One shot learning for acoustics classification of Malaysia bird species

Abstract

Malaysia is famed for its beautiful bio-diverse forest and its bird species, some of it is still understudied. Using acoustic detection, we can study these bird as current advanced in machine learning application have resulted in cutting edge performance for acoustic classification application. However, most of these applications need large amount of data for prediction to have acceptable accuracy. We, however, do not have this kind of resources. This situation is experienced by large demographic of this country, which provide importance to our studies. Adding to that problem, one common issue that come with studying of species is that we can only know the status of species in a habitat up to certain point. Such problem needs to be solve using methodologies that can cope with the fluidity of information. As such, we propose neural network framework that able to notice any changes in the class categories and learn new classes on the fly. To solve our issue, we seek to design a Siamese Style convolutional Neural Network for one-shot-learning architecture. Additionally, we would train it using base convolutional neural network with low complexity so that it can be realistically implement in hardware of low computing power. We evaluated and benchmarked our framework, showing promising results as the Network is able to classify trained bird species with accuracy of 90% or higher with only around 100 sound clips per bird species. Additionally, it is able to detect new bird species on the fly and add it to its class successfully, however, it still needs some work as the accuracy is around 50% on this part. All of these is achieved using base Convolutional Neural Networks of low complexity with only 4 layers of conv layers and 2 layers of fully connected layers. From this thesis, It is shown that this neural network can work and I am optimistic that this work can be further improved, which can be done by using a higher variety of dataset and transfer learning and a further tweaking of the base neural network architecture.