Machine Learning

A dive into Machine Learning algorithms and Artificial Intelligence systems

What is Machine Learning?

Generally, a machine follows instructions given it by humans to carry out work in a more efficient manner. Machine Learning describes the process in which an algorithm performs a task without requiring explicit instructions for each step. Instead, the algorithm learns and improves through experience and through finding patterns in the data. Machine Learning algorithms can therefore be highly adaptable to complex tasks and require minimal human interaction, improving the speed and efficiency of previous exclusively manual tasks.

Are Machine Learning and Artificial Intelligence the same thing?

Machine Learning (ML) and Artificial Intelligence (AI) are interrelated tools, and the terms are often considered interchangeable. More accurately, ML falls under the broader umbrella of AI. The larger concept of AI refers to machines that can mimic human intelligence and capabilities, compared to ML which is a tool that enables the machine to perform tasks by extracting and using information from data, therefore requiring less human involvement. We can create AI systems that do not use ML, such as those that respond to user input based on a set of rules (like many simple computer game opponents); however, in more complex tasks in which lots of data is available, ML can result in more efficient and adaptable algorithms.

What are the benefits of ML and AI for coral reef science?

With the advancement of AI applications in computer vision and robotics, AI is gaining tremendous popularity in the field of coral reef science. Marine biologists are now using ML algorithms to automatically identify and map coral benthos from captured images based on their characteristic features. We can also assess the health condition of coral reefs over time through comparing data from the same over successive surveys. Another example includes the use of ML algorithms to detect the degree of damage or recovery to coral reefs following bleaching events, where data includes photographs taken before and after a bleaching event.

Why use Machine Learning to monitor coral reefs?

Coral reef scientists have traditionally been limited in the amount of data they can collect — due to restrictions on how long the human body can stay underwater! Now, thanks to the availability of high-quality underwater cameras, collecting data in the field is much easier and much more data can be collected. Analysing these photos in the lab can be a slow and painstaking process, so while much more data can be collected, scientists are limited by the time it takes to process them. The use of ML algorithms to learn from experts to rapidly identify benthic organisms from photo imagery has vastly improved the efficiency of what we do, where what may have taken a coral reef scientist 10 to 15 minutes to annotate now takes the machine a few seconds!

What type of Machine Learning does ReefCloud use?

Image segmentation is a computer vision technique used to understand what is in each image at a pixel level. Image segmentation algorithms assign class labels to every pixel in an image based on their appearance and contextual information from the surrounding pixels. It is different than image classification, which assigns a label to an entire image; or object detection, which locates objects within an image by drawing a bounding box around them (sometimes used to classify fish from video data).

The benthic communities which make up coral reefs are complex 3D structures, often growing on and around one another. This means a single image often contains multiple individual coral colonies which cannot be easily separated by a bounding box. Segmenting regions in an image which contain different coral colonies gives us a much more nuanced view of the composition of the image, and therefore the reefs we are surveying. It is generally not feasible to regularly perform segmentation of images if we are only relying on manual analysis, however automated segmentation algorithms make the routine use of segmentation possible.

What is new about the Machine Learning available in ReefCloud?

ReefCloud can re-inference your photos extremely fast, converting image patches (256 x 256 =~65k pixels) into a feature vector that is 128 elements long. These feature vectors contain enough information for the model to classify the patch and are why it is so much faster to train and infer classifications than other methods.