SLID: Exploiting Spatial Locality in Input Data as a Computational Reuse Method for Efficient CNN

Abstract

Convolutional Neural Networks (CNNs) revolutionized computer vision and reached the state-of-the-art performance for image processing, object recognition, and video classification. Even though CNN inference is notoriously compute-intensive, as convolutions account for > 90% of the total operation tasks, the ability to tradeoff between accuracy, performance, power, and latency to meet target application makes it an open research topic. This paper proposes the Spatial Locality Input Data (SLID) method for computational reuse during the inference stage for a pre-trained network. The method exploits input data spatial locality via skipping partial processing of the multiply-and-accumulate (MAC) operations for adjacent data and equating its value to previously computed ones. SLID improves the throughput of resource-constrained devices (Internet-of-Things, edge devices) and accelerates computations during the inference phase by reducing the number of MAC operations. Such approximate computing schema does not require a similarity quantification step nor any modification for the training stage. The computational data reuse was evaluated on three well-known distinctive CNN structures and data sets with alternating layer selections: LeNet, CIFAR-10, and AlexNet. The computational data reuse method saves up to 34.9%, 49.84%, and 31.5% of MAC operations while reducing the accuracy by 8%, 3.7%, and 5.0% for the three models mentioned earlier, respectively. Besides, the proposed method saves on memory access by eliminating data fetching of skipped inputs. Furthermore, filter size, strides, and padding on the accuracy and savings of operations are analyzed. SLID is the first work to exploit the input spatial locality for savings on CNN convolution operations with minimal accuracy loss and without memory or computational overhead. This makes it a great option to support intelligence at the edge.

Performance of Image Recognition with Machine Vision: A Systematic Review

“…Igualmente, redes Neuronales convolucionales cuánticas (QOCNN) es un método que implica tomar las circunvoluciones y la agrupación típica de las CNN y traducirlas de forma cuántica [20]. YOLO (You Only Look One) es una técnica de detección de objetos múltiples extremadamente rápida en tiempo real [21]. Finalmente, las redes Neuronales de Picos (SNN) es una técnica donde las neuronas se presentan en forma puntiaguda denominadas redes neuronales de tercera generación [22].…”

“…En cuanto a la categoría de tecnología, se debe precisar aquellos programas que hacen posible la integración de los algoritmos con la implementación y el tratamiento de los datos. Por ejemplo, en [21] se logró alcanzar un buen desempeño en el sistema de detección de aves relacionado con la precisión utilizando la tecnología YOLO. Esta tecnología permitió crear un algoritmo para el aprendizaje automático programado en Python.…”

FORSA: Exploiting Filter Ordering to Reduce Switching Activity for Low Power CNNs

“…• Analyze the impact of the proposed approach on AI CIFAR-10 and examine the impact on the switching activity and reduced energy consumption. Major works in the literature on efficient CNN focus on reducing the complexity through different methods which would reflect on reduced energy, area, and speed-up in performance: Propose techniques to reduce the number of operations: quantization, pruning, fast algorithms, and computational re-use are some examples [2]- [5]. Explore other computing paradigms: In-memory computing (IMC) that is capable of performing computation on the same physical place where data is stored [6].…”

Approximate Down-Sampling Strategy for Power-Constrained Intelligent Systems

“…Therefore, often, alternative approaches, including data reuse [15] and approximate computing [16]- [24], are adopted in conjunction with data quantization to further expand the design space exploration for the specific application. Some of the referenced techniques [16]- [18] adopt operation-level approximations, while others [19]- [24] exploit data dependency across convolutional and auxiliary layers. In particular, the strategies presented in [19] and [20] are based on detection algorithms able to predict negative feature map values.…”