ARTIFICAL INTELLIGENCE - SOME BASICS

It’s been a few years since I came into contact with a company named Pelican AI, its founder Parth Desai, and some of the company’s technology experts, who have been innovators in the deployment of Artificial Intelligence technology in the financial services industry for many years.  Since that time, these people have educated and excited me about the potential for this technology, whilst at the same time alleviated the concerns I had on the impact of AI on the future of the human race.

I now spend a lot of time, travelling around presenting and speaking to financial services executives on how this technology can now help with their daily work lives across payments processing and financial crime compliance.  Naturally, there are still many people who although appreciative on how AI helps them with their personal lives, do not really understand much about the technology itself (me being one of those people before engaging with Pelican!).

To assist these people and also at the same time apologising to various publications whom I have blatantly copied text from, I detail below some key pieces of information about this exciting technology which I hope will give some background knowledge.

The goal of Artificial Intelligence is to allow computers to mimic human intelligence so that they can learn, sense, think and act in order to achieve automation and gain analytic insights.  To mimic human intelligence, AI applications adopt two computation approaches: rule-based and non-rule based. Rule-based AI ‘learns’ using pre-defined rules and knowledge, and ‘thinks’ by inferring logical causes and effects according to ‘if-then-else’ rules.  Non rule-based AI ‘learns’ with machine learning algorithms and ‘thinks’ using trained AI models.  Key enablers of AI development have been the huge increase in the amount of data available and vastly increased computing power in both traditional computers and mobile devices, along with the ongoing development of machine-learning algorithms. The adoption of Natural Language Processing (NLP) also significantly improves data pre-processing.

AI has now evolved to the point where it can interact and communicate with humans through analytics and automation.  In a nutshell, the goal of AI is to enable computers to mimic human intelligence so that they can learn, sense, think and act, performing human-like cognitive functions (e.g. recognizing patterns from data and understanding images, text or speech), as well as making predictions, recommendations and decisions to achieve automation and gain analytic insights.

While machines can ‘sense’ and ‘act’ in relation to humans and the environment through user interfaces, sensors and robotics, the core of AI is the ‘think’ component – e.g. making predictions, decisions or recommendations.  To mimic human intelligence, AI applications adopt two computation approaches: rule-based and non-rule based.

Rule-based decision making, also known as ‘expert systems’, is one of the simplest and most common types of AI.  It involves translating a fixed set of pre-defined rules and knowledge about decision making into a knowledge base and computer logic.  Rule-based AI mimics thinking by inferring logical causes and effects according to a chain of ‘if-then-else’ rules in its knowledge base.  Rule-based AI systems became popular in the 1960s and dominated the period from the 1970s to the 1990s, Applications were limited however, as it can be very time consuming and expensive to add new rules or knowledge to these systems so that they can react to the changing business environment.  Also, in many scenarios it can be very challenging to explicitly define rules in a programmatic or declarative way.  This led to a demand for, and the development of, non-rule-based AI.

Non-rule-based AI not only reacts and make decisions according to a pre-defined set of rules, it can also derive extra, meaningful insights by ‘learning’ automatically from its inputs.  Machine learning (ML) is one way to enable the ‘learn’ capability in AI.  This involves using a set of learning algorithms driven by mathematical techniques which allow machines to learn from data, instead of being explicitly programmed to perform certain tasks.  The training process uses the learning algorithm to derive relationships between data points from training data, which is commonly a subset of historical data.  The outputs of the training are trained machine learning models, which can perform predictions or make decisions according to the data patterns observed from the input data, or from queries provided by users.  ML is able to identify subtle data patterns which cannot be easily described by humans, and extract insights from less structured data.  As a result, ML has become one of the main techniques driving the current wave of AI applications – from call centre voice analysis to autonomous vehicles.  However, ML in general is data and computing-heavy.  In order to find data patterns or relationships and make accurate predictions, ML needs significant volumes of data. 

There are four types of machine learning algorithms: ‘supervised’, ‘unsupervised’, ‘semi-supervised’ and ‘reinforcement learning’:

Supervised learning is a type of machine learning algorithm in which humans guide the system by labelling the relationship between every variable of input data and every variable of output.  For example, a supervised machine learning model can learn which sanctions alerts should be classified as false positives by identifying when previous instances of the same scenario have been cleared by a user.  Techniques such as linear regression, logistic regression and decision tree classification are used to train supervised learning models.

An unsupervised learning algorithm, by contrast, identifies patterns in a set of unlabelled input data automatically.  In other words, rather than learning the relationships between a given set of input and output data, unsupervised learning algorithms are designed only to learn whether data patterns or structures exist.  One of the most common use cases is to identify different customer segments with different preferences and demographics.

The third type of machine learning algorithm is semi-supervised learning.  It is used on data sets which are only partially labelled.  As it can be costly and time-consuming to label large data sets, semi-supervised models are commonly applied to real-world machine learning problems.  Semi-supervised learning combines supervised and unsupervised learning algorithms to either infer the pattern or structure of input variables, or make predictions about the unlabeled data and feed those predictions back to the model as training data.

The last type of machine learning algorithm is reinforcement learning.  This determines optimal behaviour based on feedback from its environment.  Reinforcement learning models incentivise machines to optimise behaviour by providing positive feedback, or ‘rewards’, which teach the machine to strive towards a preferred behaviour.  In the absence of preexisting training data, a reinforcement learning algorithm must learn to predict the consequences of its actions through trial and error.  It therefore learns from its own prior experience rather than from a predetermined set of examples. Reinforcement learning is commonly used to automate the decision-making process in interactive scenarios.

However, Machine Learning can only detect patterns in raw data effectively if the algorithm can read and interpret the data.  Natural Language Processing (NLP) is a branch of AI that aids computers to understand natural human language, which is usually relatively unstructured structured and contains ‘implicit’ meanings.  Differences in authors’ writing styles or cultural backgrounds can also affect how the content of natural languages is interpreted.  Data scientists could manually tag the meaning of each sentence as part of data pre-processing, but this would be too time consuming with the large volume of training data.  NLP uses different techniques to learn and understand language, enabling computers to understand not only single words and word combinations, but also grammar and hence written or spoken sentences.  NLP categorises analyses and comprehends the following elements using a combination of statistical analysis and pre-defined rules:

• The language used, such as spelling, grammar, meaning and connotations of words
• The content of the text – namely the message the author wants to convey (i.e. logical analysis)
• The person behind the text, including their style and feelings (i.e. sentiment analysis)

In short, NLP has the ability to recognise linguistic patterns and interpret meaning in languages, which it achieves through the use of linguistic analysis tools such as partofspeech tagging, automatic text summarisation, and entity and relationship extraction.  NLP is widely used to process large volumes of unstructured text in big data to reveal patterns, to identify and tag relevant information and to summarise document contents — all with minimal human input. 

I hope the above gives the reader some understanding of this exciting technology.  My closing remark is that I hope that we all realise its potential and embrace it to improve our business lives in the same way that it is helping us in our daily lives. If you would like to know more about how AI can improve your business processes, please do not hesitate to either contact me or one of my Pelican AI colleagues.