5 Tribes of Machine Learning

machine Learning has many facets. Learning techniques are being shown to be extremely beneficial to various disciplines. When machine learning is applied to a particular discipline such as Biology or Psychology for instance it generally takes on a shape which is in accordance with the particular discipline.

The book Pedro Domingos’ book, The Master Algorithm: How the Quest of the Ultimate Learning Machine Will Change the Way We See Our World In it, he divides the various kinds of machine learning algorithms into five distinct classes and calls them machines learning tribes. Each tribe is based on the principles of each group, and from them are derived different models of machine learning.

One to many, the method is concluded with the statement every tribe would converge on a single master algorithm that can be the ultimate source of knowledge. This artificial intelligence algorithm known as known as the Master Algorithm is an AI general intelligence that many believe will arrive. To get there, Domingos wants to take some of each of the five tribes and mix them up.

The Five Tribes of Machine Learning:

  1. Symbolists
  2. Connectionists
  3. Bayesians
  4. Evolutionaries
  5. Analogizers

Symbolists

Symbolists are focused on logic. They create formal systems that create AI. From the 1950s until the 1980s in the 1980s, the concept of symbolic AI was the most popular model. A majority of computers were based on this type of thinking and over the years it was the one that was most easily coded. Symbolists employ technologies which follow a predetermined route to take an informed decision. The rules they coded in their hard-coded form would be like “Always leave when you see a stop signal”.

Systems that facilitate this type of decision-making are:

  • Decision trees
  • Random decision forests
  • Production rule systems
  • Programming inductively

Connectionists (Neuroscience)

Connectionists are the group who create models that are based in the brain. They like to highlight they believe that neural networks are a crucial component of their machine learning structure, are modeled on the way that neurons function within the brain. They make use of models such as:

  • Artificial neural nets
  • Reinforcement learning
  • Deep Learning

When they talk about connections, they don’t use analogies to describe connections as a kitten to a cat, just as the puppy is to the dog. For Connectionists, they employ connections to refer to the signals that flow between neurons. There are strengths to signals as well as the number of signals.

One of the major issues people face when using connectionist framework is Connectionist model is the method by which the decisions are made is concealed from view- the white box. Every step that is made in a tree of decision is publicly known. In Connectionism it is possible for an input to be able to produce an output, but the route that input traveled to reach the output isn’t known. The user is not able to ask “How does the modeling arrive at that conclusion?”

Bayesians (Statisticians)

The Bayesian tribe is a fan of statistics. It’s a different kind of logic that is based on probabilistic outcomes rather than in hard coded or either/or results. For Bayesians Machine learning can be described as a type of inference based on probabilities. If you input a data point there is a chance that the result will be Y. The outcome is not known; however, the possibilities are there for a possibility of occurring.

Bayesians could employ Connectionist as well as Symbolist methods to create decision models, however their primary focus is on having probabilistic outcomes. The structure to go to X towards Y’s conclusion is not nearly as crucial as how the results are interpreted. But, the Bayesians are able to use models that include:

  • Hidden Markov chains
  • Graphical models
  • Causal inference


(Source)

Evolutionaries (Biologists)

Evolutionists are fond of thinking of things that are not infrastructure or predicted outcomes, or even the interpretations of results. They are the type of people who believe in process theory. They’re concerned with processes and steps.

From the perspective of biologists, they are interested in the evolution in an AI. They’re interested in the way it develops, mutates…with how it becomes. Much like how people change with each change, the human genome is also able to have the capacity to adapt and learn and create a human that is better equipped to deal with its environment in every direction. Evolutionists seek to design machines that can, as well, can learn and evolve.

Evolutionaries employ genetic algorithms or evolutionary programs. One common use for Evolutionary AI is on learning tasks. Below is an example of how you can see how the DeepMind Team made a small stick figure that had legs, and provided it with an environment and a simple physical principles, and set the aim to move forward. After numerous trials and trial and improved to the point that it could leap over obstacles and advance forward. The use of evolutionary algorithms is extensively in game play. They’ll be among the most fundamental models to get automobiles operating independently.

Analogizers (Psychologists)

Analogizers are generally the storytellers. They are able to construct groups of entities. In the event that an input either old or new is identified as being part of one of these classes Analogizers think they are able to predict the outcome of the input similar to the outcomes of the class.

Problems that can be solved using this method are recommendeder systems. For instance If a person were to stream a romantic comedy on Netflix the Analogizer will tell the user, “Given your watching behavior and the fact that you enjoyed this romantic comedy, you’ll enjoy these other collections of films since other viewers who watched the same movie are also interested in these films.” Analogizers employ types and classes to determine certain groups of people and forecast future outcomes on the basis of the results of others in the class.

Analogizers’ machine learning models they employ are K-Nearest Neighbor algorithm along with SVMs. which are unsupervised learning methodsthat place people in their classes. The chart below places individuals into two groups:, namely 0 and 1. The second chart places a member in any of the three categories.

Five ML Tribes

Each method is useful in its own way. A combination of the above Machine Learning methods might be the solution for an AGI. Self-driving vehicles may learn to safely navigate the road using the Evolutionary method, however they’ll use Connectionist methods to provide the sensors of the car sight. The vehicle could get an immense boost in driver-user interactions by using Analogizer’s method of categorizing its drivers as types such as aggressive, defensive or passive. There are also guidelines to follow including the stop sign to stop and observing the rules which is where the techniques of the old Symbolist tribe can prove useful.

Leave a Reply

Your email address will not be published. Required fields are marked *