A Brief History of AI and ML
The term "Artificial Intelligence" (AI) was first used by John McCarthy, a Dartmouth mathematics professor, in 1956. He created the term to apply for a grant to fund the Dartmouth conference, which is now generally considered to be the major event that created AI as a research discipline. The term "Machine Learning" (ML) was an even older discipline emerged from computer science. It is a branch of AI where a computer algorithm is trained to autonomously learn from data without being explicitly programmed to do so. ML is a very broad concept and it includes Supervised and Un-supervised, semi-supervised and Reinforcement Learning or deep learning.
"Intelligence" usually means "human". Only human beings were supposed to have intelligence. The term 'artificial intelligence' means intelligence like behavior by non-human, e.g. machine. Why human being has intelligence is because we have a big and complex brain, which contains hundreds of billions of neurons. It is the networks of those inter-connected neurons that help us to process information and making decisions in an intelligent manner. Therefore, understanding of the neuro-network, how they are processing information, as well as the simulation of the neuro networks, is in the forefront of AI from the very beginning.
"The advancement in the cloud and open source computing has provided a huge momentous boost to the application of AI and ML models in our everyday life"
While the first neural network was built by Warren McCulloch, a neurophysiologist, and Walter Pitts, a mathematician in 1943 to simulate the brain, it is without surprise that two psychologists had made significant contributions to the early development of AI methodology. In 1949, Donald Hebb, a Canadian psychologist who wanted to understand how the function of neurons contributed to the psychological process of learning, wrote a book, The Organization of Behavior, that pointed out the fact that knowledge and learning occur in the brain primarily through the formation and changes of synapses between neurons. The neural pathways are strengthened each time they are used. If two nerves fire at the same time, he argued, the connection between them is enhanced. This is a concept fundamentally essential to the ways in which humans learn, and it had a huge impact on the following works in neural network and machine learning.
In 1958, Frank Rosenblatt, another psychologist, conceived of a simplified mathematical model of how the neurons in our brain operate. The Perceptron, as it was called, takes a set of binary inputs from nearby neurons, and multiplies each input by a continuous weight similar to the synapse strength to each nearby neuron. If the sum of those weights exceeds a predetermined threshold, the model output a 1, otherwise output a 0. This is similar to the way a neuron works it is either fire or not fire based on the inputs from the nearby neurons.
The Perceptron neural network model was an advancement from the Mcculoch-Pitts model since Perceptron model has the learning mechanism that the Mcculoch-Pitts model did not have. Rosenblatt, inspired by the pioneering work of Donald Hebb, came up with a method to make the artificial neuron networks learn, which was crucial for the development of a usable AI or ML algorithm.
As computers became more advanced in the 1950s, it was finally possible to simulate a hypothetical neural network using a computer instead of using the electrical circuits by Warren McCulloch in 1943.
Recent Advancements in AI and ML, and the Reasons Behind It
The early works and developments of the neural networks in AI were largely in the research and academic space. The most recent advancement in this area in the past decade is mainly in the area of 'deep learning', which is primarily based on the Artificial Neural Networks (ANN) that has been around even before the term of AI was created in 1956. What's new is that the deep learning is built on huge amount of data, including unstructured data, and the most advanced ML models usually including dynamic updating or self-learning, using up to date real-time data and they involve multiple and deep layers of the artificial neurons. The recent advancements in AI also focused on getting computers to act without being explicitly programmed and to use algorithms to learn from data interactively, such as in natural language processing (NLP), voice and facial recognition, and automatic driving machines.
Three major recent advancements in technology have driven the advancement and wide applications of the AI and ML in the past decade: 1) The availability of extremely cheap data storage; 2) the rapid development of, and easy access to, superfast computers; 3) the cheaper and fast data collection and distribution channels. The advancement in data storage made it possible to store and access to oceans of big data, particularly from the internet. The superfast computers made it possible to develop and apply complex and deep layer neural networks quickly enough for practical and real-time usage. The cheap and fast collection and distribution of big data, such as web-based search and browsing data, has made it possible to use live data to train and update AI and ML models to make real-time decision.
In addition, the advancement in the cloud and open source computing has provided a huge momentous boost to the application of AI and ML models in our everyday life. The cloud and open source computing are gradually leveling the playing ground for AI and ML development and application. Everyone, not just the big corporations could have easy access to oceans of data and a supercomputer to build and run deep learning models to solve everyday problems. It wouldn't be surprised that the future advancement on AI and ML application will be coming out of individuals or startups using cloud and open source computing other than big corporations.
With the availability of big data and super fast computing power, AI and ML scientists have been able to 'train' deep learning networks with thousands of layers deep with billions of parameters, very similar to the human brain.
While AI and Machine Learning became the most trending buzzwords in computer science nowadays, their origins were many decades old and their advancements benefited from researchers from many different areas. If history is a lesson, the future advancement of AIs, including such hot areas as NLP, voice and facial recognition, automatic driving, requires the collaboration of experts from multidisciplinary areas in computer science, mechanical science, data science, linguistics, neural science, and psychology. In particular, future AIs researches and applications should take advantage of the advancements in cognitive psychology, learning theories, and behavioral science to bridge the gap between machine and human being. The advancement in the multidisciplinary areas, such as Cognitive Computing and Behavioral and Cognitive Biometrics, have great potential in making seam changes in data security, fraud detection and prevention, better customer experience, as well as robotic development to perform human tasks that used to be hard to perform by non-human beings.
Major areas of applications of AI and ML in the financial industry
1) Marketing and Customer Service. The application of simple supervised ML model in marketing has been widely used in the past 20 years or so. The more widely spread usage of the advanced AI and ML technology has made great strides in Fintech as the transparent requirements are relatively lighter in these areas. Also, the most advanced AI technology, such as voice and facial recognition, has seen great potential in these areas.
2) Loan Underwriting and Credit Approval. The use of consumer credit profile data in automatic credit approval or loan underwriting has been around for about 30 some years. The accumulative of big consumer data and using of advanced ML model could open a front in identifying more credit behavioral patterns that might have not been possible before thus make some dents in this area. However, with the strict regulatory requirement for transparency in credit decisions, the use and impact of AI and ML in this area would be constrained and limited.
3) Fraud Detection and AML: The used of neural networks in fighting credit fraud and money laundry had already been more widely used even before the most recent advancement in the AI and ML technology. The application of advanced AI and ML technology would be a double edge sword in these areas. Fraudsters would use illegally acquired big data from data breach and AI to conduct frauds thus makes the effort in using AI algorithms to fight fraud and money laundry a must have in these areas.
4) Credit Loss and Stress testing/CCAR: After the great recession about a decade ago, the regulatory oversight on banks had got much stricter. With the implementation of CCAR/DFAST, huge resources had been polled into the modeling of credit losses in the past 10 years or so. With the upcoming rollout of CECL, using the right data and best modeling methodology to meet the tightened requirements are challenges to the big banks. While the traditional modeling techniques had played its role and paid their dues, there are great potentials to leverage the big data and advanced ML to improve the estimation of the currently expected loss using forward-looking economic scenarios.