Going deeper than deep learning for massive data analytics under physical constraints
Bita Darvish Rouhani, Azalia Mirhoseini, Farinaz Koushanfar · 2016
Deep Neural Networks (DNNs) are a set of powerful yet computationally complex learning mechanisms that are projected to dominate various artificial intelligence and massive data analytic domains. Physical viability, such as timing, memory, or energy efficiency, are standing challenges in realizing the true potential of DNNs. We propose DeLight, a set of novel methodologies which aim to bring physical constraints as design parameters in the training and execution of DNN architectures. We use physical profiling to bound the network size in accordance to the pertinent platform's characteristics. An automated customization methodology is proposed to adaptively conform the DNN configurations to meet the characterization of the underlying hardware while minimally affecting the inference accuracy. The key to our approach is a new content- and resource-aware transformation of data to a lower-dimensional embedding by which learning the correlation between data samples requires significantly smaller number of neurons. We leverage the performance gain achieved as a result of the data transformation to enable the training of multiple DNN architectures that can be aggregated to further boost the inference accuracy. An accompanying API is also developed, which can be used for rapid prototyping of an arbitrary DNN application customized to the platform. Proof-of concept evaluations for deployment of different imaging, audio, and smart-sensing applications demonstrate up to 100-fold performance improvement compared to the state-of-the-art DNN solutions.