Deep Blue Sea: Using Deep Learning to Detect Hundreds of Different Plankton Species
As part of a graduate Machine Learning course at the University of Amsterdam, students were tasked with this challenge. My team’s model finished in 3rd place among 30 other teams, with an accuracy score of 77.6% — a difference of ~0.005 between our score and the winning team.
I decided to share a bit of our strategy, our model, challenges and tools we used. We will go through the items below:
I will go ahead and assume that you are already familiar with the concept of Artificial Neural Networks (ANN). If that is the case, the concept of Deep Neural Networks is fairly simple. If that is not the case, this video has a great and detailed explanation about how it works.
Deep Neural Networks (DNN) are a specific group of ANNs that are characterised by having a significant amount of hidden layers, among other aspects.
The fact that DNNs have many layers allows them to perform well in classification or regression tasks involving pattern recognition problems within complex, unstructured data, such as image, video and audio recognition.
Convolutional Neural Networks (CNNs) are a specific architecture type of DNNs which add the concept of Convolutional Layers. These layers are extremely important to extract relevant features from images to perform image classification.
This is done by applying different kinds of transformations across each of the image’s RGB channels. So in layman terms, we can say that Convolution operations perform transformations over image channels in order to make it easier for a neural network to detect certain patterns.
While getting into the theoretical and mathematical details about Convolutional Layers is beyond the scope of this post, I can recommend this chapter from a landmark book on Deep Learning by Goodfellow et al to better understand not only Convolution Layers, but all aspects related to Deep Learning.
The picture below shows a glimpse of how a CNN is structured:
Activation functions are also important in this context, and in the case of DNNs, some of the activation functions that are commonly used are ReLU, tanh, PReLU, LReLU, Softmax and others.
These functions represented a huge shift from classical ANNs, which used to rely on the Sigmoid function for activation. This type of activation function is known to suffer from the Vanishing Gradient problem; rectifying functions such as ReLU present one of the possible solutions for this issue.
Now going back to our problem. The data is composed by training and testing sets. In the training set, there are ~24K images of plankton and in the testing set there are ~6K images of plankton.
A first look at the data shows that we have an imbalanced dataset — the plot below shows how many images from each species we have in the training set.
This is a problem, specially for species that are under represented. Most likely our model would not have enough training data to be able to detect plankton images from these classes.
The approach we used to tackle this problem was to set class weights for imbalanced classes. Basically, if we have image classes A and B, where A is underrepresented and B is overrepresented, we could want to treat every instance of class A as 50 instances of class B.
This means that in our loss function, we assign higher value to these instances.
