The Internet is inundated with articles about artificial intelligence (AI), deep learning, and machine learning. These buzz terms are used interchangeably, and often incorrectly.
I recently met a venture capitalist who firmly stated that he had stopped investing in machine-learning applications and was now focusing on AI. He was surprised to hear that machine learning is, in fact, an important component of AI. So, if you are not quite clear on all the terms, don’t worry, ―you’re not alone! Below are the basic definitions of each with respect to retail applications.
Artificial intelligence is a generic name covering a wide variety of methods, tools and techniques that mimic “cognitive” functions or tasks that people associate with the human mind, such as “learning,” “planning,” “reasoning,” or “problem solving.” AI also covers the many approaches that address these tasks, two of which are knowledge-based systems and machine learning.
With knowledge-based systems, also called expert systems, a large amount of expert knowledge is uploaded to a computer’s memory. The learning part comes from the reasoning that the program uses to solve complex problems (such as “if–then” and “inference logic rules”). One example of an expert system task in the retail industry is assortment planning or planogram planning, in which expert knowledge about product interactions can help retailers optimize their offerings.
As the name implies, machine learning means that the computers “themselves” do the learning without the need to be explicitly programmed. So, a computer will extract data patterns to give predictions and smart recommendations, constantly improving the algorithm based on experience. Market basket analysis is an example of where machine learning can be used, analyzing large amounts of purchase data to find correlated products that are frequently sold together (e.g., coffee and tea or bread and milk).
Machine learning methods are often divided into two main branches: unsupervised learning and supervised learning.
In unsupervised learning untagged data samples are introduced to a system to find significant patterns. It often involves detecting anomalies and data clusters to find positive and negative deviations from the norm and provide actionable recommendations. Fraud detection within sales data, identification of unmet demand in stores, store segmentation, and anomaly detection in inventory data are all good examples.
With supervised learning, data samples that are introduced to the system are tagged to match inputs with their desired outputs. For example, in customer lifetime value, known customers are tagged as “churned customers” or “loyal customers.” The system learns their features to help predict high value customers in advance.
Within supervised learning falls a group of neural networks inspired by the structure and function of the human brain. These models are often used to represent complex (non-linear) relationships between inputs and outputs and got elevated by the deep-learning revolution, including press coverage of high-profile applications such as Apple’s Siri.
Deep learning models are essentially neural networks with multiple processing layers. Highly complex, they require fast computer systems and a large amount of data, which became readily available with the development of big-data technologies.
Deep learning models evolved to address unsupervised learning and were successful in signal processing applications, such as image processing on e-commerce sites. However, deep-learning models do not always guarantee great results, and their “black-box” outputs can be difficult to interpret as they are not descriptive. Deep learning should be viewed as one of many important tools in the data science toolkit.
More practical examples of “Putting AI to Work in the Retail Enterprise” will be covered during professor Ben-Gal’s presentation at RetailTechCon September 6-8.