Work

Application Specific Domain Co-design Hardware Accelerator IP for Deep Learning Enabled Internet-of-Things

Year: 2022
Type: dissertation
Abstract: Currently, DeepLearning (DL) is becoming ubiquitous in everyday life.In particular, Convolutional Neural Networks (CNNs) are already present in many applications related to computer vision ranging from medicine to autonomous driving, neural language processing, and finance.However, these algorithms employ very deep networks to achieve impressive performance, requiring a significant computational power, both during the training and inference time.A single inference of a DL model may require billions of multiply-and-accumulated operations, making the DL extremely compute-and energy-hungry.The growing desire to adopt these algorithms in an increasing number of applications and the evolution of hardware devices is recently shifting the inference process from the cloud to the edge, i.e., on embedded systems with limited resources.In addition, this paradigm shift can offer the user lower latency and greater privacy.In-device CNN processing presents several benefits.Since the data does not have to be transmitted to a central processing unit, it remains local, ensuring better user privacy and more deterministic latency.In addition, the absence of transmission reduces power consumption and the need for a network infrastructure capable of handling large amounts of data.At the same time, however, edge computing demands addressing the gap between the computational and memory-hungry CNNs requirements and limited hardware and energy-constrained embedded platforms.Consequently, the need for cost-effective hardware platforms and co-design optimization techniques capable of implementing energy-efficient DL execution arises.Achieving the objective of enabling minimal resource systems for CNNs algorithms relies on the ability to optimize and co-design accelerator architectures.This dissertation proposes an in-depth analysis and design development for the acceleration of CNN algorithms for edge computing.The analysis covers all critical Contents List of Figures x List of Tables xv Pooling layers.Generally, Pooling layers come after Conv layers downsampling feature maps by summarizing statistics in their receptive fields.For the group of neurons in each receptive field, they yield a single value featuring a specific statistic of the group such as the average or the maximum value, Figure 1.9.In the Pooling case, the stride is usually equal to the size of the receptive field, providing non-overlapping windows.Gradient backpropagation is the most commonly used supervised-learning algorithm for training DNNs.The input sample is fed into the DNN that performs an inference.The output prediction is then compared to the expected output, and a loss function, such as Euclidean distance or Mean Squared Error (MSE), calculates the
Source: PORTO@Iris
Language: English
Open Access status: green
OpenAlex ID: W7216662823
Cites: 105
Cited by:
Related to: 10