Our Research Field
Large-Scale AI and Deep Learning
What Motivates Us
Most advances in novel machine learning algorithms are developed in an academic setting and based on smaller datasets. As a result, these algorithms are not ready for production use on real-world datasets. Scaling so far has been restricted to a select few industrial settings. To efficiently deploy AI at scale, we need to develop distributed machine learning algorithms and pipelines, and work on techniques to efficiently explore multidimensional hyper-parameter spaces.
One way to scale training of models is through multi-task architectures: these architectures exploit the correlation between several tasks to improve performance over individual tasks. As key enabler for wider applicability, multi-task architectures need to be able to learn from noisy, unlabeled or alternatively labeled data sets. Further, we investigate combination of distributed computing paradigms with techniques for efficient hyper-parameter optimization. To scale data preprocessing, we focus on techniques combining multiple sources of data to provide ground truth labels.
Techniques which are able to effectively explore multi-dimensional hyper-parameter spaces are a key enabler across domains. At Bosch, we focus on deployment of these techniques for semantic search capabilities on large-scale video databases. Our research aims to provide accurate frame-level and scene-level tags for video sequences, enabling efficient search of desired scenes.
Zhang et al."Deep learning on symbolic representations for large-scale heterogeneous time-series event..."
- Authors: Shengdong Zhang, Soheil Bahrampour, Naveen Ramakrishnan
- Published in IEEE Xplore in 2017
Dhar et al."Universum Learning for Multiclass SVM"
- Authors: Sauptik Dhar, Naveen Ramakrishnan, Vladimir Cherkassky, Mohak Shah
- Published in SVM in 2016
Dhar et al."ADMM based Scalable Machine Learning on Spark"
- Authors: Sauptik Dhar, Congrui Yi Naveen Ramakrishnan, Mohak Shah
- Published in ADMM in 2015
Heit et al."An architecture for the deployment of statistical models for the big data era"
- Authors: Juergen Heit, Jiayi Liu, Mohak Shah
- Published in Big Data in 2016