The limited and dynamically varied resources on edge devices motivate us to deploy an optimized deep neural network that can adapt its sub-networks to fit in different resource constraints. However, existing works often build sub-networks through searching different network architectures in a hand-crafted sampling space, which not only can result in a subpar performance but also may cause on-device re-configuration overhead. In this paper, we propose a novel training algorithm, Dynamic REal-time Sparse Subnets (DRESS). DRESS samples multiple sub-networks from the same backbone network through row-based unstructured sparsity, and jointly trains these sub-networks in parallel with weighted loss. DRESS also exploits strategies including parameter reusing and row-based fine-grained sampling for efficient storage consumption and efficient on-device adaptation. Extensive experiments on public vision datasets show that DRESS yields significantly higher accuracy than state-of-the-art sub-networks.
Low-Power Edge-AI capabilities are essential for on-device extended reality (XR) applications to support the vision of Metaverse. In this work, we investigate two representative XR workloads: (i) Hand detection and (ii) Eye segmentation, for hardware design space exploration. For both applications, we train deep neural networks and analyze the impact of quantization and hardware specific bottlenecks. Through simulations, we evaluate a CPU and two systolic inference accelerator implementations. Next, we compare these hardware solutions with advanced technology nodes. The impact of integrating state-of-the-art emerging non-volatile memory technology (STT/SOT/VGSOT MRAM) into the XR-AI inference pipeline is evaluated. We found that significant energy benefits (>=80%) can be achieved for hand detection (IPS=40) and eye segmentation (IPS=6) by introducing non-volatile memory in the memory hierarchy for designs at 7nm node while meeting minimum IPS (inference per second). Moreover, we can realize substantial reduction in area (>=30%) owing to the small form factor of MRAM compared to traditional SRAM.
Augmented Reality/Virtual Reality (AR/VR) glasses are widely foreseen as the next generation computing platform. AR/VR glasses are a complex "system of systems" which must satisfy stringent form factor, computing-, power- and thermal- requirements. In this paper, we will show that a novel distributed on-sensor compute architecture, coupled with new semiconductor technologies (such as dense 3D-IC interconnects and Spin-Transfer Torque Magneto Random Access Memory, STT-MRAM) and, most importantly, a full hardware-software co-optimization are the solutions to achieve attractive and socially acceptable AR/VR glasses. To this end, we developed a semi-analytical simulation framework to estimate the power consumption of novel AR/VR distributed on-sensor computing architectures. The model allows the optimization of the main technological features of the system modules, as well as the computer-vision algorithm partition strategy across the distributed compute architecture. We show that, in the case of the compute-intensive machine learning based Hand Tracking algorithm, the distributed on-sensor compute architecture can reduce the system power consumption compared to a centralized system, with the additional benefits in terms of latency and privacy.
Deploying Deep Neural Networks in low-power embedded devices for real time-constrained applications requires optimization of memory and computational complexity of the networks, usually by quantizing the weights. Most of the existing works employ linear quantization which causes considerable degradation in accuracy for weight bit widths lower than 8. Since the distribution of weights is usually non-uniform (with most weights concentrated around zero), other methods, such as logarithmic quantization, are more suitable as they are able to preserve the shape of the weight distribution more precise. Moreover, using base-2 logarithmic representation allows optimizing the multiplication by replacing it with bit shifting. In this paper, we explore non-linear quantization techniques for exploiting lower bit precision and identify favorable hardware implementation options. We developed the Quantization Aware Training (QAT) algorithm that allowed training of low bit width Power-of-Two (PoT) networks and achieved accuracies on par with state-of-the-art floating point models for different tasks. We explored PoT weight encoding techniques and investigated hardware designs of MAC units for three different quantization schemes - uniform, PoT and Additive-PoT (APoT) - to show the increased efficiency when using the proposed approach. Eventually, the experiments showed that for low bit width precision, non-uniform quantization performs better than uniform, and at the same time, PoT quantization vastly reduces the computational complexity of the neural network.
In this era of machine learning models, their functionality is being threatened by adversarial attacks. In the face of this struggle for making artificial neural networks robust, finding a model, resilient to these attacks, is very important. In this work, we present, for the first time, a comprehensive analysis of the behavior of more bio-plausible networks, namely Spiking Neural Network (SNN) under state-of-the-art adversarial tests. We perform a comparative study of the accuracy degradation between conventional VGG-9 Artificial Neural Network (ANN) and equivalent spiking network with CIFAR-10 dataset in both whitebox and blackbox setting for different types of single-step and multi-step FGSM (Fast Gradient Sign Method) attacks. We demonstrate that SNNs tend to show more resiliency compared to ANN under black-box attack scenario. Additionally, we find that SNN robustness is largely dependent on the corresponding training mechanism. We observe that SNNs trained by spike-based backpropagation are more adversarially robust than the ones obtained by ANN-to-SNN conversion rules in several whitebox and blackbox scenarios. Finally, we also propose a simple, yet, effective framework for crafting adversarial attacks from SNNs. Our results suggest that attacks crafted from SNNs following our proposed method are much stronger than those crafted from ANNs.
Spiking Neural Networks (SNNs) has recently emerged as a prominent neural computing paradigm. However, the typical shallow spiking network architectures have limited capacity for expressing complex representations, while training a very deep spiking network have not been successful so far. Diverse methods have been proposed to get around this issue such as converting off-line trained deep Artificial Neural Networks (ANNs) to SNNs. However, ANN-to-SNN conversion scheme fails to capture the temporal dynamics of a spiking system. On the other hand, it is still a difficult problem to directly train deep SNNs using input spike events due to the discontinuous and non-differentiable nature of the spike signals. To overcome this problem, we propose using differentiable (but approximate) activation for Leaky Integrate-and-Fire (LIF) spiking neurons to train deep convolutional SNNs with input spike events using spike-based backpropagation algorithm. Our experiments show the effectiveness of the proposed spike-based learning strategy on state-of-the-art deep networks (VGG and Residual architectures) by achieving the best classification accuracies in MNIST, SVHN and CIFAR-10 datasets compared to other SNNs trained with spike-based learning. Moreover, we analyze sparse event-driven computations to demonstrate the efficacy of proposed SNN training method for inference operation in the spiking domain.
Deep convolutional neural network (DCNN) based supervised learning is a widely practiced approach for large-scale image classification. However, retraining these large networks to accommodate new, previously unseen data demands high computational time and energy requirements. Also, previously seen training samples may not be available at the time of retraining. We propose an efficient training methodology and incrementally growing a DCNN to allow new classes to be learned while sharing part of the base network. Our proposed methodology is inspired by transfer learning techniques, although it does not forget previously learned classes. An updated network for learning new set of classes is formed using previously learned convolutional layers (shared from initial part of base network) with addition of few newly added convolutional kernels included in the later layers of the network. We evaluated the proposed scheme on several recognition applications. The classification accuracy achieved by our approach is comparable to the regular incremental learning approach (where networks are updated with new training samples only, without any network sharing).
Convolutional Neural Networks (CNN) are being increasingly used in computer vision for a wide range of classification and recognition problems. However, training these large networks demands high computational time and energy requirements; hence, their energy-efficient implementation is of great interest. In this work, we reduce the training complexity of CNNs by replacing certain weight kernels of a CNN with Gabor filters. The convolutional layers use the Gabor filters as fixed weight kernels, which extracts intrinsic features, with regular trainable weight kernels. This combination creates a balanced system that gives better training performance in terms of energy and time, compared to the standalone CNN (without any Gabor kernels), in exchange for tolerable accuracy degradation. We show that the accuracy degradation can be mitigated by partially training the Gabor kernels, for a small fraction of the total training cycles. We evaluated the proposed approach on 4 benchmark applications. Simple tasks like face detection and character recognition (MNIST and TiCH), were implemented using LeNet architecture. While a more complex task of object recognition (CIFAR10) was implemented on a state of the art deep CNN (Network in Network) architecture. The proposed approach yields 1.31-1.53x improvement in training energy in comparison to conventional CNN implementation. We also obtain improvement up to 1.4x in training time, up to 2.23x in storage requirements, and up to 2.2x in memory access energy. The accuracy degradation suffered by the approximate implementations is within 0-3% of the baseline.
Large-scale artificial neural networks have shown significant promise in addressing a wide range of classification and recognition applications. However, their large computational requirements stretch the capabilities of computing platforms. The fundamental components of these neural networks are the neurons and its synapses. The core of a digital hardware neuron consists of multiplier, accumulator and activation function. Multipliers consume most of the processing energy in the digital neurons, and thereby in the hardware implementations of artificial neural networks. We propose an approximate multiplier that utilizes the notion of computation sharing and exploits error resilience of neural network applications to achieve improved energy consumption. We also propose Multiplier-less Artificial Neuron (MAN) for even larger improvement in energy consumption and adapt the training process to ensure minimal degradation in accuracy. We evaluated the proposed design on 5 recognition applications. The results show, 35% and 60% reduction in energy consumption, for neuron sizes of 8 bits and 12 bits, respectively, with a maximum of ~2.83% loss in network accuracy, compared to a conventional neuron implementation. We also achieve 37% and 62% reduction in area for a neuron size of 8 bits and 12 bits, respectively, under iso-speed conditions.
Multilayered artificial neural networks (ANN) have found widespread utility in classification and recognition applications. The scale and complexity of such networks together with the inadequacies of general purpose computing platforms have led to a significant interest in the development of efficient hardware implementations. In this work, we focus on designing energy efficient on-chip storage for the synaptic weights. In order to minimize the power consumption of typical digital CMOS implementations of such large-scale networks, the digital neurons could be operated reliably at scaled voltages by reducing the clock frequency. On the contrary, the on-chip synaptic storage designed using a conventional 6T SRAM is susceptible to bitcell failures at reduced voltages. However, the intrinsic error resiliency of NNs to small synaptic weight perturbations enables us to scale the operating voltage of the 6TSRAM. Our analysis on a widely used digit recognition dataset indicates that the voltage can be scaled by 200mV from the nominal operating voltage (950mV) for practically no loss (less than 0.5%) in accuracy (22nm predictive technology). Scaling beyond that causes substantial performance degradation owing to increased probability of failures in the MSBs of the synaptic weights. We, therefore propose a significance driven hybrid 8T-6T SRAM, wherein the sensitive MSBs are stored in 8T bitcells that are robust at scaled voltages due to decoupled read and write paths. In an effort to further minimize the area penalty, we present a synaptic-sensitivity driven hybrid memory architecture consisting of multiple 8T-6T SRAM banks. Our circuit to system-level simulation framework shows that the proposed synaptic-sensitivity driven architecture provides a 30.91% reduction in the memory access power with a 10.41% area overhead, for less than 1% loss in the classification accuracy.