Data, Predictions, and Machine Learning
Big Data Analytics is concerned with generating insights from large amounts of operational data, often hundreds of gigabyte in size, which then form the data for decision-making processes. Only very recently, Big Data Analytics became a reality with the abundance of low-cost high-performance sensors and new computing paradigms such as e.g., GPU computing.
Right: Performance evaluation of various failure mode predictors
Left: Activation-supervised training of a Deep Neural Network
Big Data Analytics and Deep Learning.
Many real-world analytics applications rely on machine learning and signal processing on very large amounts of data. The basic idea is that a representative model is directly derived from the data rather some abstract hypothesis. The underlying data can be real-time data collected from sensors (e.g., 3D cameras or vibration sensors), video signals, or very large natural language corpora (e.g., social media data).
We leverage GPU-computing, outperforming traditional setups by several orders of magnitude, to effectively train deep neural networks and perform signal processing.
Current involvements and projects.
Natural-Language Processing. Recurrent Neural Networks are a powerful tool to process natural language. Typical applications include sentiment analysis and entity tagging, and have been successfully used to predict user behavior from social-media data.
Predictive Maintenance. Deep architectures are also a very powerful tool to capture machine/asset deterioration and predict failure modes and failure times. This data can then be leveraged in decision-making to enable proactive maintenance and streamline spare-parts logistics.