FIN 108 - ARTIFICIAL INTELLIGENCE AND ITS CAPABILITIES

John McCarthy, 1956

The term artificial intelligence (AI) was first coined by American computer scientist ______ in _____

Broad AI, Narrow AI

2 high-level categories of AI

AI

is an umbrella term that encompasses many different approaches (e.g. machine learning, neural networks, etc.) seeking to achieve a wide array of specific goals (e.g. identifying a face in a photo, translating a language etc.).

AI

a set of technologies capable of adaptive predictive power against a well-defined problem and exhibiting some degree of autonomous learning and improvement in the solving of that problem

Broad AI (General AI or Strong AI)

• Human-like intelligence

• stuff of science fiction

• Learns across domains

• Solves many problems

• Adapts like people

Narrow AI (Weak AI)

• Task-focused systems

• Design for specific job

• Uses trained models

• No true understanding

• Common today

Structured data

is highly organized; It can be codified, placed in spreadsheets, sorted, and searched.

Unstructured data

It has no predefined model or organization. The individual data points have no clear and well-defined relationship with each other, and so can’t be sorted in a neat spreadsheet or organized in by pivot table.

Supervised learning, unsupervised learning, reinforcement learning

3 types of machine learning

Supervised learning

provides the model being trained with data that has been structured and labeled by humans and where clear objectives has been outlined

Unsupervised learning

does not include labels or instructions, and sometimes does not even provide a goal, instead allowing the model to identify its own structures, patterns, and groupings within the data

Reinforcement learning

scores the performance of variations in a model against an objective to determine which model works best for a given data set

Machine learning, Neural networks (deep learning), genetics and evolutionary algorithms

3 Selected AI Techniques

1959, Arthur Samuel

Machine learning was coined in ______ by _______

Machine learning

• A technique where computers learn patterns from data and improve performance without being explicitly programmed for every task.

• parses existing data, ‘learns’ insights from it, and then makes a prediction based on that learning.

Deep Learning models, Convolutional Neural Networks

2 subsets of neural networks

train the algorithm, test the performance of the algorithm

Machine learning models are commonly trained using datasets divided into two parts: one half of the data used to ___________ and a second half used to _____________.

Neuron

The basic unit of the human nervous system

Neuron

a simple cell capable of transmitting an electrical signal in response to stimuli

Synapses

Neurons in the human brain are connected to each other via junctions called ______

Input units, hidden units, output units

3 layers/categories of artificial neurons

Input units

are designed to receive various data inputs such as any of the types of structured or unstructured data

Output units

The layer of artificial neurons that provide the results, such as predictions or decisions

Hidden units

They typically make up most of the neurons in a neural network and provide the layers of connectivity between the input units and the output units.

Neural networks

are AI systems inspired by the structure of the human brain

Genetics and Evolutionary Algorithms

This approach applies the principles of evolution found in nature to the process of training an AI model by incorporating features such as Darwinian natural selection and the randomness of mutations.

Natural Language Processing (NLP), Machine Vision

2 Selected/Major AI Capabilities

Machine Vision

is the sub-branch of artificial intelligence concerned about the analysis and interpretation of images and videos; seeks to analyze the pixels of a photo to identify groupings and patterns of data

Optical Character Recognition

where pictures of text are converted to machine readable and searchable text

Natural Language Processing (NLP)

is the branch of artificial intelligence focused on enabling computers to interpret and process both spoken and written human language. It does this by combining the power of various AI models, including machine learning and deep learning, with the principles of linguistic structure, to break sentences down into their elemental pieces and identify semantic relationships

Automation, customization, improved decision-making, new value propositions

4 strategies that are particularly relevant to financial institutions today

Applied Mathematics, Data Science

the building blocks of artificial intelligence according to Jody Kochansky, BlackRock’s Chief Engineer

Data

is an essential input to the development of any useful application of AI

Human Talent

is critical to the successful implementation of AI within any environment

AI-Enabled Automation

The main goal is to improve the speed and efficiency of a process by reducing or altogether eliminating human intervention, which significantly lowers operational costs and enhances the user experience.

AI-Enabled Improved Decision-Making

AI technologies give financial institutions the capability to incorporate much broader and less structured sets of data into their analytics processes, theoretically enabling significantly improvised foresight.

AI-Enabled Customization

• Artificial intelligence technology enables financial institutions to break this tradeoff theoretially enabling the deployment of fully personalized financial products services at zero marginal cost once a system is in place.

• Traditional Trade-Off

  • Only accessible to those able to pay higher fees, or to particularly large and important clients.

AI-Enabled New Value Propositions

Final AI-enabled strategy that financial institutions can deploy. A financial institution may find that it is in possession of a unique set of data streams that place it in an advantageous position either to deploy monetizable insights such as more detailed macro economic reports or to build out suites of AI-enabled services that support the automation, customization, and decision-making aims of other financial institutions.

Data Challenges, Technology Challenges, Talent Challenges, Regulatory Challenges

4 Key Challenges to the Deployment of AI in Financial Institution

Data Challenges

• Highly Fragmented Data Systems

• Poor Data Quality and Inconsistent Standards

• Incomplete Digitization of Data Processes

• Difficulty Integrating Third-Party Data

Technology Challenges

• Complex ‘spaghetti’ of legacy technology systems

• Inability of on-premises mainframe systems to support required technology environment

• Lack of migration to cloud-based architecture

Talent Challenges

• Shortage of Qualified Senior Leaders

• Difficulty Acquiring and Retaining Top AI Talent

• Resistance from Existing Employees

• Staff Redeployment and Retraining Difficulties

Regulatory Challenges

• Risk of unwanted biases in AI-enabled systems

• Auditability and interpretability of AI systems

• Regulatory fragmentation and liability issues related to personal data use

• Evolution of consumer financial protection frameworks

Training Data

The foundational dataset used to teach the model how to perform a specific task (e.g., recognizing images or translating languages).