Manan Suri

Vivek Parmar, Syed Shakib Sarwar, Ziyun Li et al.

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 (>=24%) can be achieved for hand detection (IPS=10) and eye segmentation (IPS=0.1) 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.

Khushal Sethi, Vivek Parmar, Manan Suri

Deep learning research has generated widespread interest leading to emergence of a large variety of technological innovations and applications. As significant proportion of deep learning research focuses on vision based applications, there exists a potential for using some of these techniques to enable low-power portable health-care diagnostic support solutions. In this paper, we propose an embedded-hardware-based implementation of microscopy diagnostic support system for PoC case study on: (a) Malaria in thick blood smears, (b) Tuberculosis in sputum samples, and (c) Intestinal parasite infection in stool samples. We use a Squeeze-Net based model to reduce the network size and computation time. We also utilize the Trained Quantization technique to further reduce memory footprint of the learned models. This enables microscopy-based detection of pathogens that classifies with laboratory expert level accuracy as a standalone embedded hardware platform. The proposed implementation is 6x more power-efficient compared to conventional CPU-based implementation and has an inference time of $\sim$ 3 ms/sample.

Sai Sukruth Bezugam, Ahmed Shaban, Manan Suri

EMG (Electromyograph) signal based gesture recognition can prove vital for applications such as smart wearables and bio-medical neuro-prosthetic control. Spiking Neural Networks (SNNs) are promising for low-power, real-time EMG gesture recognition, owing to their inherent spike/event driven spatio-temporal dynamics. In literature, there are limited demonstrations of neuromorphic hardware implementation (at full chip/board/system scale) for EMG gesture classification. Moreover, most literature attempts exploit primitive SNNs based on LIF (Leaky Integrate and Fire) neurons. In this work, we address the aforementioned gaps with following key contributions: (1) Low-power, high accuracy demonstration of EMG-signal based gesture recognition using neuromorphic Recurrent Spiking Neural Networks (RSNN). In particular, we propose a multi-time scale recurrent neuromorphic system based on special double-exponential adaptive threshold (DEXAT) neurons. Our network achieves state-of-the-art classification accuracy (90%) while using ~53% lesser neurons than best reported prior art on Roshambo EMG dataset. (2) A new multi-channel spike encoder scheme for efficient processing of real-valued EMG data on neuromorphic systems. (3) Unique multi-compartment methodology to implement complex adaptive neurons on Intel's dedicated neuromorphic Loihi chip is shown. (4) RSNN implementation on Loihi (Nahuku 32) achieves significant energy/latency benefits of ~983X/19X compared to GPU for batch size as 50.

Aryan Vats, Manan Suri

The use of Deep Learning techniques for classification in Hyperspectral Imaging (HSI) is rapidly growing and achieving improved performances. Due to the nature of the data captured by sensors that produce HSI images, a common issue is the dimensionality of the bands that may or may not contribute to the label class distinction. Due to the widespread nature of class labels, Principal Component Analysis is a common method used for reducing the dimensionality. However,there may exist methods that incorporate all bands of the Hyperspectral image with the help of the Attention mechanism. Furthermore, to yield better spectral spatial feature extraction, recent methods have also explored the usage of Graph Convolution Networks and their unique ability to use node features in prediction, which is akin to the pixel spectral makeup. In this survey we present a comprehensive summary of Graph based and Attention based methods to perform Hyperspectral Image Classification for remote sensing and aerial HSI images. We also summarize relevant datasets on which these techniques have been evaluated and benchmark the processing techniques.

Vivek Parmar, Dwijay Bane, Syed Shakib Sarwar et al.

With the emergence of the Metaverse and focus on wearable devices in the recent years gesture based human-computer interaction has gained significance. To enable gesture recognition for VR/AR headsets and glasses several datasets focusing on egocentric i.e. first-person view have emerged in recent years. However, standard frame-based vision suffers from limitations in data bandwidth requirements as well as ability to capture fast motions. To overcome these limitation bio-inspired approaches such as event-based cameras present an attractive alternative. In this work, we present the first event-camera based egocentric gesture dataset for enabling neuromorphic, low-power solutions for XR-centric gesture recognition. The dataset has been made available publicly at the following URL: https://gitlab.com/NVM_IITD_Research/xrage.

Narayani Bhatia, Devang Mahesh, Jashandeep Singh et al.

Avian botulism is a paralytic bacterial disease in birds often leading to high fatality. In-vitro diagnostic techniques such as Mouse Bioassay, ELISA, PCR are usually non-preventive, post-mortem in nature, and require invasive sample collection from affected sites or dead birds. In this study, we build a first-ever multi-spectral, remote-sensing imagery based global Bird-Area Water-bodies Dataset (BAWD) (i.e. fused satellite images of warm-water lakes/marshy-lands or similar water-body sites that are important for avian fauna) backed by on-ground reporting evidence of outbreaks. BAWD consists of 16 topographically diverse global sites monitored over a time-span of 4 years (2016-2021). We propose a first-ever Artificial Intelligence based (AI) model to predict potential outbreak of Avian botulism called AVI-BoT (Aerosol Visible, Infra-red (NIR/SWIR) and Bands of Thermal). We also train and investigate a simpler (5-band) Causative-Factor model (based on prominent physiological factors reported in literature) to predict Avian botulism. AVI-BoT demonstrates a training accuracy of 0.96 and validation accuracy of 0.989 on BAWD, far superior in comparison to our model based on causative factors. We also perform an ablation study and perform a detailed feature-space analysis. We further analyze three test case study locations - Lower Klamath National Wildlife Refuge and Langvlei and Rondevlei lakes where an outbreak had occurred, and Pong Dam where an outbreak had not occurred and confirm predictions with on-ground reportings. The proposed technique presents a scale-able, low-cost, non-invasive methodology for continuous monitoring of bird-habitats against botulism outbreaks with the potential of saving valuable fauna lives.

Vivek Parmar, Narayani Bhatia, Shubham Negi et al.

In this paper, we present an analysis on the impact of network parameters for semantic segmentation architectures in context of UAV data processing. We present the analysis on the DroneDeploy Segmentation benchmark. Based on the comparative analysis we identify the optimal network architecture to be FPN-EfficientNetB3 with pretrained encoder backbones based on Imagenet Dataset. The network achieves IoU score of 0.65 and F1-score of 0.71 over the validation dataset. We also compare the various architectures in terms of their memory footprint and inference latency with further exploration of the impact of TensorRT based optimizations. We achieve memory savings of ~4.1x and latency improvement of 10% compared to Model: FPN and Backbone: InceptionResnetV2.

Sandeep Kaur Kingra, Vivek Parmar, Shubham Negi et al.

In this paper, we present an efficient hardware mapping methodology for realizing vector matrix multiplication (VMM) on resistive memory (RRAM) arrays. Using the proposed VMM computation technique, we experimentally demonstrate a binarized-ADALINE (Adaptive Linear) classifier on an OxRAM crossbar. An 8x8 OxRAM crossbar with Ni/3-nm HfO2/7 nm Al-doped-TiO2/TiN device stack is used. Weight training for the binarized-ADALINE classifier is performed ex-situ on UCI cancer dataset. Post weight generation the OxRAM array is carefully programmed to binary weight-states using the proposed weight mapping technique on a custom-built testbench. Our VMM powered binarized-ADALINE network achieves a classification accuracy of 78% in simulation and 67% in experiments. Experimental accuracy was found to drop mainly due to crossbar inherent sneak-path issues and RRAM device programming variability.

Abhilash Garg, Supriya Chakraborty, Manoj Malik et al.

Semiconductor NAND Flash based memory technology dominates the electronic Non-Volatile storage media market. Though NAND Flash offers superior performance and reliability over conventional magnetic HDDs, yet it suffers from certain data-security vulnerabilities. Such vulnerabilities can expose sensitive information stored on the media to security risks. It is thus necessary to study in detail the fundamental reasons behind data-security vulnerabilities of NAND Flash for use in critical applications. In this paper, the problem of unreliable data-deletion/sanitization in commercial NAND Flash media is investigated along with the fundamental reasons leading to such vulnerabilities. Exhaustive software based data recovery experiments (multiple iterations) has been carried out on commercial NAND Flash storage media (8 GB and 16 GB) for different types of filesystems (NTFS and FAT) and OS specific delete/Erase instructions. 100 % data recovery is obtained for windows and linux based delete/Erase commands. Inverse effect of performance enhancement techniques like wear levelling, bad block management etc. is also observed with the help of software based recovery experiments.

Vivek Parmar, Manan Suri

Deep Learning and its applications have gained tremendous interest recently in both academia and industry. Restricted Boltzmann Machines (RBMs) offer a key methodology to implement deep learning paradigms. This paper presents a novel approach for realizing hybrid CMOS-OxRAM based deep generative models (DGM). In our proposed hybrid DGM architectures, HfOx based (filamentary-type switching) OxRAM devices are extensively used for realizing multiple computational and non-computational functions such as: (i) Synapses (weights), (ii) internal neuron-state storage, (iii) stochastic neuron activation and (iv) programmable signal normalization. To validate the proposed scheme we have simulated two different architectures: (i) Deep Belief Network (DBN) and (ii) Stacked Denoising Autoencoder for classification and reconstruction of hand-written digits from a reduced MNIST dataset of 6000 images. Contrastive-divergence (CD) specially optimized for OxRAM devices was used to drive the synaptic weight update mechanism of each layer in the network. Overall learning rule was based on greedy-layer wise learning with no back propagation which allows the network to be trained to a good pre-training stage. Performance of the simulated hybrid CMOS-RRAM DGM model matches closely with software based model for a 2-layers deep network. Top-3 test accuracy achieved by the DBN was 95.5%. MSE of the SDA network was 0.003, lower than software based approach. Endurance analysis of the simulated architectures show that for 200 epochs of training (single RBM layer), maximum switching events/per OxRAM device was ~ 7000 cycles.