Jae-sun Seo

Jason Yik, Korneel Van den Berghe, Douwe den Blanken et al. · eth-zurich

Neuromorphic computing shows promise for advancing computing efficiency and capabilities of AI applications using brain-inspired principles. However, the neuromorphic research field currently lacks standardized benchmarks, making it difficult to accurately measure technological advancements, compare performance with conventional methods, and identify promising future research directions. Prior neuromorphic computing benchmark efforts have not seen widespread adoption due to a lack of inclusive, actionable, and iterative benchmark design and guidelines. To address these shortcomings, we present NeuroBench: a benchmark framework for neuromorphic computing algorithms and systems. NeuroBench is a collaboratively-designed effort from an open community of researchers across industry and academia, aiming to provide a representative structure for standardizing the evaluation of neuromorphic approaches. The NeuroBench framework introduces a common set of tools and systematic methodology for inclusive benchmark measurement, delivering an objective reference framework for quantifying neuromorphic approaches in both hardware-independent (algorithm track) and hardware-dependent (system track) settings. In this article, we outline tasks and guidelines for benchmarks across multiple application domains, and present initial performance baselines across neuromorphic and conventional approaches for both benchmark tracks. NeuroBench is intended to continually expand its benchmarks and features to foster and track the progress made by the research community.

Tianyi Zhang, Kishore Kasichainula, Yaoxin Zhuo et al.

Early object detection (OD) is a crucial task for the safety of many dynamic systems. Current OD algorithms have limited success for small objects at a long distance. To improve the accuracy and efficiency of such a task, we propose a novel set of algorithms that divide the image into patches, select patches with objects at various scales, elaborate the details of a small object, and detect it as early as possible. Our approach is built upon a transformer-based network and integrates the diffusion model to improve the detection accuracy. As demonstrated on BDD100K, our algorithms enhance the mAP for small objects from 1.03 to 8.93, and reduce the data volume in computation by more than 77\%. The source code is available at \href{https://github.com/destiny301/dpr}{https://github.com/destiny301/dpr}

Jian Meng, Yuan Liao, Anupreetham Anupreetham et al.

The development of model compression is continuously motivated by the evolution of various neural network accelerators with ASIC or FPGA. On the algorithm side, the ultimate goal of quantization or pruning is accelerating the expensive DNN computations on low-power hardware. However, such a "design-and-deploy" workflow faces under-explored challenges in the current hardware-algorithm co-design community. First, although the state-of-the-art quantization algorithm can achieve low precision with negligible degradation of accuracy, the latest deep learning framework (e.g., PyTorch) can only support non-customizable 8-bit precision, data format, and parameter extraction. Secondly, the objective of quantization is to enable the computation with low-precision data. However, the current SoTA algorithm treats the quantized integer as an intermediate result, while the final output of the quantizer is the "discretized" floating-point values, ignoring the practical needs and adding additional workload to hardware designers for integer parameter extraction and layer fusion. Finally, the compression toolkits designed by the industry are constrained to their in-house product or a handful of algorithms. The limited degree of freedom in the current toolkit and the under-explored customization hinder the prototype ASIC or FPGA-based accelerator design. To resolve these challenges, we propose Torch2Chip, an open-sourced, fully customizable, and high-performance toolkit that supports user-designed compression followed by automatic model fusion and parameter extraction. Torch2Chip incorporates the hierarchical design workflow, and the user-customized compression algorithm will be directly packed into the deployment-ready format for prototype chip verification with either CNN or vision transformer (ViT). The code is available at https://github.com/SeoLabCornell/torch2chip.

Tianyi Zhang, Kishore Kasichainula, Yaoxin Zhuo et al.

Conventional super-resolution methods suffer from two drawbacks: substantial computational cost in upscaling an entire large image, and the introduction of extraneous or potentially detrimental information for downstream computer vision tasks during the refinement of the background. To solve these issues, we propose a novel transformer-based algorithm, Selective Super-Resolution (SSR), which partitions images into non-overlapping tiles, selects tiles of interest at various scales with a pyramid architecture, and exclusively reconstructs these selected tiles with deep features. Experimental results on three datasets demonstrate the efficiency and robust performance of our approach for super-resolution. Compared to the state-of-the-art methods, the FID score is reduced from 26.78 to 10.41 with 40% reduction in computation cost for the BDD100K dataset. The source code is available at https://github.com/destiny301/SSR.

Hoigi Seo, Wongi Jeong, Jae-sun Seo et al.

Large-scale text encoders in text-to-image (T2I) diffusion models have demonstrated exceptional performance in generating high-quality images from textual prompts. Unlike denoising modules that rely on multiple iterative steps, text encoders require only a single forward pass to produce text embeddings. However, despite their minimal contribution to total inference time and floating-point operations (FLOPs), text encoders demand significantly higher memory usage, up to eight times more than denoising modules. To address this inefficiency, we propose Skip and Re-use layers (Skrr), a simple yet effective pruning strategy specifically designed for text encoders in T2I diffusion models. Skrr exploits the inherent redundancy in transformer blocks by selectively skipping or reusing certain layers in a manner tailored for T2I tasks, thereby reducing memory consumption without compromising performance. Extensive experiments demonstrate that Skrr maintains image quality comparable to the original model even under high sparsity levels, outperforming existing blockwise pruning methods. Furthermore, Skrr achieves state-of-the-art memory efficiency while preserving performance across multiple evaluation metrics, including the FID, CLIP, DreamSim, and GenEval scores.

Tianyi Zhang, Baoxin Li, Jae-sun Seo et al.

In recent years, the long-range attention mechanism of vision transformers has driven significant performance breakthroughs across various computer vision tasks. However, the traditional self-attention mechanism, which processes both informative and non-informative tokens, suffers from inefficiency and inaccuracies. While sparse attention mechanisms have been introduced to mitigate these issues by pruning tokens involved in attention, they often lack context-awareness and intelligence. These mechanisms frequently apply a uniform token selection strategy across different inputs for batch training or optimize efficiency only for the inference stage. To overcome these challenges, we propose a novel algorithm: Select and Pack Attention (SPA). SPA dynamically selects informative tokens using a low-cost gating layer supervised by selection labels and packs these tokens into new batches, enabling a variable number of tokens to be used in parallelized GPU batch training and inference. Extensive experiments across diverse datasets and computer vision tasks demonstrate that SPA delivers superior performance and efficiency, including a 0.6 mAP improvement in object detection and a 16.4% reduction in computational costs.

Gokul Krishnan, Sumit K. Mandal, Chaitali Chakrabarti et al.

With the widespread use of Deep Neural Networks (DNNs), machine learning algorithms have evolved in two diverse directions -- one with ever-increasing connection density for better accuracy and the other with more compact sizing for energy efficiency. The increase in connection density increases on-chip data movement, which makes efficient on-chip communication a critical function of the DNN accelerator. The contribution of this work is threefold. First, we illustrate that the point-to-point (P2P)-based interconnect is incapable of handling a high volume of on-chip data movement for DNNs. Second, we evaluate P2P and network-on-chip (NoC) interconnect (with a regular topology such as a mesh) for SRAM- and ReRAM-based in-memory computing (IMC) architectures for a range of DNNs. This analysis shows the necessity for the optimal interconnect choice for an IMC DNN accelerator. Finally, we perform an experimental evaluation for different DNNs to empirically obtain the performance of the IMC architecture with both NoC-tree and NoC-mesh. We conclude that, at the tile level, NoC-tree is appropriate for compact DNNs employed at the edge, and NoC-mesh is necessary to accelerate DNNs with high connection density. Furthermore, we propose a technique to determine the optimal choice of interconnect for any given DNN. In this technique, we use analytical models of NoC to evaluate end-to-end communication latency of any given DNN. We demonstrate that the interconnect optimization in the IMC architecture results in up to 6$\times$ improvement in energy-delay-area product for VGG-19 inference compared to the state-of-the-art ReRAM-based IMC architectures.

Adnan Siraj Rakin, Li Yang, Jingtao Li et al.

Recently developed adversarial weight attack, a.k.a. bit-flip attack (BFA), has shown enormous success in compromising Deep Neural Network (DNN) performance with an extremely small amount of model parameter perturbation. To defend against this threat, we propose RA-BNN that adopts a complete binary (i.e., for both weights and activation) neural network (BNN) to significantly improve DNN model robustness (defined as the number of bit-flips required to degrade the accuracy to as low as a random guess). However, such an aggressive low bit-width model suffers from poor clean (i.e., no attack) inference accuracy. To counter this, we propose a novel and efficient two-stage network growing method, named Early-Growth. It selectively grows the channel size of each BNN layer based on channel-wise binary masks training with Gumbel-Sigmoid function. Apart from recovering the inference accuracy, our RA-BNN after growing also shows significantly higher resistance to BFA. Our evaluation of the CIFAR-10 dataset shows that the proposed RA-BNN can improve the clean model accuracy by ~2-8 %, compared with a baseline BNN, while simultaneously improving the resistance to BFA by more than 125 x. Moreover, on ImageNet, with a sufficiently large (e.g., 5,000) amount of bit-flips, the baseline BNN accuracy drops to 4.3 % from 51.9 %, while our RA-BNN accuracy only drops to 37.1 % from 60.9 % (9 % clean accuracy improvement).

Vinay Joshi, Wangxin He, Jae-sun Seo et al.

The cost involved in training deep neural networks (DNNs) on von-Neumann architectures has motivated the development of novel solutions for efficient DNN training accelerators. We propose a hybrid in-memory computing (HIC) architecture for the training of DNNs on hardware accelerators that results in memory-efficient inference and outperforms baseline software accuracy in benchmark tasks. We introduce a weight representation technique that exploits both binary and multi-level phase-change memory (PCM) devices, and this leads to a memory-efficient inference accelerator. Unlike previous in-memory computing-based implementations, we use a low precision weight update accumulator that results in more memory savings. We trained the ResNet-32 network to classify CIFAR-10 images using HIC. For a comparable model size, HIC-based training outperforms baseline network, trained in floating-point 32-bit (FP32) precision, by leveraging appropriate network width multiplier. Furthermore, we observe that HIC-based training results in about 50% less inference model size to achieve baseline comparable accuracy. We also show that the temporal drift in PCM devices has a negligible effect on post-training inference accuracy for extended periods (year). Finally, our simulations indicate HIC-based training naturally ensures that the number of write-erase cycles seen by the devices is a small fraction of the endurance limit of PCM, demonstrating the feasibility of this architecture for achieving hardware platforms that can learn in the field.

Colby R. Banbury, Vijay Janapa Reddi, Max Lam et al.

Recent advancements in ultra-low-power machine learning (TinyML) hardware promises to unlock an entirely new class of smart applications. However, continued progress is limited by the lack of a widely accepted benchmark for these systems. Benchmarking allows us to measure and thereby systematically compare, evaluate, and improve the performance of systems and is therefore fundamental to a field reaching maturity. In this position paper, we present the current landscape of TinyML and discuss the challenges and direction towards developing a fair and useful hardware benchmark for TinyML workloads. Furthermore, we present our four benchmarks and discuss our selection methodology. Our viewpoints reflect the collective thoughts of the TinyMLPerf working group that is comprised of over 30 organizations.

Shihui Yin, Xiaoyu Sun, Shimeng Yu et al.

Deep learning hardware designs have been bottlenecked by conventional memories such as SRAM due to density, leakage and parallel computing challenges. Resistive devices can address the density and volatility issues, but have been limited by peripheral circuit integration. In this work, we demonstrate a scalable RRAM based in-memory computing design, termed XNOR-RRAM, which is fabricated in a 90nm CMOS technology with monolithic integration of RRAM devices between metal 1 and 2. We integrated a 128x64 RRAM array with CMOS peripheral circuits including row/column decoders and flash analog-to-digital converters (ADCs), which collectively become a core component for scalable RRAM-based in-memory computing towards large deep neural networks (DNNs). To maximize the parallelism of in-memory computing, we assert all 128 wordlines of the RRAM array simultaneously, perform analog computing along the bitlines, and digitize the bitline voltages using ADCs. The resistance distribution of low resistance states is tightened by write-verify scheme, and the ADC offset is calibrated. Prototype chip measurements show that the proposed design achieves high binary DNN accuracy of 98.5% for MNIST and 83.5% for CIFAR-10 datasets, respectively, with energy efficiency of 24 TOPS/W and 158 GOPS throughput. This represents 5.6X, 3.2X, 14.1X improvements in throughput, energy-delay product (EDP), and energy-delay-squared product (ED2P), respectively, compared to the state-of-the-art literature. The proposed XNOR-RRAM can enable intelligent functionalities for area-/energy-constrained edge computing devices.

Shreyas Kolala Venkataramanaiah, Yufei Ma, Shihui Yin et al.

Training of convolutional neural networks (CNNs)on embedded platforms to support on-device learning is earning vital importance in recent days. Designing flexible training hard-ware is much more challenging than inference hardware, due to design complexity and large computation/memory requirement. In this work, we present an automatic compiler-based FPGA accelerator with 16-bit fixed-point precision for complete CNNtraining, including Forward Pass (FP), Backward Pass (BP) and Weight Update (WU). We implemented an optimized RTL library to perform training-specific tasks and developed an RTL compiler to automatically generate FPGA-synthesizable RTL based on user-defined constraints. We present a new cyclic weight storage/access scheme for on-chip BRAM and off-chip DRAMto efficiently implement non-transpose and transpose operations during FP and BP phases, respectively. Representative CNNs for CIFAR-10 dataset are implemented and trained on Intel Stratix 10-GX FPGA using proposed hardware architecture, demonstrating up to 479 GOPS performance.

Paul N. Whatmough, Chuteng Zhou, Patrick Hansen et al.

The computational demands of computer vision tasks based on state-of-the-art Convolutional Neural Network (CNN) image classification far exceed the energy budgets of mobile devices. This paper proposes FixyNN, which consists of a fixed-weight feature extractor that generates ubiquitous CNN features, and a conventional programmable CNN accelerator which processes a dataset-specific CNN. Image classification models for FixyNN are trained end-to-end via transfer learning, with the common feature extractor representing the transfered part, and the programmable part being learnt on the target dataset. Experimental results demonstrate FixyNN hardware can achieve very high energy efficiencies up to 26.6 TOPS/W ($4.81 \times$ better than iso-area programmable accelerator). Over a suite of six datasets we trained models via transfer learning with an accuracy loss of $<1\%$ resulting in up to 11.2 TOPS/W - nearly $2 \times$ more efficient than a conventional programmable CNN accelerator of the same area.

Chetan Singh Thakur, Jamal Molin, Gert Cauwenberghs et al.

Neuromorphic engineering (NE) encompasses a diverse range of approaches to information processing that are inspired by neurobiological systems, and this feature distinguishes neuromorphic systems from conventional computing systems. The brain has evolved over billions of years to solve difficult engineering problems by using efficient, parallel, low-power computation. The goal of NE is to design systems capable of brain-like computation. Numerous large-scale neuromorphic projects have emerged recently. This interdisciplinary field was listed among the top 10 technology breakthroughs of 2014 by the MIT Technology Review and among the top 10 emerging technologies of 2015 by the World Economic Forum. NE has two-way goals: one, a scientific goal to understand the computational properties of biological neural systems by using models implemented in integrated circuits (ICs); second, an engineering goal to exploit the known properties of biological systems to design and implement efficient devices for engineering applications. Building hardware neural emulators can be extremely useful for simulating large-scale neural models to explain how intelligent behavior arises in the brain. The principle advantages of neuromorphic emulators are that they are highly energy efficient, parallel and distributed, and require a small silicon area. Thus, compared to conventional CPUs, these neuromorphic emulators are beneficial in many engineering applications such as for the porting of deep learning algorithms for various recognitions tasks. In this review article, we describe some of the most significant neuromorphic spiking emulators, compare the different architectures and approaches used by them, illustrate their advantages and drawbacks, and highlight the capabilities that each can deliver to neural modelers.

Shihui Yin, Gaurav Srivastava, Shreyas K. Venkataramanaiah et al.

Deep learning algorithms have shown tremendous success in many recognition tasks; however, these algorithms typically include a deep neural network (DNN) structure and a large number of parameters, which makes it challenging to implement them on power/area-constrained embedded platforms. To reduce the network size, several studies investigated compression by introducing element-wise or row-/column-/block-wise sparsity via pruning and regularization. In addition, many recent works have focused on reducing precision of activations and weights with some reducing down to a single bit. However, combining various sparsity structures with binarized or very-low-precision (2-3 bit) neural networks have not been comprehensively explored. In this work, we present design techniques for minimum-area/-energy DNN hardware with minimal degradation in accuracy. During training, both binarization/low-precision and structured sparsity are applied as constraints to find the smallest memory footprint for a given deep learning algorithm. The DNN model for CIFAR-10 dataset with weight memory reduction of 50X exhibits accuracy comparable to that of the floating-point counterpart. Area, performance and energy results of DNN hardware in 40nm CMOS are reported for the MNIST dataset. The optimized DNN that combines 8X structured compression and 3-bit weight precision showed 98.4% accuracy at 20nJ per classification.

Shihui Yin, Shreyas K. Venkataramanaiah, Gregory K. Chen et al.

We present a new back propagation based training algorithm for discrete-time spiking neural networks (SNN). Inspired by recent deep learning algorithms on binarized neural networks, binary activation with a straight-through gradient estimator is used to model the leaky integrate-fire spiking neuron, overcoming the difficulty in training SNNs using back propagation. Two SNN training algorithms are proposed: (1) SNN with discontinuous integration, which is suitable for rate-coded input spikes, and (2) SNN with continuous integration, which is more general and can handle input spikes with temporal information. Neuromorphic hardware designed in 40nm CMOS exploits the spike sparsity and demonstrates high classification accuracy (>98% on MNIST) and low energy (48.4-773 nJ/image).

R. Tapiador, A. Rios-Navarro, A. Linares-Barranco et al.

Deep learning has significantly advanced the state of the art in artificial intelligence, gaining wide popularity from both industry and academia. Special interest is around Convolutional Neural Networks (CNN), which take inspiration from the hierarchical structure of the visual cortex, to form deep layers of convolutional operations, along with fully connected classifiers. Hardware implementations of these deep CNN architectures are challenged with memory bottlenecks that require many convolution and fully-connected layers demanding large amount of communication for parallel computation. Multi-core CPU based solutions have demonstrated their inadequacy for this problem due to the memory wall and low parallelism. Many-core GPU architectures show superior performance but they consume high power and also have memory constraints due to inconsistencies between cache and main memory. FPGA design solutions are also actively being explored, which allow implementing the memory hierarchy using embedded BlockRAM. This boosts the parallel use of shared memory elements between multiple processing units, avoiding data replicability and inconsistencies. This makes FPGAs potentially powerful solutions for real-time classification of CNNs. Both Altera and Xilinx have adopted OpenCL co-design framework from GPU for FPGA designs as a pseudo-automatic development solution. In this paper, a comprehensive evaluation and comparison of Altera and Xilinx OpenCL frameworks for a 5-layer deep CNN is presented. Hardware resources, temporal performance and the OpenCL architecture for CNNs are discussed. Xilinx demonstrates faster synthesis, better FPGA resource utilization and more compact boards. Altera provides multi-platforms tools, mature design community and better execution times.

Ming Tu, Visar Berisha, Yu Cao et al.

Deep neural networks are typically represented by a much larger number of parameters than shallow models, making them prohibitive for small footprint devices. Recent research shows that there is considerable redundancy in the parameter space of deep neural networks. In this paper, we propose a method to compress deep neural networks by using the Fisher Information metric, which we estimate through a stochastic optimization method that keeps track of second-order information in the network. We first remove unimportant parameters and then use non-uniform fixed point quantization to assign more bits to parameters with higher Fisher Information estimates. We evaluate our method on a classification task with a convolutional neural network trained on the MNIST data set. Experimental results show that our method outperforms existing methods for both network pruning and quantization.