In this paper we present a convolutional neural network architecture that supports user customization through incremental transfer learning. The architecture consists of a large basic inference engine and a small augmenting engine. After training the basic inference engine and augmenting engine on a large general dataset, the basic inference engine is fixed. For user customization, only the augmenting engine is re-trained on-device using a small user specific dataset provided by the user. To accelerate the training of the augmenting engine we map this to a coarsegrained reconfigurable array processor. The complete network architecture is evaluated using the Caffe framework, and a C-code equivalent network is implemented and tested on a CGRA processor. Experiments with NIST'19 and our user-specific datasets show an increase in accuracy of the system from 76.3% to 93.2% after user customization. Mapping this code to a CGRA gives us a speed up of 45x and a 49-and 3-fold reduced energy consumption over an ARMv7 processor and a 3-way VLIW processor, respectively, showing the potential of CGRAs as DNN processors.

Modulo-scheduled course-grain reconfigurable array (CGRA) processors excel at exploiting loop-level parallelism at a high performance per watt ratio. The frequent reconfiguration of the array, however, causes between 25% and 45% of the consumed chip energy to be spent on the instruction memory and fetches therefrom. This article presents a hardware/software codesign methodology for such architectures that is able to reduce both the size required to store the modulo-scheduled loops and the energy consumed by the instruction decode logic. The hardware modifications improve the spatial organization of a CGRA’s execution plan by reorganizing the configuration memory into separate partitions based on a statistical analysis of code. A compiler technique optimizes the generated code in the temporal dimension by minimizing the number of signal changes. The optimizations achieve, on average, a reduction in code size of more than 63% and in energy consumed by the instruction decode logic by 70% for a wide variety of application domains. Decompression of the compressed loops can be performed in hardware with no additional latency, rendering the presented method ideal for low-power CGRAs running at high frequencies. The presented technique is orthogonal to dictionary-based compression schemes and can be combined to achieve a further reduction in code size.

This paper presents a convolutional neural network architecture that supports transfer learning for user customization. The architecture consists of a large basic inference engine and a small augmenting engine. Initially, both engines are trained using a large dataset. Only the augmenting engine is tuned to the user-specific dataset. To preserve the accuracy for the original dataset, the novel concept of quality factor is proposed. The final network is evaluated with the Caffe framework, and our own implementation on a coarse-grained reconfigurable array (CGRA) processor. Experiments with MNIST, NIST'19, and our user-specific datasets show the effectiveness of the proposed approach and the potential of CGRAs as DNN processors.

There has been a body of research to use stochastic computing (SC) for the implementation of neural networks, in the hope that it will reduce the area cost and energy consumption. However, no working neural network system based on stochastic computing has been demonstrated to support the viability of SC-based deep neural networks in terms of both recognition accuracy and cost/energy efficiency. In this demonstration we present an SC-based deep nenural network system that is highly accurate and efficient. Our system takes an input image and processes it with a convolutional neural network implemented on an FPGA using stochastic computing to recognize the input image, with nearly the same accuracy as conventional binary implementations.

Coarse-Grained Reconfigurable Architectures (CGRAs) have become popular in recent times as the increased transistor densities have enabled greater integration of increasingly complex “compute cores”. These devices pack massive compute power and can be effectively used to build efficient solutions for applications which have a significant degree of parallelism. In many cases, these CGRAs are also partially reconfigurable. Clearly to make effective use of these highly “parallel compute platforms”, a good mapping flow is required to map the parallelism that is present in a target application.

Programmable hardware built on a regular architecture can partially alleviate the problem of increased defect densities associated with transistor scaling by dynamically wiring around the defects [1]. The fine granularity of FPGAs is however unsuitable for effectively exploiting runtime reconfiguration because of the high overheads involved. A coarse grain reconfigurable array with malleable communication links - reMORPH - is proposed in this paper. The compute tile uses DSP48E and BRAM embedded blocks in a Xilinx FPGA and has a very low footprint of about 200 slice LUTs. The semi-systolic near neighbour communication interconnect can be dynamically reconfigured for each “epoch” of computation. The “epoch” or phases of the application are obtained via profiling or static data flow analysis. Some of the links between the compute tiles are changed during the reconfiguration phase which drastically reduces the context switch overhead enabling high performance/area applications to be built on this fabric.