State of The Art Machine Learning Methods For Classification of Animals In Ocean Acoustics Data

JASCO是一家为海洋建筑和石油勘探行业提供被动声学监测解决方案的领先供应商，总部位于NS的哈利法克斯。JASCO定期与政府和学术界合作开展音景和海洋哺乳动物栖息地研究项目。作为这些项目的一部分，JASCO收集了连续的声学数据集，这些数据集从一天持续到一年，用于广泛地区的多个记录器。迄今为止，JASCO积累了60万个来自数百种海洋哺乳动物物种和噪音源的“真实数据”注释。JASCO已经确定需要改进其海洋哺乳动物检测和分类算法。目前的检测器/分类器是有效的，但人们期望新的机器学习技术既能提高检测概率，又能减少误报。能够迅速向大学、政府和行业提供可靠的结果是JASCO认为必须不断改进以保持其市场领导者地位的独特产品。虽然JASCO在海洋声学数据方面拥有丰富而广泛的内部专业知识，但它将受益于由达尔豪西大学贡献的一项关于在分类器开发中使用最先进的机器学习专业知识的试点研究。我们将在标记的数据上应用几个机器学习分类器。此外，我们将尝试从深度学习的角度来看，当表示直接从用于给定任务（这里是物种分类）的数据中学习时，已知表现良好的数据重新表示技术。卷积神经网络（CNN）是一种很有前途的方法。我们的假设是，CNN可以产生一个紧凑的，但同时高性能的表示，以供上述机器学习分类器之一使用。最后，很可能由于数据的规模，效率问题可能会出现wrt至少一些themods（如聚类）。如果是这样的话，我们将研究使用基于GPU的解决方案，无论是通过专用的专有资源，或计算加拿大的设施与多个GPU的可用性。

JASCO is a leading provider of passive acoustic monitoring solutions to the marine construction and petroleumexploration industries from our headquarters in Halifax, NS. JASCO regularly partners with government andacademia on soundscape and marine mammal habitat research projects. As part of these projects JASCOcollects continuous acoustic datasets that last from a day to year for multiple recorders over wide areas. To dateJASCO accumulated 600,000 'truth-data' annotations from hundreds of types of marine mammal species andnoise sources. JASCO has identified a need to improve their marine mammal detection and classificationalgorithms. Current detector/classifiers are effective, but the expectation is that new machine learningtechniques will both improve the probability of detection and will reduce false alarms. Being able to providereliable results quickly to universities, government and industry is a unique offering that JASCO believes itmust continually improve to maintain it's market leader position. While JASCO has rich and extensive internalexpertise in ocean acoustic data, it will benefit from a pilot study on the use of state-of-the-art Machinelearning expertise in classifier development, contributed by Dalhousie University. We will apply severalMachine Learning classifiers on the labeled data. Moreover, we will experiment with data re-representationtechniques known to perform well-from a Deep Learning perspective - when the representation is directlylearned from the data used for a given task (here, species classification). Convolutional Neural Networks(CNNs) are a promising approach. The hypothesis is that CNNs can produce a compact but at the same timehighly performing representation to be used by one of the machine learning classifiers mentioned above.Finally, it is likely that due to the scale of the data, efficiency problems may arise wrt at least some of themethods (eg clustering). If that will be the case, we will look into the use of GPU-based solutions either thruthe dedicated proprietary resources, or on Compute Canada's facilities with multiple GPU availability.