The term artificial intelligence broadly refers to applications of technology to perform tasks that resemble human cognitive function and is generally defined as the capability of a machine to imitate intelligent human behavior.
AI typically involves the theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages. John McCarthy, one of the founders of AI research, “once defined the field as getting a computer to do things which, when done by people, are said to involve intelligence.”
Scope
While the definitions for AI discussed above provide a general outline of the meaning of the term, there is no single universally agreed upon definition of AI.
In practice, AI is used as an umbrella term that encompasses a broad spectrum of different technologies and applications, some of which are described below.
- Machine Learning (ML): Machine learning is a field of computer science that uses algorithms to process large amounts of data and learn from it. Unlike traditional rules-based programming, ML models10 learn from input data to make predictions or identify meaningful patterns without being explicitly programmed to do so. There are different types of ML models, depending on their intended function and structure:
- Supervised Machine Learning: In supervised ML, the model is trained with labeled input data that correlates to a specified output. For example, a dataset of animal photos (input data) can be labeled as “cats” or “not cats” (output data). The model is continuously refined to provide more accurate output as additional training data becomes available. After the model has learned from the patterns in the training data, it can then analyze additional data to produce the desired output. Results of supervised ML models are typically reviewed by humans for accuracy and fed back into the model for further refinement. Supervised ML is successful when the model can consistently produce accurate predictions when provided with new datasets. For example, the ML model learns to recognize if a new picture is a cat or not.
- Unsupervised Machine Learning: In unsupervised ML, the input data is not labeled nor is the output specified. Instead, the models are fed large amounts of raw data and the algorithms are designed to identify any underlying meaningful patterns. The algorithms may cluster similar data but do so without any preconceived notion of the output. For example, a time series of trade events can be inputted into an unsupervised model, with the model identifying groups of similar trades as well as outliers. Results of unsupervised machine learning models are then interpreted by humans to determine if they are meaningful and relevant.
- Reinforcement Learning: In reinforcement learning, the model learns dynamically to achieve the desired output through trial and error. If the model algorithm performs correctly and achieves the intended output, it is rewarded. Conversely, if it does not produce the desired output, it is penalized. Accordingly, the model learns over time to perform in a way that maximizes the net reward. For example, in the securities industry, reinforcement learning models are being explored for options pricing and hedging.
- Deep Learning: A deep learning model is built on an artificial neural network, in which algorithms process large amounts of unlabeled or unstructured data through multiple layers of learning in a manner inspired by how neural networks function in the brain. These models are typically used when the underlying data is significantly large in volume, obtained from disparate sources, and may have different formats (e.g., text, voice, and video). For example, some firms in the securities industry are developing surveillance and conduct monitoring tools built on deep learning models. Deep learning applications can be supervised, unsupervised, or reinforcement based.
- Natural Language Processing (NLP): NLP is a form of AI that enables machines to read or recognize text and voice, extract value from it, and potentially convert information into a desired output format, such as text or voice. Examples of NLP applications in the securities industry range from keyword extraction from legal documents and language translation to more complex tasks, such as sentiment analysis and providing relevant information through chat-boxes and virtual assistants.
- Computer Vision (CV): CV (also referred to as machine vision) is a “field of computer science that works on enabling computers to see, identify and process images in the same way that human vision does, and then provide appropriate output.” Frequently a CV application will use ML models to interpret what it “sees” and make predictions or determinations. Examples of CV-based applications include facial recognition, fingerprint recognition, optical character recognition, and other biometric tools to verify user identity.
- Robotics Process Automation (RPA): RPA refers to the use of preprogrammed software tools that interact with other applications to automate labor-intensive tasks, resulting in increased accuracy, speed, and cost-savings. RPA tools are generally used for high-volume, repetitive processes involving structured data, such as account reconciliation, accounts payable processing, and depositing of checks. Some market participants do not consider RPA to be a form of AI because its focus is on automation of processes in a manner more akin to a rules-based system. However, others consider it to be a rudimentary form of AI, particularly when it is combined with other technologies such as ML.
Key Components of AI Applications
AI applications generally involve the use of data, algorithms, and human feedback. Ensuring each of these components is appropriately structured and validated is important for the development and implementation of AI applications. The discussion that follows highlights how each of these components influences the development of AI applications.
- Data: Data generation in the financial services industry has grown exponentially over the past decade, in part due to the use of mobile technologies and the digitization of data. The importance of data has likewise rapidly increased, and some have even referred to data as a more valuable resource than oil. Furthermore, cloud technology has enabled firms to collect, store, and analyze significantly large datasets at very low costs. Firms in the financial services industry now collect data from a variety of internal sources (e.g., trading desks, customer account history, and communications) and external sources (e.g., public filings, social media platforms, and satellite images) in both structured and unstructured formats, and analyze this data to identify opportunities for revenue generation as well as cost-savings. This explosion of data in the financial services industry is one of the key factors contributing to the increased exploration of AI in the industry.
Data plays a critical role in the training and success of any AI application. AI applications are generally designed to analyze data by identifying patterns and to make determinations or predictions based on those patterns.
The applications continuously and iteratively learn from any inaccurate determinations made by such applications, typically identified through human reviews as well as from new information, and refine the outputs accordingly. Therefore, AI applications are generally best positioned to yield meaningful results when the underlying datasets are substantially large, valid, and current.
- Algorithms: An algorithm is a set of well-defined, step-by-step instructions for a machine to solve a specific problem and generate an output using a set of input data. AI algorithms, particularly those used for ML, involve complex mathematical code designed to enable the machines to continuously learn from new input data and develop new or adjusted output based on the learnings. An AI algorithm is “not programmed to perform a task, but is programmed to learn to perform the task.” The availability of open-source AI algorithms, including those from some of the largest technology companies, has helped fueled AI innovation and made the technology more accessible to the financial industry.
- Human interaction: Human involvement is imperative throughout the lifecycle of any AI application, from preparing the data and the algorithms to testing the output, retraining the model, and verifying results. As data is collected and prepared, human reviews are essential to curate the data as appropriate for the application. As algorithms sift through data and generate output (e.g., classifications, outliers, and predictions), the next critical component is human review of the output for relevancy, accuracy, and usefulness. Business and technology stakeholders typically work together to analyze AI-based output and give appropriate feedback to the AI systems for refinement of the model. Absence of such human review and feedback may lead to irrelevant, incorrect, or inappropriate results from the AI systems, potentially creating inefficiencies, foregone opportunities, or new risks if actions are taken based on faulty results.
