Keshab K. Parhi

Xingyi Liu, Keshab K. Parhi

Modern neural networks have revolutionized the fields of computer vision (CV) and Natural Language Processing (NLP). They are widely used for solving complex CV tasks and NLP tasks such as image classification, image generation, and machine translation. Most state-of-the-art neural networks are over-parameterized and require a high computational cost. One straightforward solution is to replace the layers of the networks with their low-rank tensor approximations using different tensor decomposition methods. This paper reviews six tensor decomposition methods and illustrates their ability to compress model parameters of convolutional neural networks (CNNs), recurrent neural networks (RNNs) and Transformers. The accuracy of some compressed models can be higher than the original versions. Evaluations indicate that tensor decompositions can achieve significant reductions in model size, run-time and energy consumption, and are well suited for implementing neural networks on edge devices.

Keshab K. Parhi

Fast time-domain algorithms have been developed in signal processing applications to reduce the multiplication complexity. For example, fast convolution structures using Cook-Toom and Winograd algorithms are well understood. Short length fast convolutions can be iterated to obtain fast convolution structures for long lengths. In this paper, we show that well known fast convolution structures form the basis for design of fast algorithms in four other problem domains: fast parallel filters, fast polynomial modular multiplication, and fast pointwise multiplication in the DFT and NTT domains. Fast polynomial modular multiplication and fast pointwise multiplication problems are important for cryptosystem applications such as post-quantum cryptography and homomorphic encryption. By establishing the equivalence of these problems, we show that a fast structure from one domain can be used to design a fast structure for another domain. This understanding is important as there are many well known solutions for fast convolution that can be used in other signal processing and cryptosystem applications.

Nanda K. Unnikrishnan, Keshab K. Parhi

In our prior work, LayerPipe, we had introduced an approach to accelerate training of convolutional, fully connected, and spiking neural networks by overlapping forward and backward computation. However, despite empirical success, a principled understanding of how much gradient delay needs to be introduced at each layer to achieve desired level of pipelining was not addressed. This paper, LayerPipe2, fills that gap by formally deriving LayerPipe using variable delayed gradient adaptation and retiming. We identify where delays may be legally inserted and show that the required amount of delay follows directly from the network structure where inner layers require fewer delays and outer layers require longer delays. When pipelining is applied at every layer, the amount of delay depends only on the number of remaining downstream stages. When layers are pipelined in groups, all layers in the group share the same assignment of delays. These insights not only explain previously observed scheduling patterns but also expose an often overlooked challenge that pipelining implicitly requires storage of historical weights. We overcome this storage bottleneck by developing a pipeline--aware moving average that reconstructs the required past states rather than storing them explicitly. This reduces memory cost without sacrificing the accuracy guarantees that makes pipelined learning viable. The result is a principled framework that illustrates how to construct LayerPipe architectures, predicts their delay requirements, and mitigates their storage burden, thereby enabling scalable pipelined training with controlled communication computation tradeoffs.

Lulu Ge, Keshab K. Parhi

This paper addresses the clustering of data in the hyperdimensional computing (HDC) domain. In prior work, an HDC-based clustering framework, referred to as HDCluster, has been proposed. However, the performance of the existing HDCluster is not robust. The performance of HDCluster is degraded as the hypervectors for the clusters are chosen at random during the initialization step. To overcome this bottleneck, we assign the initial cluster hypervectors by exploring the similarity of the encoded data, referred to as \textit{query} hypervectors. Intra-cluster hypervectors have a higher similarity than inter-cluster hypervectors. Harnessing the similarity results among query hypervectors, this paper proposes four HDC-based clustering algorithms: similarity-based k-means, equal bin-width histogram, equal bin-height histogram, and similarity-based affinity propagation. Experimental results illustrate that: (i) Compared to the existing HDCluster, our proposed HDC-based clustering algorithms can achieve better accuracy, more robust performance, fewer iterations, and less execution time. Similarity-based affinity propagation outperforms the other three HDC-based clustering algorithms on eight datasets by 2~38% in clustering accuracy. (ii) Even for one-pass clustering, i.e., without any iterative update of the cluster hypervectors, our proposed algorithms can provide more robust clustering accuracy than HDCluster. (iii) Over eight datasets, five out of eight can achieve higher or comparable accuracy when projected onto the hyperdimensional space. Traditional clustering is more desirable than HDC when the number of clusters, $k$, is large.

Wenrui Xu, Keshab K. Parhi

Large language models (LLMs) and LLM-based agents have been widely deployed in a wide range of applications in the real world, including healthcare diagnostics, financial analysis, customer support, robotics, and autonomous driving, expanding their powerful capability of understanding, reasoning, and generating natural languages. However, the wide deployment of LLM-based applications exposes critical security and reliability risks, such as the potential for malicious misuse, privacy leakage, and service disruption that weaken user trust and undermine societal safety. This paper provides a systematic overview of the details of adversarial attacks targeting both LLMs and LLM-based agents. These attacks are organized into three phases in LLMs: Training-Phase Attacks, Inference-Phase Attacks, and Availability & Integrity Attacks. For each phase, we analyze the details of representative and recently introduced attack methods along with their corresponding defenses. We hope our survey will provide a good tutorial and a comprehensive understanding of LLM security, especially for attacks on LLMs. We desire to raise attention to the risks inherent in widely deployed LLM-based applications and highlight the urgent need for robust mitigation strategies for evolving threats.

Abhijit Chatterjee, Niraj K. Jha, Jonathan D. Cohen et al.

The field of artificial intelligence (AI) has taken a tight hold on broad aspects of society, industry, business, and governance in ways that dictate the prosperity and might of the world's economies. The AI market size is projected to grow from 189 billion USD in 2023 to 4.8 trillion USD by 2033. Currently, AI is dominated by large language models that exhibit linguistic and visual intelligence. However, training these models requires a massive amount of data scraped from the web as well as large amounts of energy (50--60 GWh to train GPT-4). Despite these costs, these models often hallucinate, a characteristic that prevents them from being deployed in critical application domains. In contrast, the human brain consumes only 20~W of power. What is needed is the next level of AI evolution in which lightweight domain-specific multimodal models with higher levels of intelligence can reason, plan, and make decisions in dynamic environments with real-time data and prior knowledge, while learning continuously and evolving in ways that enhance future decision-making capability. This will define the next wave of AI, progressing from today's large models, trained with vast amounts of data, to nimble energy-efficient domain-specific agents that can reason and think in a world full of uncertainty. To support such agents, hardware will need to be reimagined to allow energy efficiencies greater than 1000x over the state of the art. Such a vision of future AI systems is developed in this work.

Weihang Tan, Antian Wang, Yingjie Lao et al.

This paper presents a low-latency hardware accelerator for modular polynomial multiplication for lattice-based post-quantum cryptography and homomorphic encryption applications. The proposed novel modular polynomial multiplier exploits the fast finite impulse response (FIR) filter architecture to reduce the computational complexity of the schoolbook modular polynomial multiplication. We also extend this structure to fast $M$-parallel architectures while achieving low-latency, high-speed, and full hardware utilization. We comprehensively evaluate the performance of the proposed architectures under various polynomial settings as well as in the Saber scheme for post-quantum cryptography as a case study. The experimental results show that our proposed modular polynomial multiplier reduces the computation time and area-time product, respectively, compared to the state-of-the-art designs.

Nanda K. Unnikrishnan, Keshab K. Parhi

The time required for training the neural networks increases with size, complexity, and depth. Training model parameters by backpropagation inherently creates feedback loops. These loops hinder efficient pipelining and scheduling of the tasks within the layer and between consecutive layers. Prior approaches, such as PipeDream, have exploited the use of delayed gradient to achieve inter-layer pipelining. However, these approaches treat the entire backpropagation as a single task; this leads to an increase in computation time and processor underutilization. This paper presents novel optimization approaches where the gradient computations with respect to the weights and the activation functions are considered independently; therefore, these can be computed in parallel. This is referred to as intra-layer optimization. Additionally, the gradient computation with respect to the activation function is further divided into two parts and distributed to two consecutive layers. This leads to balanced scheduling where the computation time of each layer is the same. This is referred to as inter-layer optimization. The proposed system, referred to as LayerPipe, reduces the number of clock cycles required for training while maximizing processor utilization with minimal inter-processor communication overhead. LayerPipe achieves an average speedup of 25% and upwards of 80% with 7 to 9 processors with less communication overhead when compared to PipeDream.

Chunhua Deng, Siyu Liao, Yi Xie et al.

Deep neural network (DNN) has emerged as the most important and popular artificial intelligent (AI) technique. The growth of model size poses a key energy efficiency challenge for the underlying computing platform. Thus, model compression becomes a crucial problem. However, the current approaches are limited by various drawbacks. Specifically, network sparsification approach suffers from irregularity, heuristic nature and large indexing overhead. On the other hand, the recent structured matrix-based approach (i.e., CirCNN) is limited by the relatively complex arithmetic computation (i.e., FFT), less flexible compression ratio, and its inability to fully utilize input sparsity. To address these drawbacks, this paper proposes PermDNN, a novel approach to generate and execute hardware-friendly structured sparse DNN models using permuted diagonal matrices. Compared with unstructured sparsification approach, PermDNN eliminates the drawbacks of indexing overhead, non-heuristic compression effects and time-consuming retraining. Compared with circulant structure-imposing approach, PermDNN enjoys the benefits of higher reduction in computational complexity, flexible compression ratio, simple arithmetic computation and full utilization of input sparsity. We propose PermDNN architecture, a multi-processing element (PE) fully-connected (FC) layer-targeted computing engine. The entire architecture is highly scalable and flexible, and hence it can support the needs of different applications with different model configurations. We implement a 32-PE design using CMOS 28nm technology. Compared with EIE, PermDNN achieves 3.3x~4.8x higher throughout, 5.9x~8.5x better area efficiency and 2.8x~4.0x better energy efficiency on different workloads. Compared with CirCNN, PermDNN achieves 11.51x higher throughput and 3.89x better energy efficiency.

Lulu Ge, Keshab K. Parhi

Hyperdimensional (HD) computing is built upon its unique data type referred to as hypervectors. The dimension of these hypervectors is typically in the range of tens of thousands. Proposed to solve cognitive tasks, HD computing aims at calculating similarity among its data. Data transformation is realized by three operations, including addition, multiplication and permutation. Its ultra-wide data representation introduces redundancy against noise. Since information is evenly distributed over every bit of the hypervectors, HD computing is inherently robust. Additionally, due to the nature of those three operations, HD computing leads to fast learning ability, high energy efficiency and acceptable accuracy in learning and classification tasks. This paper introduces the background of HD computing, and reviews the data representation, data transformation, and similarity measurement. The orthogonality in high dimensions presents opportunities for flexible computing. To balance the tradeoff between accuracy and efficiency, strategies include but are not limited to encoding, retraining, binarization and hardware acceleration. Evaluations indicate that HD computing shows great potential in addressing problems using data in the form of letters, signals and images. HD computing especially shows significant promise to replace machine learning algorithms as a light-weight classifier in the field of internet of things (IoTs).

Sohini Roychowdhury, Dara D. Koozekanani, Michael Reinsbach et al.

This paper presents a novel automated system that segments six sub-retinal layers from optical coherence tomography (OCT) image stacks of healthy patients and patients with diabetic macular edema (DME). First, each image in the OCT stack is denoised using a Wiener deconvolution algorithm that estimates the additive speckle noise variance using a novel Fourier-domain based structural error. This denoising method enhances the image SNR by an average of 12dB. Next, the denoised images are subjected to an iterative multi-resolution high-pass filtering algorithm that detects seven sub-retinal surfaces in six iterative steps. The thicknesses of each sub-retinal layer for all scans from a particular OCT stack are then compared to the manually marked groundtruth. The proposed system uses adaptive thresholds for denoising and segmenting each image and hence it is robust to disruptions in the retinal micro-structure due to DME. The proposed denoising and segmentation system has an average error of 1.2-5.8 $μm$ and 3.5-26$μm$ for segmenting sub-retinal surfaces in normal and abnormal images with DME, respectively. For estimating the sub-retinal layer thicknesses, the proposed system has an average error of 0.2-2.5 $μm$ and 1.8-18 $μm$ in normal and abnormal images, respectively. Additionally, the average inner sub-retinal layer thickness in abnormal images is estimated as 275$μm (r=0.92)$ with an average error of 9.3 $μm$, while the average thickness of the outer layers in abnormal images is estimated as 57.4$μm (r=0.74)$ with an average error of 3.5 $μm$. The proposed system can be useful for tracking the disease progression for DME over a period of time.