Unpacking the Foundations of AI

Artificial Intelligence (AI) is a topic generating a lot of buzz and our CTO, Timothy Ay, presented about the foundations of AI at Cairo Dreamin’ earlier this year. Below is a synopsis of what he covered, with the recording capturing his entire presentation. 

What is Artificial Intelligence?

At its heart, AI is about the simulation of human intelligence by machines. The goal is to enable machines to perform tasks that typically require human cognition, such as:

  • Reasoning
  • Learning
  • Problem-solving
  • Perception
  • Natural language understanding

This involves using technology to further our own cognitive skills, leveraging data, information, and knowledge to make decisions or carry out actions. A basic AI model involves data, the analytics performed on that data, and the action taken based on those analytics.

AI is Everywhere

While we once only thought of complex systems, AI is integrated into many things we use daily without even realizing it. A simple example is the capability of your phone to unlock itself using facial recognition. We are seeing AI show up in various areas like vision, speech, planning, robotics, and even coding co-pilots.

AI is commonly divided into three main categories:

  • Narrow AI: These are systems designed to perform specific tasks, such as chatbots and recommendation engines (like Netflix). Another prime example is the movie AlphaGo.
  • General AI: This is a hypothetical stage where AI would possess human-like cognitive abilities, capable of understanding and performing a wide range of tasks. This is the type of AI people imagine when they think of something that can handle all their business data and answer questions about it.
  • Superintelligent AI: While currently theoretical, this level describes AI that qualitatively surpasses all human intelligence. Reaching this level is currently widely debated by industry experts, with time frames anywhere from 5 to 20 years.

The Pillars of AI: Algorithms and Data

An AI model is essentially a combination of algorithms and data.

  • The Role of Algorithms: Algorithms provide the method for AI to process information and algorithms fall into different categories:
    • Supervised Learning: This is where the model is trained on labeled data, learning from examples of correct inputs and outputs. This is used for tasks like regression and classification, such as detecting fraud. 
    • Unsupervised Learning: In this approach, the model learns from unlabeled data, finding patterns and relationships on its own. This is used for tasks like clustering and dimensionality reduction. Netflix’s recommendation algorithm learns your likes and dislikes by your actions as you use their system. 
    • Reinforcement Learning: Here, the model learns through trial and error, receiving rewards or penalties based on its actions. AlphaZero utilized reinforcement learning once the rules of the game were “modeled.”
  • The Importance of Data: Data is described as the “fuel” for AI models. The quality and quantity of data are key factors for model performance. Data needs to be clean, normalized, and potentially enhanced through feature engineering. Data preprocessing is a crucial step. 

Training and Testing the Model

Building an effective AI model involves a training process. This includes initializing the model parameters, running data through it (forward propagation) to make predictions, calculating errors (backpropagation), and optimizing the model to minimize errors. It’s important to note that the data used is divided into two groups – training data and testing data. Training data teaches the model to identify patterns, while testing data evaluates how well the model generalizes to new, unseen data. Both data sets should ideally come from the same dataset to avoid anomalies. 

Machine Learning vs. Deep Learning

Machine Learning is a domain within Artificial Intelligence. It involves systems learning patterns from data to make decisions without explicit programming. Deep Learning is a sub-domain within Machine Learning that specifically uses neural networks with multiple layers (hence “deep”) to learn complex patterns directly from raw data. Neural networks function by taking inputs (data), running them through interconnected nodes (which have weights and parameters), and producing an output value. The process of optimizing the model involves adjusting these weights based on the error rate compared to actual values, and as you can imagine, requires significant computational power.

Challenges in Managing AI Models

Once deployed, AI models require management. Data drift occurs when the patterns or distribution of the input data changes over time, potentially making the model less accurate. Similarly, model drift happens when the task or environment the model was designed for changes. Managing AI involves tracking experiments and handling these changes, a field sometimes referred to as MLOps or XOps.

The foundations of AI involve simulating human cognition through a combination of data and algorithms. Understanding the different types of AI, the role of data and algorithms, and the nuances of training and managing models provides a solid basis for appreciating this rapidly evolving technology. If you have questions about putting AI into practice at your organization, reach out to us.

Agile Cloud Consulting​


Schedule time with us