Speaker
Description
The problem of object recognition is computationally expensive, especially when large amounts of data is involved. Recently, techniques in deep neural networks (DNN) - including convolutional neural networks and residual neural networks - have shown great recognition accuracy compared to traditional methods (artificial neural networks, decision tress, etc.). However, experience reveals that there are still a number of factors that limit scientists from deriving the full performance benefits of large, DNNs. We summarize these challenges as follows: (1) large number of hyperparameters that have to be tuned against the DNN during training phase, leading to several data re-computations over a large design-space, (2) the share volume of data used for training, resulting in prolonged training time, (3) how to effectively utilize underlying hardware (compute, network and storage) to achieve maximum performance during this training phase. In this presentation, we discuss a cross-layer perspective into realizing efficient DNNs for classification of physics objects (in particular, Higgs). We describe how we compose hardware, software and algorithmic components to derive efficient and optimized DNN models that are not only efficient, but can also be rapidly re-purposed for other tasks, such as muon identification, or assignment of transverse momentum to these muons. This work is an extension of the previous work to design a generalized hardware-software framework that simplifies the usage of deep learning techniques in big data problems
