With the rapid expansion in the scale of large language models (LLMs), enabling efficient distributed inference across multiple computing units has become increasingly critical. However, communication overheads from popular distributed inference techniques such as Tensor Parallelism pose a significant challenge to achieve scalability and low latency. Therefore, we introduce a novel optimization technique, Sync-Point Drop (SPD), to reduce communication overheads in tensor parallelism by selectively dropping synchronization on attention outputs. In detail, we first propose a block design that allows execution to proceed without communication through SPD. Second, we apply different SPD strategies to attention blocks based on their sensitivity to the model accuracy. The proposed methods effectively alleviate communication bottlenecks while minimizing accuracy degradation during LLM inference, offering a scalable solution for diverse distributed environments: SPD offered about 20% overall inference latency reduction with $<$ 1% accuracy regression for LLaMA2-70B inference over 8 GPUs.

Using a vision-inspired keyword spotting framework, we propose an architecture with input-dependent dynamic depth capable of processing streaming audio. Specifically, we extend a conformer encoder with trainable binary gates that allow us to dynamically skip network modules according to the input audio. Our approach improves detection and localization accuracy on continuous speech using Librispeech top1000 most frequent words while maintaining a small memory footprint. The inclusion of gates also reduces the average amount of processing without affecting the overall performance. These benefits are shown to be even more pronounced using the Google speech commands dataset placed over background noise where up to 97% of the processing is skipped on non-speech inputs, therefore making our method particularly interesting for an always-on keyword spotter.

Binary neural network (BNN) delivers increased compute intensity and reduces memory/data requirements for computation. Scalable BNN enables inference in a limited time due to different constraints. This paper explores the application of Scalable BNN in oblivious inference, a service provided by a server to mistrusting clients. Using this service, a client can obtain the inference result on his/her data by a trained model held by the server without disclosing the data or learning the model parameters. Two contributions of this paper are: (1) we devise lightweight cryptographic protocols explicitly designed to exploit the unique characteristics of BNNs. (2) we present an advanced dynamic exploration of the runtime-accuracy tradeoff of scalable BNNs in a single-shot training process. While previous works trained multiple BNNs with different computational complexities (which is cumbersome due to the slow convergence of BNNs), we train a single BNN that can perform inference under various computational budgets. Compared to CryptFlow2, the state-of-the-art technique in the oblivious inference of non-binary DNNs, our approach reaches 3× faster inference while keeping the same accuracy. Compared to XONN, the state-of-the-art technique in the oblivious inference of binary networks, we achieve 2× to 12× faster inference while obtaining higher accuracy.

Streaming keyword spotting is a widely used solution for activating voice assistants. Methods based on Deep Neural Networks with Hidden Markov Model (DNN-HMM) have proven to be efficient and widely adopted in this space, primarily because of the ability to detect and identify the start and end of the wake-up word at low compute cost. However, such hybrid systems suffer from loss metric mismatch when the DNN and HMM are trained independently. Sequence discriminative training cannot fully mitigate the loss-metric mismatch due to the inherent Markovian style of the operation. We propose an low footprint CNN model, called HEiMDaL, to detect and localize keywords in streaming conditions. We introduce an alignment-based classification loss to detect the occurrence of the keyword along with an offset loss to predict the start of the keyword. HEiMDaL shows 73% improvement in detection metrics along with equivalent localization accuracy and with the same memory footprint as existing DNN-HMM style models for a given wake-word.

This paper explores the possibility of using visual object detection techniques for word localization in speech data. Object detection has been thoroughly studied in the contemporary literature for visual data. Noting that an audio can be interpreted as a 1-dimensional image, object localization techniques can be fundamentally useful for word localization. Building upon this idea, we propose a lightweight solution for word detection and localization. We use bounding box regression for word localization, which enables our model to detect the occurrence, offset, and duration of keywords in a given audio stream. We experiment with LibriSpeech and train a model to localize 1000 words. Compared to existing work [1], our method reduces model size by 94%, and improves the F1 score by 6.5%.

We introduce COINN - an efficient, accurate, and scalable framework for oblivious deep neural network (DNN) inference in the two-party setting. In our system, DNN inference is performed without revealing the client's private inputs to the server or revealing server's proprietary DNN weights to the client. To speedup the secure inference while maintaining a high accuracy, we make three interlinked innovations in the plaintext and ciphertext domains: (i) we develop a new domain-specific low-bit quantization scheme tailored for high-efficiency ciphertext computation, (ii) we construct novel techniques for increasing data re-use in secure matrix multiplication allowing us to gain significant performance boosts through factored operations, and (iii) we propose customized cryptographic protocols that complement our optimized DNNs in the ciphertext domain. By co-optimization of the aforesaid components, COINN brings an unprecedented level of efficiency to the setting of oblivious DNN inference, achieving an end-to-end runtime speedup of 4.7×14.4× over the state-of-the-art. We demonstrate the scalability of our proposed methods by optimizing complex DNNs with over 100 layers and performing oblivious inference in the Billion-operation regime for the challenging ImageNet dataset. Our framework is available at https://github.com/ACESLabUCSD/COINN.git.

Recent advances in adversarial Deep Learning (DL) have opened up a new and largely unexplored surface for malicious attacks jeopardizing the integrity of autonomous DL systems. This article introduces CuRTAIL, a novel end-to-end computing framework to characterize and thwart potential adversarial attacks and significantly improve the reliability (safety) of a victim DL model. We formalize the goal of preventing adversarial attacks as an optimization problem to minimize the rarely observed regions in the latent feature space spanned by a DL network. To solve the aforementioned minimization problem, a set of complementary but disjoint modular redundancies are trained to validate the legitimacy of the input samples. The proposed countermeasure is unsupervised, meaning that no adversarial sample is leveraged to train modular redundancies. This, in turn, ensures the effectiveness of the defense in the face of generic attacks. We evaluate the robustness of our proposed methodology against the state-of-the-art adaptive attacks in a white-box setting considering that the adversary knows everything about the victim model and its defenders. Extensive evaluations for analyzing MNIST, CIFAR10, and ImageNet data corroborate the effectiveness of CuRTAIL framework against adversarial samples. The computations in each modular redundancy can be performed independently of the other redundancy modules. As such, CuRTAIL detection algorithm can be completely parallelized among multiple hardware settings to achieve maximum throughput. We further provide an open-source Application Programming Interface (API) to facilitate the adoption of the proposed framework for various applications.

In the contemporary big data realm, Deep Neural Networks (DNNs) are evolving towards more complex architectures to achieve higher inference accuracy. Model compression techniques can be leveraged to efficiently deploy these compute-intensive architectures on resource-limited mobile devices. Such methods comprise various hyperparameters that require per-layer customization to ensure high accuracy. Choosing the hyperparameters is cumbersome as the pertinent search space grows exponentially with model layers. This paper introduces GeneCAI, a novel optimization method that automatically learns how to tune per-layer compression hyperparameters. We devise a bijective translation scheme that encodes compressed DNNs to the genotype space. Each genotype's optimality is measured using a multi-objective score based on the accuracy and number of floating-point operations. We develop customized genetic operations to iteratively evolve the non-dominated solutions towards the optimal Pareto front, thus, capturing the optimal trade-off between model accuracy and complexity. GeneCAI optimization method is highly scalable and can achieve a near-linear performance boost on distributed multi-GPU platforms. Our extensive evaluations demonstrate that GeneCAI outperforms existing rule-based and reinforcement learning methods in DNN compression by finding models that lie on a better accuracy/complexity Pareto curve.

Deep neural networks (DNN) have demonstrated effectiveness for various applications such as image processing, video segmentation, and speech recognition. Running state-of-theart DNNs on current systems mostly relies on either generalpurpose processors, ASIC designs, or FPGA accelerators, all of which suffer from data movements due to the limited on-chip memory and data transfer bandwidth. In this work, we propose a novel framework, called RAPIDNN, which performs neuron-to-memory transformation in order to accelerate DNNs in a highly parallel architecture. RAPIDNN reinterprets a DNN model and maps it into a specialized accelerator, which is designed using non-volatile memory blocks that model four fundamental DNN operations, i.e., multiplication, addition, activation functions, and pooling. The framework extracts representative operands of a DNN model, e.g., weights and input values, using clustering methods to optimize the model for in-memory processing. Then, it maps the extracted operands and their pre-computed results into the accelerator memory blocks. At runtime, the accelerator identifies computation results based on efficient in-memory search capability which also provides tunability of approximation to improve computation efficiency further. Our evaluation shows that RAPIDNN achieves 68.4×, 49.5× energy efficiency improvement and 48.1×, 10.9× speedup as compared to ISAAC and PipeLayer, the state-of-the-art DNN accelerators, while ensuring less than 0.5% quality loss.

We propose CleaNN, the first end-to-end framework that enables online mitigation of Trojans for embedded Deep Neural Network (DNN) applications. A Trojan attack works by injecting a backdoor in the DNN while training; during inference, the Trojan can be activated by the specific backdoor trigger. What differentiates CleaNN from the prior work is its lightweight methodology which recovers the ground-truth class of Trojan samples without the need for labeled data, model retraining, or prior assumptions on the trigger or the attack. We leverage dictionary learning and sparse approximation to characterize the statistical behavior of benign data and identify Trojan triggers. CleaNN is devised based on algorithm/hardware co-design and is equipped with specialized hardware to enable efficient real-time execution on resource-constrained embedded platforms. Proof of concept evaluations on CleaNN for the state-of-the-art Neural Trojan attacks on visual benchmarks demonstrate its competitive advantage in terms of attack resiliency and execution overhead.

This article proposes EncoDeep, an end-to-end framework that facilitates encoding, bitwidth customization, fine-tuning, and implementation of neural networks on FPGA platforms. EncoDeep incorporates nonlinear encoding to the computation flow of neural networks to save memory. The encoded features demand significantly lower storage compared to the raw full-precision activation values; therefore, the execution flow of EncoDeep hardware engine is completely performed within the FPGA using on-chip streaming buffers with no access to the off-chip DRAM. We further propose a fully automated optimization algorithm that determines the flexible encoding bitwidths across network layers. EncoDeep full-stack framework comprises a compiler that takes a high-level Python description of an arbitrary neural network. The compiler then instantiates the corresponding elements from EncoDeep Hardware library for FPGA implementation. Our evaluations on MNIST, SVHN, and CIFAR-10 datasets demonstrate an average of 4.65× throughput improvement compared to stand-alone weight encoding. We further compare EncoDeep with six FPGA accelerators on ImageNet, showing an average of 3.6× and 2.54× improvement in throughput and performance-per-watt, respectively.

Hyperdimensional (HD) computing is gaining traction as an alternative light-way machine learning approach for cognition tasks. Inspired by the neural activity patterns of the brain, HD computing performs cognition tasks by exploiting longsize vectors, namely hypervectors, rather than working with scalar numbers as used in conventional computing. Since a hypervector is represented by thousands of dimensions (elements), the majority of prior work assume binary elements to simplify the computation and alleviate the processing cost. In this paper, we first demonstrate that the dimensions need to have more than one bit to provide an acceptable accuracy to make HD computing applicable to real-world cognitive tasks. Increasing the bit-width, however, sacrifices energy efficiency and performance, even when using low-bit integers as the hypervector elements. To address this issue, we propose a framework for HD acceleration, dubbed SparseHD, that leverages the advantages of sparsity to improve the efficiency of HD computing. Essentially, SparseHD takes account of statistical properties of a trained HD model and drops the least effective elements of the model, augmented by iterative retraining to compensate the possible quality loss raised by sparsity. Thanks to the bit-level manipulability and abounding parallelism granted by FPGAs, we also propose a novel FPGAbased accelerator to effectively utilize the advantage of sparsity in HD computation. We evaluate the efficiency of our framework for practical classification problems. We observe that SparseHD makes the HD model up to 90% sparse while affording a minimal quality loss (less than 1%) compared to the non-sparse baseline model. Our evaluation shows that, on average, SparseHD provides 48.5× and 15.0× lower energy consumption and faster execution as compared to the AMD R390 GPU implementation.

This paper proposes ReBNet, an end-to-end framework for training reconfigurable binary neural networks on software and developing efficient accelerators for execution on FPGA. Binary neural networks offer an intriguing opportunity for deploying large-scale deep learning models on resource-constrained devices. Binarization reduces the memory footprint and replaces the power-hungry matrix-multiplication with light-weight XnorPopcount operations. However, binary networks suffer from a degraded accuracy compared to their fixed-point counterparts. We show that the state-of-the-art methods for optimizing binary networks accuracy, significantly increase the implementation cost and complexity. To compensate for the degraded accuracy while adhering to the simplicity of binary networks, we devise the first reconfigurable scheme that can adjust the classification accuracy based on the application. Our proposition improves the classification accuracy by representing features with multiple levels of residual binarization. Unlike previous methods, our approach does not exacerbate the area cost of the hardware accelerator. Instead, it provides a tradeoff between throughput and accuracy while the area overhead of multi-level binarization is negligible.

Deep Learning (DL) models have been shown to be vulnerable to adversarial attacks. In light of the adversarial attacks, it is critical to reliably quantify the confidence of the prediction in a neural network to enable safe adoption of DL models in autonomous sensitive tasks (e.g., unmanned vehicles and drones). This article discusses recent research advances for unsupervised model assurance against the strongest adversarial attacks known to date and quantitatively compare their performance. Given the widespread usage of DL models, it is imperative to provide model assurance by carefully looking into the feature maps automatically learned within D1 models instead of looking back with regret when deep learning systems are compromised by adversaries.

Recent advances in adversarial Deep Learning (DL) have opened up a largely unexplored surface for malicious attacks jeopardizing the integrity of autonomous DL systems. With the wide-spread usage of DL in critical and time-sensitive applications, including unmanned vehicles, drones, and video surveillance systems, online detection of malicious inputs is of utmost importance. We propose DeepFense, the first end-to-end automated framework that simultaneously enables efficient and safe execution of DL models. DeepFense formalizes the goal of thwarting adversarial attacks as an optimization problem that minimizes the rarely observed regions in the latent feature space spanned by a DL network. To solve the aforementioned minimization problem, a set of complementary but disjoint modular redundancies are trained to validate the legitimacy of the input samples in parallel with the victim DL model. DeepFense leverages hardware/software/algorithm co-design and customized acceleration to achieve just-in-time performance in resource-constrained settings. The proposed countermeasure is unsupervised, meaning that no adversarial sample is leveraged to train modular redundancies. We further provide an accompanying API to reduce the non-recurring engineering cost and ensure automated adaptation to various platforms. Extensive evaluations on FPGAs and GPUs demonstrate up to two orders of magnitude performance improvement while enabling online adversarial sample detection.

Neural networks are machine learning models that have been successfully used in many applications. Due to the high computational complexity of neural networks, deploying such models on embedded devices with severe power/resource constraints is troublesome. Neural networks are inherently approximate and can be simplified. We propose LookNN, a methodology to replace floating-point multiplications with lookup table search. First, we devise an algorithmic solution to adapt conventional neural networks to LookNN such that the model's accuracy is minimally affected. We provide experimental results and theoretical analysis demonstrating the applicability of the method. Next, we design enhanced general purpose processors for searching look-up tables: each processing element of our GPU has access to a small associative memory, enabling it to bypass redundant computations. Our evaluations on AMD Southern Island GPU architecture shows that LookNN results in 2.2x energy saving and 2.5x speedup running four different neural network applications with zero additive error. For the same four applications, if we tolerate an additive error of less than 0.2%, LookNN can achieve an average of 3x energy improvement and 2.6x speedup compared to the traditional GPU architecture.

We propose a novel end-to-end framework to customize execution of deep neural networks on FPGA platforms. Our framework employs a reconfigurable clustering approach that encodes the parameters of deep neural networks in accordance with the application's accuracy requirement and the underlying platform constraints. The throughput of FPGA-based realizations of neural networks is often bounded by the memory access bandwidth. The use of encoded parameters reduces both the required memory bandwidth and the computational complexity of neural networks, increasing the effective throughput. Our framework enables systematic customization of encoded deep neural networks for different FPGA platforms. Proof-of-concept evaluations on four different applications demonstrate up to 9-fold reduction in memory footprint and 15-fold improvement in the operational throughput while the drop in accuracy remains below 0.1%.

We introduce CAMsure, the first realization of secure Content Addressable Memory (CAM) in the context of approximate search using near-neighbor algorithms. CAMsure provides a lightweight solution for practical secure (approximate) search with a minimal drop in the accuracy of the search results. CAM has traditionally been used as a hardware search engine that explores the entire memory in a single clock cycle. However, there has been little attention to the security of the data stored in CAM. Our approach stores distance-preserving hash embeddings within CAM to ensure data privacy. The hashing method provides data confidentiality while preserving similarity in the sense that a high resemblance in the data domain is translated to a small Hamming distance in the hash domain. Consequently, the objective of near-neighbor search is converted to approximate lookup table search which is compatible with the realizations of emerging content addressable memories. Our methodology delivers on average two orders of magnitude faster response time compared to RAM-based solutions that preserve the privacy of data owners.