Many will recognize the term AI or Artificial Intelligence, understanding that this broad term applies to almost any technique which allows for computers to mimic human behavior. Machine Learning is a subset of AI which includes abstruse statistical techniques, supporting gradual task improvement through experience gained. These broad sets of algorithms are used to extract useful models from raw data which are in turn used for a variety of mining tasks & synthetic tasks. In this blog we aim to explain both the various types of machine learning algorithms including supervised, unsupervised and reinforcement learning, as well as highlighting its business use examples. Many of the terms used in the blog can be further understood through Google’s extensive ML glossary, found here.
“Machine learning research is part of research on artificial intelligence, seeking to provide knowledge to computers through data, observations and interacting with the world. That acquired knowledge allows computers to correctly generalize to new settings”
- Yoshua Bengio
Recognized as the most common type of Machine Learning, supervised learning algorithms are designed to learn through example, hence the term ‘supervised’. To achieve this, the algorithm uses provided input and output data. This provided data is labeled to provide a base for future data processing. Using this data, the goal is to produce an accurate mapping function which in-turn allows for prediction of the desired output. Supervised learning is then further compartmentalized into a range of algorithms, including, but not limited to decision trees, logistic regression & support vector machines. Of course, as with many facets of AI, supervised learning has both advantages and disadvantages. Firstly, supervised learning is a simple process to understand and is extremely useful in classification problems. That said, supervised learning is not ‘real-time’ data, meaning that there will be delays in results required. Alongside this, supervised learning requires substantial computation time for training and is considerably more complex in comparison to unsupervised learning due to the need for labeling all inputs.
Unsupervised learning refers to the process of giving an algorithm no labeled data and leaving it to structure its own output. Through this lack of labeling, models using unsupervised learning can suggest subtle trends that would otherwise be unfound, especially when using semi-supervised learning. Unsupervised Learning can be seen as extremely beneficial, as it then becomes possible to uncover previously unknown patterns in data. The downside? Unsupervised learning results make it hard to find meaning in the data due to the lack of answer labels and this also makes it harder to compare to supervised learning tasks. Applications of unsupervised learning include clustering, anomaly detection and association mining.
Before moving on to reinforcement learning, it is important to also address semi-supervised learning. Semi-supervised learning sits between the two aforementioned methods, using a mixture of both tagged and untagged data to fit models. This technique is best suited for a large amount of data with both tagged and untagged sections. An example of this? Amazon’s Alexa! Jeff Bezos has previously spoken very highly of semi-supervised learning, suggesting that the reduced amount of labeled data needed to achieve the same accuracy improvement by 40 times.
A recent buzzword, reinforcement learning is a technique used to aid the development of 'learning' in an environment, through the process of trial and error. This in turn, uses 'feedback' to correct itself. Unlike the above, whereby feedback provided to the agent is a correct set of actions for performing a task, Reinforcement Learning uses reward and punishment as signals for positive and negative behavior, with the goal to find a suitable action model that would maximize the total cumulative reward of the agent. This feedback design acts as a motivational factor for the RL-agent, whereby an understanding of outcome pushes the agent to learn the method of maximizing accumulated rewards over time. Applications of reinforcement learning are often seen in robotics for industrial automation, for data processing and in the creation of training systems. This technique has many positives, including being seen to solve various complicated problems which cannot be solved with conventional techniques including robotic movement and video game completion, similar to human problem solving in regard to process and repeat and the lessening in the potential for repeat mistakes.
A challenge that should be recognized when looking at RL is the time it takes to generate data. This seems to be commonplace in the majority of keynote discussions surrounding different algorithms and subsets of AI. Alongside this, it is suggested that RL assumes the world is Markovian, which it is not. The Markovian model describes a sequence of possible events in which the probability of each event depends only on the state attained in the previous event.
Machine Learning, in its many forms, is seen as a staple in AI, with the potential to scale, a key factor in the growth of intelligent machines. A number of industries utilizing large swathes of data already recognize the potential of ML, as it’s regularly used in financial services, healthcare, government processes, transportation, and more. Using ML not only increases the potential effectiveness of the product, but can aid in a competitive advantage over those proceeding without.
Currently a Project Manager at Sama, Taylor Rouleau has a passion for ensuring ethical and sustainable practices in tech. After 5 years leading production teams for our customers, Taylor's expertise is applied internally in our Project Management Office. She heads up efforts to maintain our industry-leading data training processes with a special focus on Security & Compliance.