Machine Learning and Artificial Intelligence

by Francisco Maturana MAusIMM(CP)

What is Machine Learning (ML)?

ML Is a branch of Artificial Intelligence (AI), which aims at developing techniques that allow computers to learn. In other words, ML comprises algorithms capable of generating desirable behaviours and recognise patterns from input data containing relevant examples. A common application of ML is autonomous driving. Early examples are the DARPA Grand Challenges in Nevada, USA (2004 and 2005). The 2005 version[1] had 23 autonomous cars aiming to drive 132 miles in less than 10 hours. The road had bends, tunnels and obstacles. Only five vehicles completed the race. The winner “Stanley’ (Stanford University) completed the task in 6 hours and 53 minutes. Stanley’s training was achieved by driving the vehicle by a human under different road conditions, allowing the system to compile data from a number of sensors. The data was then used to build a model that identified obstacles (unsafe zones). In just 15 minutes of training, Stanley reduced false positives from one in 8 to 1 in 50 thousand. In other words, Stanley learned to drive.

So, how can we understand the concept of learning from a computer point of view? A definition of learning is “an increase, through experience, of problem-solving ability.”[2] Similarly, ML enables a computer to learn from experiences, not by recognising patterns that have been previously programmed. Common scenarios where ML is applied are:

  • Regression: Predict a specific value. For instance, the value of a stock based on its historical behaviour
  • Classification: Classify an object using predefined categories. For example, classify an email as spam, or whether a news is sports, politics, etc.
  • Ranking: Predict the optimal order for a set of objects according to predefined definitions of relevance. A typical example here would be results retrieved from a Google search.

Many types of learning have been implemented for ML techniques. The most common ones are:

Supervised learning: Uses a function to establish the relationship between inputs and outputs by clearly defined examples (“you’ve got labelled data”). Typical uses are regression and classification.

Unsupervised learning: The modelling process is based on a set of examples fed into the system without knowing what a correct classification looks like. In other words, “You’ve got data but it’s not labelled. See if there’s a structure in there”. Typical uses are clustering and feature extraction.

Reinforment learning: In this case, algorithms learn by observing the surrounding world with a continuous flow of information (from the world to the machine and from the machine to the world) applying trial and error and reinforcing those actions that produce positive responses in the world. This option is useful when learning data in not available or there is no time to compile it, or data may change rapidly causing the outcome to change more rapidly than a typical model refresh cycle can accommodate.

Generative learning: In this case, algorithms use two models: a generative model that generates candidates and a discriminative model that evaluates if the candidates come from the data distribution. When the two models are trained, the generative model is capable of generating new synthetic samples. Typical uses include generating synthetic data, combining two datasets and completing missing data

How does ML fit in the world of AI?

AI and ML are expressions used more and more every day. Although often used interchangeably they are not the same. AI is a broader concept referring to an environment where machines are able to perform tasks in an “intelligent” manner. ML, on the other side, is an application of AI based on the idea that machines should be able to take data and learn by themselves using algorithms. In summary, ML is AI’s enabler.

Applications for Geoscience

Considering the examples and cases presented above there are many instances where these techniques would be extremely useful for Geoscientists. Classification is a key task we apply routinely in our jobs. How about reliable core logging without subjective interpretations? Or being able to identify alteration or grade trends within an ore body that are not visible with our standard practices. Data integration and assimilation by ML has also a great potential in combining different sources of data. ML provides an excellent framework to build risk classification models for seismicity and other geotechnical parameters, applications to geometallurgy and grade estimation. The options are countless. However, in order to realise the benefits of ML/AI we would need to overcome some obstacles such as overall lack of relevant skills and lack willingness of organisations to invest in developing those skills. In addition, the topic is not significantly present in industry or formal Geoscience education and training. Finally, negative attitudes towards implementing these technologies could be a serious roadblock. Are we up to the challenge?


[2] Washburne, J. N. (1936). The definition of learning. Journal of Educational Psychology, 27(8), 603-611.


Note: If you would be interested in a webinar in the area of machine learning (for example applications to orebody knowledge) please express your interest via the Secretariat 

Related News


Applied Earth Science - new issue released


Geoscientist Survey: Interdisciplinary interaction in your first 5 years of professional experience

Survey closes 31 July 2019


Watch Ross Large's Distinguished Lecture

Relationship between Oxygenation and Ore Deposit Cycles through Earth History