Exploring the World of Artificial Intelligence

Understanding the Two Major Approaches to Artificial Intelligence

When it comes to addressing real-world problems, there is no singular "AI". The term refers to a diverse portfolio of approaches and techniques, each with its own strengths and weaknesses. We can broadly categorise AI into two major families of approaches: data-driven (including machine learning and deep learning) and model-based.

{Disclaimer: For the purposes of this post, we are only considering narrow AI and not AGI (Artificial General Intelligence)}. 

Data-driven AI Approaches

Data-driven techniques rely on access to vast amounts of high-quality data. They learn patterns from this data to tackle similar problems or situations. This approach proves extremely valuable when the solution to a specific problem is unclear or when numerous potential contributing factors are involved. A prime example is the use of data-driven approaches in medicine. By harnessing hospital data systems storing medical records of large patient cohorts, these approaches can assist physicians by suggesting likely diagnoses.

ChatGPT represents another example of a data-driven approach. It employs the vast expanse of the internet to train a large languages model, enabling it to generate human-like text in response to provided input (prompts).

Model-based AI Approaches

Model-based approaches, on the other hand, adopt a fundamentally different perspective. Instead of learning everything from scratch using raw data, they capitalize on pre-existing knowledge of how to address a given problem. This knowledge is encoded in a manner that smart AI-based systems can leverage to enhance our problem-solving capabilities. These approaches describe the problem and outline what effective solutions should look like, sparing us from relearning everything from the ground up.

This approach is particularly valuable in domains where domain experts already possess a deep understanding of how to address problems, such as optimizing manufacturing and logistics processes. For instance, it's evident to an expert how to load trucks efficiently or ensure accurate package deliveries. Encoding this information in a suitable format facilitates problem-solving.

The Crucial Difference

One crucial distinction between these two AI families lies in their ability to explain the reasoning behind their outputs. Generally, data-driven approaches struggle to provide explanations for generated solutions and may even have difficulty ensuring the validity of their solutions. This is because the learning process creates a complex black-box model that is challenging to dissect. In contrast, model-based approaches rely on encoded knowledge, which can be verified and validated to ensure solution validity. In most cases, model-based approaches also support the explanation of the solution generation process and facilitate what-if analyses.

Choosing the Right Path for AI Problem-Solving

The above overview simplifies a complex landscape but serves as a foundation for understanding the differences in AI approaches and informs the selection of the most suitable approach for addressing new problems. If you already know how to solve a problem, consider starting with model-based AI approaches. If you're uncertain about a solution or dealing with a complex, multifaceted issue, data-driven approaches are often the way to go.

 

And Finally

The tables below give a snapshot of the pros and cons of both approaches as well as an overview of the current capabilities and potential future developments:

 

Data-Driven AI
Pros	Cons
Adaptability: Data driven AI excels at learning from large datasets, making it highly adaptable to a variety of tasks.	Explainability: One of the major drawbacks is the lack of explainability; it's often hard to understand why the AI made a particular decision.
Problem Solving: This approach is particularly useful when the solution to a problem is not clear cut, as it can analyse numerous variables to find the best outcome.	Data Dependency: These models require large amounts of high-quality data, which can be a limitation in some cases.
Flexibility: Data driven AI has broad applications, from healthcare and finance to customer service and entertainment.	Validation Challenges: Ensuring the validity of solutions can be difficult due to the "black-box" nature of the algorithms.
Model-Based AI
Pros	Cons
Explainability: Unlike data-driven AI, model-based approaches are often easier to understand and interpret.	Flexibility: These models are generally less flexible than data-driven models, as they rely on pre-defined rules and knowledge.
Efficiency: These models can be more efficient as they leverage pre-existing knowledge and don't have to learn everything from scratch.	Complexity: Creating a model-based AI system can be complex and time-consuming, often requiring domain experts to encode their knowledge into the system.
Domain Expertise: Model-based AI is particularly useful in fields where there's already a deep understanding of the problems, such as logistics or manufacturing.	Adaptability: They may not adapt well to new or unexpected situations that were not initially programmed into the model.

 

Data-Driven AI	Model-Based AI
Current capabilities
Chatbots for customer service (OpenAI, Google)	Supply chain optimisation for logistics improvements (SAP, Oracle)
Image recognition for diagnosis and facial recognition (Google, IBM)	Weather forecasting for short-term, accurate predictions (IBM, AccuWeather)
Recommendation systems for personalized suggestions (Netflix, Amazon)	Quality control for automated manufacturing inspection (GE, Siemens)
Fraud detection for identifying suspicious transactions (Darktrace, Kaspersky)	Energy optimisation in smart homes and industry (Google, Schneider Electric)
In Five Years
Advanced healthcare through predictive analytics	Predictive maintenance to anticipate equipment failures
Autonomous vehicles with improved safety	Expert systems expanded to new domains like law
More natural language chatbots	Simulation models for climate, social systems
Highly personalized digital assistants	Personalised medicine using health models
In Ten Years
Generalised learning with minimal data	Cross-industry optimisation via interconnected models
Real time translation between languages	Incorporating ethics into decision-making models
Emotional intelligence for detecting human emotions	Human-AI collaboration on problem solving
Collaborative robots working alongside humans	Sustainability models to optimise resources