Artificial Intelligence: Machine Learning

To most people, AI and Machine Learning are synonymous. However, while machine learning is present in many artificial intelligence algorithms, it is just a subset of AI, which itself is an extremely broad topic. AI is generally defined as the simulation of human intelligence using computer algorithms and processes, but machine learning is more specific, referring to computer algorithms and models that improve themselves through training, experience, and learning. As with all concepts in data and computer science, machine learning is not a strict subject, containing subfields and subfields of those subfields, but ML algorithms typically follow similar steps. First, data is collected and prepared. Then, the programmer chooses a machine learning model that would accurately and logically help solve the problem and trains the model with the training data. After the model is trained, it is then evaluated and tested, and using the results of this testing phase its parameters and specifications are fine-tuned to yield a better result. 

While the steps that most machine learning algorithms follow might be similar, the models and manner in which they process the data vary depending on the purpose and context. The two most common types of learning are Supervised Learning and Unsupervised Learning. The former describes machine learning where the model is trained on given input and output “training” data and generates reasonable predictions for new data that could be passed in. To accomplish this task, supervised learning algorithms typically incorporate classification or regression techniques into their processes. If the data that the software developer is dealing with can be categorized, classification techniques are efficient and effective, doing essentially that and allowing the model to categorize new and unassigned data. Medical imaging, speech recognition, and credit scoring usually employ classification techniques. If the data cannot easily or logically be classified, and the programmer is instead dealing with a range of data or needs to output a real number, then regression techniques are more applicable. These predict continuous responses, such as changes in temperature or demand, and are common in fields like algorithmic trading and sales forecasting.

Unsupervised learning is when the model is trained using unclassified and unlabeled input data and tries to find hidden patterns and structures within the data. The most common task employing unsupervised learning is clustering. Self-explanatory, clustering is when the AI groups the data based on common factors. This is extremely helpful in areas like gene sequence analysis, object identification, or market research. 

Supervised and unsupervised learning are just two of the many types of machine learning. There’s reinforcement learning, where the model is goal-oriented and uses reward feedback to find maximum benefits or costs. There’s semi-supervised learning, which falls somewhere in between supervised and unsupervised learning. There’s self-learning, which incorporates neural networks to solve tasks without rewards or external supervision. There’s feature learning, sparse dictionary learning, and robot learning. And within all these different types of learning, the models employ a similarly diverse array of algorithms: neural networks, decision trees, Bayesian networks, etc. The number of options and possibilities that a programmer needs to consider whenever creating an algorithm is huge, and many times can only be optimized with trial-and-error. The developer also needs to be careful not to over-train his model, fitting it exactly to the training data without leaving any room for future predictions. However, once all these factors are chosen, tweaked, and perfected, machine learning has futuristic and life-changing powers. For example, Rutgers researchers developed an algorithm that can analyze art and classify various works by style, genre, and artist more accurately than a human can. The UK’s RAC is using machine learning to detect low-speed car crashes, enabling faster response times and lower insurance costs. A company named BuildingIQ has a platform that optimizes HVAC energy usage in large buildings, in some cases reducing usage by 10-25%.