Unveiling Artificial Intelligence

Understanding the What, How and Why Behind Artificial Intelligence

The Essence of AI: What Is It?

Intelligence is such an alluring and elusive concept that defines our human civilization. In our everyday discussions, we talk about intelligence in a manner that suggests it to be a quality which organisms have more or less of. Examining our deep fascination with intelligence, it is no surprise that we marvel and hold in high regard individuals who display prodigious amounts of this beautiful quality. Wouldn’t it be great if we could create thinking machines? It was a question along these lines that motivated a group of researchers to convene at Darthmouth College in the summer of 1956 for the Dartmouth Summer Research Project on Artificial Intelligence. One of those researchers was John McCarthy who is credited with coining the term “Artificial Intelligence”.

The field of Artificial Intelligence(AI) revolves around designing systems and entities that replicate the spectrum of human cognitive abilities. Over the past century, we have come to view Intelligence not as a singular quality which can be easily measured, but as a multidimensional concept. Indeed, intelligence is an umbrella term encompassing emotional understanding, problem solving, creativity and more. Hence, articulating a rigorous and encompassing definition of intelligence that everyone accepts has been problematic. To make progress in our understanding of intelligence, many have found it useful to ignore the issue of having a definition and instead focus on the practical effects we can expect from an building entity which appears “intelligent” to us. After all, it is only through the observation of the behaviors of an entity that we conceive the concept of Intelligence.

Computational Agents

A computational agent is an entity that exists in some environment, makes observations and acts based on them. The entity may or may not be physical. When the entities are physical, we enter the field of robotics. Fundamentally, agents make decisions. It is this decision making process that can be decomposed into the processes of observing and acting in their environment.

There is a profound relationship between computational agents and their environments which is pivotal in shaping the agent’s intelligence. Indeed, the challenges presented by an environment strongly influence an agent's cognitive structures and problem-solving methodologies. This dynamic is epitomized in the concept of 'umwelt' - a term borrowed from biology that refers to the world as it's perceived by a specific organism. For an AI agent, its umwelt is limited by the data it receives and the agent’s ability to act. I will devote a future blog post to further outline and decompose the decisions an agent makes to solve problems and knowledge representation.

Understanding the capabilities of these agents is paramount. AI specialists often distinguish AI based on its capabilities and range of tasks. At one end of the spectrum, we have 'Narrow or Weak AI', designed to perform a specific task (like recommending music or recognizing images). Then, there's 'General or Strong AI', a hypothetical AI with cognitive functions akin to a human, enabling it to perform any intellectual task a human can do. General AI is often referred to as Artificial General Intelligence(AGI) in the literature as is the stated goal of companies like OpenAI.

Engineering AI: How Does It Work?

The question of how to build AI elicits strong responses from researchers which often lead to heated debates. For example, leading voices like Gary Marcus propose a modular and hybrid approach to building intelligent systems. On the other extreme, you will find those who advocate for end-to-end neural network based systems. There is also disagreement on whether and to what extent to draw inspiration from biology. I find the discussions stimulating and will expound my views in a future post.

While there’s great variety of opinions and research directions on how to build an AI agent, there is widespread recognition of some(by no means exhaustive) of the following areas of AI:

• Knowledge Representation: As mentioned earlier, this field is concerned with how an agent represents itself and its environment. This includes how it represents other agents and their actions. Knowledge representation is closely related to the concept of Umwelt.

• Machine Learning: Probably the most discussed field is focused on how an agent progressively improves it’s performance on tasks with experience.

• Reasoning, Planning and Search: I am lumping reasoning, planning and search together even though they could all each stand alone. I did this because in my view they are concerned with how an agent achieves it’s goals in a complex strategic space. Note that search here it not restricted to physically searching a physical space but refers to the more abstract problem of searching a “state space” given a problem representation.

• Computer Vision: Computer Vision is about endowing agents with the ability to perceive the world satisfactorily. Vision is a deceptively hard problem but much progress has been made over the past decade. It was the application of machine learning techniques-specifically Neural Networks to the ImageNet challenge in 2012 that launched the deep learning revolution.

• Natural Language Processing: Natural Language Processing(NLP) focuses on enabling agents to understand, interpret, and generate human language. The release of ChatGPT by OpenAI in November 2022, brought this field to public consciousness.

It’s important to note there exists significant overlap and merging of the areas listed above. Currently, Machine Learning(ML) enjoys tremendous popularity and research progress in ML tends to permeate other sub fields of AI. Given it’s popularity, I think it’s worthwhile to spell out the three main areas of ML for the general reader.

• Supervised: Supervised learning trains models using labeled data, where both input and desired output are provided. A common application is image classification.

• Unsupervised: Unsupervised learning deals with unlabeled data, focusing on uncovering hidden patterns or structures within. An example application is market segmentation.

• Reinforcement learning: Reinforcement learning involves agents who take actions in an environment to maximize cumulative reward. The agent learns from trial and error, receiving feedback through rewards or penalties.

Furthermore, a framework is shared across the learning paradigms, whether that's reinforcement, supervised, or unsupervised learning. The framework is described below:

• First, data is represented as numbers or vectors. Think of vectors as a list of numbers that we hope captures the essence of an issue.

• Once framed this way, an objective is specified, which is what we want our AI to achieve. In the AI field, the objective is often referred to as a loss function.

• With the data and objective in hand, the next step is optimization, refining the AI model to perform better and better until it can accomplish the task efficiently. The matter of optimization is a whole field in itself and there are families of optimization algorithms. In Neural Networks, the backpropagation algorithm utilizing gradient descent is the de facto optimization technique.

The Ethics and Purpose of AI: Why Do We Need It?

The growth and advancements of AI naturally prompt the question: What is AI for? At its best, AI has the potential to aid humanity in addressing some of our most complex challenges. Imagine a future where AI can predict natural disasters with pinpoint accuracy, tailor education to individual students' needs in real-time, or revolutionize drug discovery, medical diagnoses and treatment. Likewise in robotics, it is common to talk about the 3 D’s- Dull, Dirty and Dangerous. The 3 D’s are a reference to the kind of tasks we’d rather have performed by a robot. In addition, Stuart Russell in his book “Human Compatible” points out that AI judiciously applied to the goal of raising living standards could lead to a tenfold increase in global gross domestic product(GDP), from $76 trillion to $750 trillion per year. While I haven’t examined the source and methods behind the calculation of that estimate, I mention it to give readers a sense of how the opportunities around AI is perceived.

However, as with all powerful tools, AI prompts essential ethical considerations. Who controls AI? As we integrate AI deeper into societal infrastructures, deciding on control becomes paramount. It determines who benefits – and, conversely, who might lose out. Will it be a tool for the few, concentrating power and wealth further? Or will it be democratized, benefiting society at large? As we stand at this technological crossroad, the decisions we make today will shape the AI-driven world of tomorrow.

My aim in this inaugural post, is to provide readers with a foundational understanding of the AI landscape. Future articles will dive into the technical, economic, and political facets surrounding AI. As we navigate further into the AI era, I am committed to offering a balanced viewpoint. AI undoubtedly presents unique opportunities, but it also brings forth considerable technical and ethical challenges. As we shape our AI-augmented future, it's crucial that we stay informed and engaged, ensuring that AI's evolution reflects the very best of human values and aspirations.