211
Machine learning has already been used within the mining industry. But the application of
this technology to geology, and to geological samples and observations is very limited. This
paper undertakes a high-level review of the machine learning technology that could be applied
to geological data. It highlights areas where the technology could add value to a resource business. And finally, it demonstrates the effectiveness of the technology by presenting a research
project that was able to successfully isolate the areas of core within core tray imagery.
2 AN OVERVIEW OF MACHINE LEARNING
Machine learning algorithms use input data to learn patterns and use inference to build
mathematical models. These models are used to make predictions or decisions on the input
data (Bishop, 2006).
There are four main types of machine learning systems: supervised, semi-supervised, unsupervised, and reinforcement learning. Supervised learning algorithms create models using
training data that contains inputs and labels. The iterative optimization of an objective function allows supervised learning algorithms to learn and predict the output from input data
(Mohri et al, 2012) to the point where it can correctly determine the output for data not in
the initial training dataset.
In unsupervised learning, no labelled data is used. Instead the system attempts to learn by
following the output of functions. Clustering is an example of unsupervised learning where
an algorithm tries to detect similar features in input data and group them into unlabelled
groups. For example, if this style of algorithm was used to sort multiple pictures of a group
of people, the algorithm would group similar images and each group would contain images
of the same person.
The semi-supervised learning is a combination of supervised and unsupervised learning.
Face recognition algorithms are a good example of this ML type. Using our earlier example,
the unsupervised part of the algorithm detects and groups similar human faces from the set of
pictures. If the user then labels the whole set of pre-grouped images, then this second part of the
process is supervised, and the overall classification for the algorithm becomes semi-supervised.
The last type of reinforcement learning is where an algorithm learns by experimentation.
The algorithm requires a scoring function to determine if it is making good or bad decisions.
When this ML type is used to teach an algorithm for a self-driving car, the function gives a
high score when the vehicle is driving within marked road lines, but if the car unintentionally goes outside the lines (it is not deliberately changing lanes) it receives a low score. This
reliance on learning from bad decisions makes it difficult to practice reinforcement learning
methods in the real-physical world. However, there have been some good results using this
type in the virtual world. For example, OpenAI has prepared a reinforcement learning algorithm that can compete with some of the best players in a complex strategic game (https://
blog.openai.com/openai-five/). This opens the possibility of using this ML type in the mine
environment by training an algorithm on high quality digital twins of the mine.
Deep learning algorithms analyse an image by examining it through a series of layers
(sometimes also called filters). The lower layers define the simple features while the higher
layers composite the lower layers to create more abstract features (Bengio, 2015). For example, the first presentation layer may abstract the pixels and encode edges within the image.
The next layer might combine arrangements of edges together. The third layer might encode
the combined edges into wheels, a cab, or a tray. The final layer might recognise the image
contains a haul truck.
Most deep learning algorithms are based on artificial neural networks. The design of artificial neural networks was inspired from the biological neural networks of animal brains
(Coolen et al., 2005). An artificial neural network contains connected nodes called artificial
neurons which are analogous to the neurons in a biological brain. Just like the synapses in a
biological brain, the connections transmit a signal from one node to another. The signal is
normally a real number, and the node computes a sum of the inputs with a non-linear function. The connections are called edges, and the nodes and the edges are weighted to adjust the
Machine learning has already been used within the mining industry. But the application of
this technology to geology, and to geological samples and observations is very limited. This
paper undertakes a high-level review of the machine learning technology that could be applied
to geological data. It highlights areas where the technology could add value to a resource business. And finally, it demonstrates the effectiveness of the technology by presenting a research
project that was able to successfully isolate the areas of core within core tray imagery.
2 AN OVERVIEW OF MACHINE LEARNING
Machine learning algorithms use input data to learn patterns and use inference to build
mathematical models. These models are used to make predictions or decisions on the input
data (Bishop, 2006).
There are four main types of machine learning systems: supervised, semi-supervised, unsupervised, and reinforcement learning. Supervised learning algorithms create models using
training data that contains inputs and labels. The iterative optimization of an objective function allows supervised learning algorithms to learn and predict the output from input data
(Mohri et al, 2012) to the point where it can correctly determine the output for data not in
the initial training dataset.
In unsupervised learning, no labelled data is used. Instead the system attempts to learn by
following the output of functions. Clustering is an example of unsupervised learning where
an algorithm tries to detect similar features in input data and group them into unlabelled
groups. For example, if this style of algorithm was used to sort multiple pictures of a group
of people, the algorithm would group similar images and each group would contain images
of the same person.
The semi-supervised learning is a combination of supervised and unsupervised learning.
Face recognition algorithms are a good example of this ML type. Using our earlier example,
the unsupervised part of the algorithm detects and groups similar human faces from the set of
pictures. If the user then labels the whole set of pre-grouped images, then this second part of the
process is supervised, and the overall classification for the algorithm becomes semi-supervised.
The last type of reinforcement learning is where an algorithm learns by experimentation.
The algorithm requires a scoring function to determine if it is making good or bad decisions.
When this ML type is used to teach an algorithm for a self-driving car, the function gives a
high score when the vehicle is driving within marked road lines, but if the car unintentionally goes outside the lines (it is not deliberately changing lanes) it receives a low score. This
reliance on learning from bad decisions makes it difficult to practice reinforcement learning
methods in the real-physical world. However, there have been some good results using this
type in the virtual world. For example, OpenAI has prepared a reinforcement learning algorithm that can compete with some of the best players in a complex strategic game (https://
blog.openai.com/openai-five/). This opens the possibility of using this ML type in the mine
environment by training an algorithm on high quality digital twins of the mine.
Deep learning algorithms analyse an image by examining it through a series of layers
(sometimes also called filters). The lower layers define the simple features while the higher
layers composite the lower layers to create more abstract features (Bengio, 2015). For example, the first presentation layer may abstract the pixels and encode edges within the image.
The next layer might combine arrangements of edges together. The third layer might encode
the combined edges into wheels, a cab, or a tray. The final layer might recognise the image
contains a haul truck.
Most deep learning algorithms are based on artificial neural networks. The design of artificial neural networks was inspired from the biological neural networks of animal brains
(Coolen et al., 2005). An artificial neural network contains connected nodes called artificial
neurons which are analogous to the neurons in a biological brain. Just like the synapses in a
biological brain, the connections transmit a signal from one node to another. The signal is
normally a real number, and the node computes a sum of the inputs with a non-linear function. The connections are called edges, and the nodes and the edges are weighted to adjust the