5 Tribes of AI - The Perspective
The concept of the tribes was introduced by Pedro Domingos in “The Master Algorithm : How the Quest for the Ultimate Learning Machine Will Remake Our World”. It provides a good explanation about the characteristics of each of the tribes and which Machine Learning algorithm falls under what category. In this blog we will try to explain every tribe as simply as possible with an example to understand the tribes better.
The five tribes can be visualised by three parameters -
- Representation: This describes how the algorithm associated with each tribe can be represented.
- Evaluation: This describes how the algorithm is evaluated depending upon the tribe it belongs to.
- Optimisation Driver: This describes how an algorithm can be optimised to achieve optimum result.
The term “Master Equation” in the centre of the diagram above refers to the ultimate goal of creating an algorithm which can possess characteristics of all the tribes.
In the following section, we will explain all the characteristics of each tribe and how the functioning of an algorithm relates to it.basis
The Symbolists
This tribe focuses on symbol manipulation where questions can be presented as equations and can be answered using expressions. They work on the concept of inverse deduction where inverse deduction means starting with certain knowledge and exponentially gaining the latter by answering the aforementioned questions.
Decision Tree
One of the best algorithms to visualise the characteristics of this tribe is Decision Tree. Decision Tree starts with certain knowledge which can be achieved using Information Gain or Gini Index to identify the root node and gathers further information using equations to produce child nodes and the final leaf node (decision).
Based on the three parameters mentioned above the symbolists can be,
Represented: Using logic, which is a tree like structure which makes it easier for humans to interpret.
Evaluated: Using accuracy, which describes how accurate the result of the tree is.
Optimised: Using inverse deduction, where the decision tree uses the concept of pruning.
The Connectionists
This tribe focuses on concepts of how a human brain functions and tries to mimic its functionalities by reverse-engineering and trying to build the neurons in the brain artificially which is referred as Perceptrons. The collection of multiple perceptrons is called an Artificial Neural Network.
Convolutional Neural Networks (CNN)
One of the most commonly used algorithms which relates to the characteristics of this tribe is Convolutional Neural networks, which is a collection of perceptrons that understands and translates data input of one form to the desired output of another form.
Based on the three parameters mentioned above the Connectionist tribe can be,
Represented: Using neural networks, which are layers of perceptrons passing input from one layer to the next.
Evaluated: Using squared error, lower the error, better the performance of the neural network.
Optimised: Using gradient descent, where the neural network uses the concept of weights and adjusts them to reduce the error and optimises the neural network .
The Evolutionaries
The Evolutionaries follow the enhancement of DNA by a natural selection process where a pool of genes (usually a set of instructions), i.e., population. From this population of genes, either two individuals mate to form a new one or one individual mutates to form a the new generation. This is coined crossover and mutation respectively.
Genetic Programming
Genetic programming is a technique of evolving programs following the process of genes, crossover and mutation as described above for a particular task and creates multiple solutions for every program and the best solution is selected.
Based on the three parameters mentioned above the Evolutionary tribe can be,
Represented: Using genetic programming, which is described above.
Evaluated: Using fitness function, chooses the best solution from pool of solutions.
Optimised: Using genetic search, this helps select the most optimised solution for the problem at hand.
The Bayesians
For the Bayesians, all real world data is uncertain and knowledge can be derived using this uncertainity. This uncertainity can be reduced by relating two or more events to formulate incomplete data by identifying the conditional probability between those events.
Bayesian Network
The Bayes Net considers an event occurance and predicts that one of the known causes was the contributing factor. It works on concepts of probabilistic inference and is based on the Bayes Theorem.
These networks are represented using Directed Acyclic Graphs with every node having a probability table.
Based on the three parameters mentioned above the Bayesian tribe can be,
Represented: Using graphical models, such as directed acyclic graphs.
Evaluated: Using posterior probability, which helps determine that an event will happen after all evidence or background information has been taken into account.
Optimised: Using probabilistic inference, which can be defined as the task of deriving the probability of one or more random variables taking a specific value or set of values.
The Analogizers
The analogizers work on concepts of identifying similarities between situations or things. The main challenge is to identify how similar these situations are. Similarity is also one of the central ideas of machine learning.
K-Nearest Neighbour
K-Nearest Neighbour algorithm also known as KNN is one of the common examples for the analogizers. This algorithm calculates the similarities between two items using a distance measure such as Euclidean distance or Manhattan distance.
The K in the KNN is the factor that decides how many datapoints to consider while finding the cluster for that data point.
Based on the three parameters mentioned above the Analogizer tribe can be,
Represented: Using support vectors, these vectors differentiate between two data points for the task at hand.
Evaluated: Using margins, which acts as a boundary line to decide the category of a data point.
Optimised:Using constrained optimization, which is the K factor in KNN.
These are the five tribes of AI with examples and their characteristics. Next time you work on a machine learning problem try to connect the choice of your algorithm to identify and understand the similarities between the characteristics of the tribe and the algorithm.
