April 12, 2019
By: Yasmine Hakimpour, CPA, CA, Principal, Audit & Assurance, Research, Guidance & Support Group, CPA Canada
Artificial Intelligence (AI) is everywhere, and news articles and broadcasts, blog posts and podcasts constantly remind us how AI is “upgrading” just about everything. But what exactly is AI? CPA Canada in collaboration with the American Institute of CPAs (AICPA) recently released A CPA’s Introduction to AI: From Algorithms to Deep Learning, What You Need to Know to explain “buzzwords” such as machine learning, deep learning, robotic process automation, and computer vision, and to discuss the evolution of data, AI and computing power. The report is the first of a planned series of publications to explore AI and its impact on the accountancy profession.
Here we share the key takeaways.
AI means different things to different people, depending on your particular area of focus, but the broadest and perhaps simplest definition describes AI as the science of teaching programs and machines to complete tasks that normally require human intelligence. There are essentially two kinds of AI: Narrow AI and General AI. As the name suggests, Narrow AI is made up of narrowly intelligent systems that can exceed humans in specific tasks, such as playing chess or making medical diagnoses. These narrow capabilities are not transferrable.
General AI refers to human-level intelligence that is able to transfer knowledge between domains. While Narrow AI is all around us in language systems, vision recognition systems, and recommendation engines, General AI is still the stuff of science fiction— for now.
Why should professional accountants care about AI? Because it’s already impacting how we work. Its ability to enable innovation provides CPAs with the opportunity to improve efficiencies and quality to make better, more informed decisions fast—if we embrace it.
How We Got Here
It all starts with big data – AI cannot perform without it. The basic idea behind AI is to let a machine statistically analyze all the data being collected to derive insights much faster and more accurately than otherwise possible. Historically, data gathering was an explicit exercise with no guarantee that the data collected represented what was actually happening.
Today, thanks to the digitization of business processes (via the Internet of Things, cloud, mobile computing and social media), the advancement and availability of computing power, the maturity of algorithms and AI models, and the huge surge in investment in AI, data is being created and collected at exponentially increasing volumes.
Analyzing Big Data
According to the International Data Centre, the world is on track to create 44 zettabytes of data by 2020. Organizations are tapping into increasingly sophisticated analytics techniques to get closer to customers, to set strategy, to innovate, and to grow.
The four main categories of analytics are:
– Descriptive analytics, which provide insights into events of the past.
– Diagnostics analytics, which examine data to answer why an outcome happened.
– Predictive analytics, which look into the future to anticipate outcomes, such as demand forecasting for a supply chain operation.
– Prescriptive analytics, which provide possible outcome solutions that guide predictions into actions, such as generating ways to optimize production or inventory.
How AI Works
In order for AI programs to navigate through situational complexities, different approaches to creating software, with the ability to determine different outcomes, are necessary. The Logic and Rules-based approach uses conditional instructions and defined rules to carry out a task or to solve a problem such as “if this, then that”. This approach has been in practice for a long time and has been the underlying premise for AI until recent advances in machine learning and deep learning, which are techniques within AI.
Machine learning is the ability of algorithms to learn from experience rather than being provided with instructions. Algorithms create computational models that process large data sets to predict outputs and make inferences. More data leads to more examples, which helps the algorithm to finely tune its output/insight over time. The insights are fed back to further refine the algorithmic models making them more accurate over time.
Three different techniques are most commonly used for a machine to “learn” the problem and become smart at providing the answer:
– Supervised learning is a method to teach AI systems by example. The systems are provided with data points that are tied to expected outcomes. Once trained, the systems can take in data and provide an output that is in line with the learned model.
– Unsupervised learning requires algorithms to draw inferences from data sets by identifying patterns and looking for similarities by which that data can be grouped.
– Reinforcement learning is a technique by which an AI system learns under its own supervision by making predictions, validating them against reality, and continually adjusting itself for a better output next time.
Deep learning is an emerging and particularly exciting subset of machine learning that uses algorithms that roughly approximate the structures and functions of the human brain. The idea is to create algorithms that can simulate an array of neurons in an artificial neural network that learns from vast sources of data that would be impossible for humans to process. In addition to recognizing images and patterns, deep learning appears to be a promising way to approach complex challenges such as speech comprehension, human-machine conversation, language translation and vehicle navigation.
Robotic Process Automation (RPA)
RPA is garnering a lot of interest in the profession because of its ability to handle high-volume, repeatable tasks such as answering questions, making calculations, maintaining records, and recording transactions. In effect, RPA mimics tasks performed by humans and automates them digitally.
While the technology sounds like AI, it’s not. RPA, on its own, requires significant human involvement in the form of detailed programmed instructions. No learning occurs as a result of performing the tasks. As technology progresses, RPA systems are being coupled with algorithms to work with unstructured data related to vision, images and Natural Language Processing in order to work on more judgment-related activities.