The Five Schools of Thought in AI/ML.

The Five Cults of AI/ML

If you take a guided tour of the landscape of artificial intelligence (AI) and machine learning (ML), you will find that it is a non-homogenous ground marked by diverse methodologies, philosophies, and techniques - each aimed at forging intelligent systems in their own interpretation.?

This landscape is inhabited by five principal "cults" or schools of thought, namely. Symbolists, Bayesians, Analogists, Connectionists, and Evolutionaries.?

These schools, with their distinct approaches and belief systems, underscore the diverse endeavors of "replicating or surpassing human intelligence" through computational means.

Symbolists stand as the custodians of logic and rule-based reasoning.

Their domain is characterized by the use of symbolic representations, where problems are articulated through if-then rules, decision trees, and knowledge bases. They strive for models that are both interpretable and precise.?

The Symbolists' toolkit includes Expert Systems, which encode domain knowledge into a set of rules, and Logic Programming, with languages like Prolog, enabling problem-solving via logical inference.?

In the standard ML models, their go to tool would be Decision Trees, which also graphically represent decisions and their possible outcomes.?

A great example of Symbolist achievement is MYCIN, an early AI system designed for diagnosing bacterial infections and recommending antibiotics, showcasing the strength of this approach in domains where expert knowledge can be codified.

However, Symbolists grapple with scalability and the challenge of handling uncertainty without integrating additional frameworks.

Bayesians chart their course through the uncertain terrains of AI with the compass of probability theory.?

They model uncertainty using Bayesian inference, updating the probability of hypotheses as new evidence emerges.

This school's main tools include Bayesian Networks, which graphically model dependencies among variables, and Na?ve Bayes classifiers, simple yet effective probabilistic classifiers that assume independence between features.?

Hidden Markov Models (HMMs) are also pivotal for analyzing sequence data.?

A quintessential application of Bayesian methods is in spam filtering, where Na?ve Bayes classifiers discern spam from legitimate emails based on word frequencies.?

Bayesians are celebrated for their adeptness at handling uncertainty and incorporating prior knowledge.?

However, they face challenges with the computational demands of large networks and the reliance on accurate prior distributions.

Analogists derive their strength from the past, employing case-based reasoning and analogies to address new challenges.?

This approach is underpinned by a memory of past experiences, with techniques like Case-Based Reasoning (CBR) for storing and retrieving cases, and K-Nearest Neighbors (KNN), a straightforward algorithm that classifies based on the closest training examples in feature space.?

Analogists excel in complex, unstructured problems, as seen in legal reasoning where decisions are informed by precedents.?

Despite their intuitive and adaptable methodology, the requirement for a vast, comprehensive case memory and the high cost of compute of large database searches are their significant hurdles. The latest GenSQL invention by #MIT may prove to be a significant net positive for this school of thought in the very near future.

Connectionists draw inspiration from the brain's neural architecture, utilizing neural networks to learn from data.?

Their approach is embodied in Artificial Neural Networks (ANNs), with interconnected nodes or neurons arranged in layers to progressively extract features and patterns.?

Convolutional Neural Networks (CNNs), specialized for processing grid-like data such as images, and Recurrent Neural Networks (RNNs), suitable for sequence data, are key tools in their methodology.?

The use of CNNs for image recognition exemplifies their capability to handle large, complex, and unstructured data, learning hierarchical feature representations autonomously.?

However, the "black box" nature of neural networks and the substantial requirements for data and computational power present notable challenges. The biggest opponents of this school of thought are the xAI cults ( which are still in search of their God )

Evolutionaries invoke the principles of natural selection to evolve solutions over time, employing Genetic Algorithms (GAs), Genetic Programming (GP), and Evolution Strategies (ES) to refine a population of candidate solutions through selection, crossover, and mutation.?

This methodology shines in domains like robotics, where genetic algorithms evolve control strategies for autonomous robots, showcasing the potential to optimize complex problems and navigate deceptive landscapes.?

Yet, the computational intensity and sensitivity to parameter tuning and operator choices still remain as the biggest challenges for this evolutionary approach.

The diverse and some times mutually antagonistic schools of thought in AI/ML — Symbolists with their logical models, Bayesians with probabilistic frameworks, Analogists drawing on past experiences, Connectionists leveraging neural networks, and Evolutionaries optimizing through natural selection — each contribute unique strengths and perspectives to the quest for artificial intelligence. This rich mosaic of methodologies not only highlights the complexity of creating intelligent systems but also offers a comprehensive view of the strategies employed in the pursuit of AI, highlighting the multifaceted and collaborative/competing nature of this fascinating field.

If you are not a member of these schools of thought and if you are generally enjoying the use of KNNs, CNNS, Decision Trees, DQNs and all the ML models in the larger areas of Supervised Learning (Regression, Classification), Unsupervised Learning ( Clustering, Dimensionality Reduction, Association), Reinforcement Learning ( Value-Based, Model-Based, Policy-Based ), Ensemble Learning ( Bagging, Stacking, Boosting ) and also the Neural Networks ( Feed-Forward, Recurrent, Generative and Specialized ), consider yourself as AI/ML Practitioner. Otherwise, you are doomed to be an AI/ML Scientist.

Co-Contributor: Chuan ( August ) Sun