Convolutional neural networks have recently achieved significant breakthroughs in various image classification tasks. However, they are computationally expensive, which can make their feasible implementation on embedded and low-power devices difficult. In this paper convolutional neural network binarization is implemented on GPU-based platforms for real-time inference on resource constrained devices. In binarized networks, all weights and intermediate computations between layers are quantized to +1 and -1, allowing multiplications and additions to be replaced with bit-wise operations between 32-bit words. This representation completely eliminates the need for floating point multiplications and additions and decreases both the computational load and the memory footprint compared to a full-precision network implemented in floating point, making it well-suited for resourceconstrained environments. We compare the performance of our implementation with an equivalent floating point implementation on one desktop and two embedded GPU platforms. Our implementation achieves a maximum speed up of 7.4× with only 4.4% loss in accuracy compared to a reference implementation.

The dataflow concept has been successfully used for modeling and synthesizing signal processing applications since decades, and recently, dataflow has also been discovered to match the computation model of machine learning applications, leading to extremely successful dataflow based application design frameworks. One of the most attractive features of dataflow, especially for signal processing, is related to its formal nature: when properly defined, a dataflow-based application model can be analytically verified for correctness at the stage of application design. This paper proposes VR-PRUNE, a novel dataflow model of computation that is aimed for design of high-performance signal processing software, together with runtime support that allows efficient application deployment to heterogeneous GPU-equipped platforms. Compared to prior work, VR-PRUNE features variable token rate processing, which enables designing adaptive signal processing applications, and implementing solutions that, e.g., allow trading-off between power consumption and filtering bandwidth at runtime. The paper presents the formal concepts of VR-PRUNE, as well as four application examples from domains related to signal processing, accompanied with quantitative results, which show that using VR-PRUNE enables, for example, application power-performance scaling, and on the other hand describing adaptive application behavior with 59% fewer dataflow graph components compared to previous work.

This is a self-archived -parallel published version of this article in the publication archive of the University of Vaasa. It might differ from the original.

No abstract