Machine Learning

Machine learning (ML) is a subset of Al (as illustrated in Figure 1.1) that uses algorithms to analyze data to carry out specific tasks, such as making predictions, without relying on explicit programming as in rule-based expert systems. Machine learning uses pattern recognition and inference to learn from data. The larger the data set, the more examples from which the algorithm can learn through trial and error (CPA Canada, & A1CPA, 2019). Machine learning is used by Netflix to make recommendations for TV shows and movies based on what the user previously viewed, thereby enhancing the user experience. The more content that a machine learning Al views, the more refined and accurate its predictions will be. To cite another example, Amazon uses machine learning to analyze purchasing data on its products to forecast demand, identify fraudulent purchases, and provide customized recommendations and promotions (Camhi & Pandolh, 2017).

We have so far discussed machine learning in consumer-driven fields, but what does it have to offer the field of accounting?

Since machine learning enables systems to be adaptative (Negnevitsky, 2011), computers equipped with ML can learn from experience, resulting in improved performance over time. As an example, K.PMG³ employs IBM’s machine learning algorithms to assist their clients with compliance with the IFRS 16 lease accounting standard (Samuel, 2018). Specifically, the K.PMG Contract Abstraction Tool extracts data from tens of thousands of contracts, each of which may be hundreds of pages long, and analyzes the data for compliance with the leasing standard. As the program examines more contracts, it becomes “smarter” and can offer more refined and targeted feedback. Markus Kreher, Global Head of Accounting Advisory Services and Head of Finance Advisory, KPMG in Germany, said: “We’ve effectively trained the solution to read and understand contracts just like an attorney would” (Samuel, 2018, para. 4).

Similarly, Deloitte uses an award-winning machine learning tool called Argus to review, identify, and extract key accounting information from any type of electronic document. Argus learns from every human interaction and can analyze sales and leasing contracts, employment agreements, invoices, meeting minutes, financial statements, and legal letters (Deloitte, 2015), eliminating a time-consuming and costly manual task normally performed by auditors. Thus, Argus enables Deloitte auditors to be more productive by spending their time interpreting results and exercising professional skepticism.

Possible applications for ML in accounting include computers that learn from audit failures by analyzing data sets that were the basis for incorrectly issuing a clean audit opinion. Another application of ML could be on tax positions taken by firms that were rejected by the federal and state tax authorities.

There are four main types of machine learning: supervised, unsupervised, semi-supervised, and reinforced.

Supervised Learning

Supervised learning is a process in which a computer algorithm learns from a set of training data that is labeled (tagged with certain attributes such as poor credit or good credit) and paired as input and output variables (X results in Y). Supervised learning is often used for classification (e.g., identifying fraudulent and non-fraudulent transactions) and prediction (e.g., predicting the number of uncollectible accounts based on the aging of individual customers). For example, suppose an auditor wants to predict the number of uncollectible accounts based on the aging of customer balances.

A developer would provide the model with historical data that includes the input variable (aging of individual customers) and the output data (amount of uncollectible accounts written off). A supervised ML model would then “learn” the rules based on the historical data and then predict the amount of uncollectible accounts when new aging data is provided.

Unsupervised Learning

L'nsupervised learning, on the other hand, is a process in which the computer algorithm only learns from unlabeled input data (i.e., the outputs are unknown). For example, consider a controller of an accounting department who wants to better understand why customers do not pay within 30 days. An unsupervised machine learning model would teach the algorithm to identify trends and patterns related to the input data. The algorithm might identify that customers with outstanding accounts receivable balances over 120 days might have similar zip codes, income levels, or credit scores. One of the most common approaches for unsupervised learning is cluster analysis, whereby input data is grouped based on similarities.

Semi-supervised Learning

Semi-supervised learning is a hybrid of supervised and unsupervised learning using both labeled and unlabeled data during the training process. As many real-world data sets often contain incorrect or missing labels (Yao et al., 2018), using a semi-supervised approach could lead to improved learning compared to a purely supervised or unsupervised method.

Reinforcement Learning

Reinforcement learning is a process in which a computer algorithm trains itself, learning from data through trial and error. The agent is rewarded for performing specific tasks correctly and penalized for performing tasks incorrectly. These algorithms are designed to maximize rewards and minimize penalties.

Deep Learning

Deep learning (DL) is a subset of machine learning (as illustrated in Figure 1.1) that uses artificial neural networks to discover patterns from data. Artificial neural networks (ANNs) are a collection of artificial neurons (also known as units) that receive data as an input, and then logic is applied to produce an output. ANNs are frequently used for pattern recognition such as facial recognition in pictures and videos, speech recognition, social networks, etc., frequently accelerated by tensor processing. A tensor processing unit (TPU) is an applicationspecific circuit developed by Google that is used to increase processing speeds for machine learning and deep learning applications. Whereas traditional machine learning relies heavily on the guidance provided by human programmers, DL models use ANNs (inspired by the biological neural networks found in human brains) to analyze large volumes of data to teach itself.

Deep learning can be used in a variety of applications, such as facial recognition, image classification, speech recognition, and text translation. Current applications of DL include self-driving vehicles that slowdown as it approaches a pedestrian cross-walk, ATM machines that reject a counterfeit banknote, and a smartphone app that translates the image of a street sign in a foreign language (The Mathworks & MATLAB, 2018).

EY uses deep learning technology to reconstruct documents (e.g., invoices and contracts used during an audit) that were poorly-scanned using optical recognition software (Duffy, 2019). Researcher Sophia Sun (2019) promotes the use of DL to support audit decision making in two key areas: information identification and judgment support. DL algorithms can analyze semi-structured data (e.g., text data from supporting documents such as invoices, contracts, etc.) and unstructured data (such as images, audio, and video from the internet, social media, etc.), which will aid in information identification. Further, Sun notes that the predictive performance of DL models is superior to classic machine learning w-hen the volume of data and the number of variables is large. To maximize the use of DL in an audit context, Sun suggests designing and developing a data warehouse to train the algorithms to make predictions and assist with audit procedures.

Natural Language Processing

Natural language processing (NLP) is a subfield of Al that focuses on the interaction of computers and people using human languages. The Siri application on Apple’s iPhone and Amazon’s Alexa uses NLP to understand requests for completing a task.

NLP can be broken down into two types: natural language understanding and natural language generation.

Natural language understanding (NLU) enables computers to understand instructions provided in human language. For example, when a user asks their iPhone for weather conditions by using the following voice command, “Hey Siri, what’s the latest weather?” Siri understands their command and retrieves the information by searching on the Internet.

Natural language generation (NLG) enables computers to produce human language so that people can understand computers. After Siri performed a search of the latest weather, it replies with an answer such as the following example: “it’s currently clear and 28 degrees in Philadelphia. Expect partly cloudy skies starting tonight. Today’s high will be 33 degrees, and the low will be 23 degrees”. Siri understood that the user was asking it to look up the weather conditions at a specific location and time, and then returned the results in spoken English words. NLG can also convert data visualizations, such as charts and graphs, into a verbal description that can be understood by humans (CPA Canada & A1CPA, 2019).

The potential for using NLP in accounting is substantial, as it can be used to process and analyze large amounts of data. NLP has been used to analyze various textual documents related to corporate financial performance and compliance with accounting standards and regulations (Fisher et al., 2016). Text mining uses NLP to extract meaning from textual data. Applications of NLP include text mining components such as semantics analysis, text classification, text summarization, and text translation. Additional applications include speech recognition, question-answering systems, and chatbots.