those images are very different from each other.

Natural Language Processing

Another common example is what’s called natural language processing, or a computer’s ability to learn and understand language. Again, language is something people can manage quite well, especially when it comes to subtleties like irony, sarcasm and humor. For computers, it’s a complex problem only the advent of AI and machine learning has allowed them to tackle. Like image recognition, it also has applications at a scale far beyond what’s possible for humans, such as translating languages on the fly, monitoring social media or providing voice assistance.

What do these examples tell us about machine learning? That machines are able to learn. They may not do it quite like humans do, but ML allows them to take in information, extract knowledge from it and, over time, get better at solving a particular problem or completing a particular task.

_______________________________________________________________________

Supervised Learning: Supervised learning involves starting with raw data that is labeled (such as a medical record for a given patient might be labeled as healthy or sick). The algorithm then determines what features and patterns in the data for the same label are missing from the other groups. This creates a model, which can be used to predict the categories of future data.

Self-supervised learning is the latest application of supervised learning and refers to a way to train computers to do things without providing them with labeled data, but with data that has some underlying structure/meaning (like images or text). In other words, machines use the format of the data to learn underlying representations and extract knowledge from the data. For example, it might remove random words from text and then learn to predict which words are missing. This type of machine learning has actually allowed a computer to create a unique image based on a text description !

However, not all data is labeled. In that case, unsupervised learning is used.?

Unsupervised Learning: Unsupervised learning algorithms focus on creating their own categories for raw data (clustering), determining how they are associated (association learning), detecting anomalies (anomaly detection), simplifying the data set (dimensionality reduction) and a few other, less-common methods. With clustering, for example, the categories are created by finding similar characteristics in the data. If some known features exist, those features could be used to create the basis for the model. Varying how many features the algorithm starts with is a good way to marry conventional knowledge with discovery.

Reinforcement Learning: Reinforcement learning is analogous to our traditional notions of learning. An algorithm is given a goal, a current state and a set of actions it can make. Then, it is given feedback about whether its attempts are progressing towards that goal. The algorithm uses the feedback to refine the attempts and favor some strategies over others.

___________________________________________________________

How does machine learning differ from software development?

This is a question we get a lot at AltaML. After all, every business uses computer technology to solve problems, which leads people to wonder how machine learning differs from software development. While the ultimate goal of both software development and machine learning is to solve a problem, the way that goal is reached is very different. Software development creates a program that can deliver some sort of result. However, with machine learning, past results are one of the inputs. Traditional programming uses input data to produce output; whereas machine learning uses input data and past results to create a model (learning), which is then used on new data to create an output (prediction).

In software development, you are essentially just making known procedures more efficient and faster. This is not so for machine learning. Machine learning is a process of discovery. It uses data to learn and build a model with the aim of solving a business problem. For example, we might take data about people’s driving habits to see if we can predict who is most likely to get in an accident or look for patterns in medical imaging to try to accurately predict the likelihood of disease. These problems would be close to impossible to solve with traditional programming.?

Essentially, a machine learning algorithm looks at the data and creates a model to try to solve that problem. In the process, it can also pull up other patterns that might lead to other potential models and use cases.

Where can machine learning be used?

Wherever there is data, machine learning can help us understand that data better. This can turn legacy data repositories (like all that customer or sales data you’ve been stockpiling) into untapped assets. The key is to identify how the data could improve decision-making, efficiency, or competitiveness in a way that has a return on investment.

For example, machine learning could be used to assess the effectiveness of a company’s bidding process. The algorithm could look at the history of a company’s winning proposals and identify key metrics that the winning proposals shared. This could shed light on a company’s competitive edge. But that’s just one example.?

Here are few other ways that AltaML has applied machine learning to business problems:

• Intelligent triage of tasks, thus allowing more time for higher-value work
• Automation of elements of contract review
• Detecting safety issues in heavy industry workplaces using image data?
• Predicting load on sections of the electrical grid
• Predicting recurrence of cancer
• Better understanding of the likelihood a loan will default

And many more… As you can see from this varied list, there is pretty much no industry that doesn't have applications using AI/ML.?

Machine learning is about discovery, but the process doesn’t have to be a black box. Knowing how models are created is important if the model will guide a company’s decision-making. This is particularly important to industries where companies need to be clear on how decisions were made. For example, in partnership with DynaLIFE , AltaML leveraged machine learning to improve medical screening, but the screening tool wasn’t designed to remove or replace medical professionals - it helped them do their job better. In most cases, machine learning does not replace expert knowledge, nor would we want it to. It simply creates tools to increase accuracy and speed up processes.

Finally: Why Machine Learning Requires an Experienced Team

Algorithms can uncover new connections, but they can also see false patterns. Therefore, machine learning is actually a very human-in-the-loop process. As an algorithm discovers new characteristics of the data, those characteristics need to be assessed. Machine learning developers weigh the importance of discovered characteristics to better reflect real patterns in the data. This ensures that false patterns are not propagated and decision-making is well-informed.

Identifying hidden gems in repositories of data is difficult. Remember when we said you didn’t need to understand the ins and outs of machine learning? This is where expert advice comes in. Machine learning developers express business problems in a way that machine learning can solve. This requires an understanding of what data is available, what data is needed and what types of questions can be answered with different data sets.?

A machine learning developer also helps create a model responsibly by ensuring that the technology is being developed in a way that is ethical, transparent and accountable. Because AI has the potential to be so disruptive, careful oversight and governance can help prevent the use of biased algorithms, protect individual privacy and generally maintain user trust in AI solutions.

Finally, AI developers are there to deploy AI solutions in an organization in a way that is as seamless as possible.

At AltaML, this is all part of our process. We work with organizations from start to finish, helping them define use cases, experiment with and refine the ones with the highest potential, and then pilot and deploy the best options.

So, what’s next?

Interested in learning more about machine learning and how AI could help move your business forward? Read Is My Organization Ready for AI?