Canada, and specifically, Montréal, buzzes with massive investments and numerous initiatives in artificial intelligence (AI). This is why Raymond Chabot Grant Thornton started AI and advanced analytics practice this year to help SMEs take advantage of this latest technological revolution.

I’ll be leading the section on AI in the Tax-R&D practice’s newsletter. I’ll be delving into the questions you should be asking, practices to adopt, examples that could serve as inspirations for your business. I’ll also be looking at disruptive technologies (ones that displace established technologies and shake up the industry or ground-breaking products that create a completely new industry) that could improve, enhance or even revolutionize business practices.

Marvin Lee Minsky, one of the creators of AI, defined it as the theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition and decision-making. In this first article, we’ll be looking at using AI and machine learning to create agents (autonomous software) able to think and reason like humans.

What is machine learning?

Machine learning is a set of techniques (linear regression, Bayesian classification, boosting, neural networks, etc.) used to give a machine the ability to learn from past experiences so it can deduce rules that will form new knowledge and serve as a basis to analyze new situations. For example, using analytical data collected by ecommerce websites, a machine learning algorithm can determine the rules that characterize users most likely to delete their account. Using those rules, the algorithm can analyze users’ actions to offer promotions just before they would take the critical decision. If the calculations of former rules are systematically based on the matchings, set by experts, between analytical data and users, the learning process is said to be supervised. Semi-supervised learning occurs when rules’ calculations partially rely on matchings set by experts. When there are no matchings used in the calculations, the learning process is unsupervised. In the case of reinforced learning, the algorithm results are reused to guide the calculation of the next predictions.

A neural network is a network of computing units (neurons) operating in parallel and arranged in tiers to perform complex functions. Each successive tier uses the output received from the tier preceding it. When a neural network is used for visual recognition, the first tiers identify, for example, lines, curves and angles, the middle tiers identify shapes while the last tiers identify objects such as eyes or wheels. The resulting automatic learning method is known as deep learning. Although deep learning has existed from some 10 years, it has grown in leaps and bounds with the increase in computational power, in particular, the ability to use GPU-accelerated computing (graphics processing unit) for general processing and the advent of large databases.

State a machine learning problem

Machine learning can be used in business to resolve several problems. In order to state the problem properly, you need to determine to which of the four categories of machine learning algorithms it fits.

Raymond Chabot Grant Thornton - image

Classification: The goal of problems in this category is to classify or label each object in a data set in supervised or semi-supervised mode, i.e. with the help of experts. They can be unary (e.g. can unusual client transactions be detected?), binary (e.g. will lead Doe be converted or not?) or multiple (e.g. what type of product is user John most likely to buy: computer, portable or smartphone?).

Clustering: The main difference with problems in this category is that there is no human intervention for determining the classification rules and classes. In order words, it is an unsupervised learning process. Problems in this category include portioning of users for marketing purposes (e.g., what are our main market segments, based on our clients’ demographics and their purchases?) or understanding their behaviour (e.g., how can we classify the key words used for searches in our website?).

Regression: This category includes problems to predict or calculate numeric values rather than classes. This includes price calculations (e.g. what is a fair sale price considering the various production constraints?), product demand predictions (e.g., considering the last marketing campaign, how many products will we sell next month?), etc.

Ranking: In this category, the importance of an object compared with other objects in the same data set is calculated. Examples include recommendations (what five products should be displayed for a user, based on purchasing history?), website layout (how should displays be organized on the website considering users’ browsing history?), etc.

So, the answer to the question in the title is YES. Artificial intelligence can be embedded into your business as long as you can state your problems in terms of the four categories. Initially, you should start with simple projects with considerable added value for your business. Stating the problem is all the more important because it will influence the data collection procedures you will need to implement, which we’ll be seeing in the forthcoming article, with emphasis on the quality and relevance of captured data.

01 Nov 2017  |  Written by :

Jean-François Djoufak, manager, taxation, Raymond Chabot Grant Thornton

See the profile