Diffusion-based text-to-image generation models trade latency for quality: small models are fast but generate lower quality images, while large models produce better images but are slow. We present MoDM, a novel caching-based serving system for diffusion models that dynamically balances latency and quality through a mixture of diffusion models. Unlike prior approaches that rely on model-specific internal features, MoDM caches final images, allowing seamless retrieval and reuse across multiple diffusion model families. This design enables adaptive serving by dynamically balancing latency and image quality: using smaller models for cache-hit requests to reduce latency while reserving larger models for cache-miss requests to maintain quality. Small model image quality is preserved using retrieved cached images. We design a global monitor that optimally allocates GPU resources and balances inference workload, ensuring high throughput while meeting Service-Level Objectives (SLOs) under varying request rates. Our evaluations show that MoDM significantly reduces average serving time by 2.5× while retaining image quality, making it a practical solution for scalable and resource-efficient model deployment.

Despite their remarkable progress in multimodal understanding tasks, large vision language models (LVLMs) often suffer from "hallucination", generating texts misaligned with the visual context.Existing methods aimed at reducing hallucinations through inference time intervention incur a significant increase in latency.To mitigate this, we present SPIN, a task-agnostic attention-guided head suppression strategy that can be seamlessly integrated during inference without incurring any significant compute or latency overhead.We investigate whether hallucination in LVLMs can be linked to specific model components.Our analysis suggests that hallucinations can be attributed to a dynamic subset of attention heads in each layer.Leveraging this insight, for each text query token, we selectively suppress attention heads that exhibit low attention to image tokens, keeping the top-k attention heads intact.Extensive evaluations on visual question answering and image description tasks demonstrate the efficacy of SPIN in reducing hallucination scores up to 2.7× while maintaining F1, and improving throughput by 1.8× compared to existing alternatives.Code is available here.

The inherent quadratic complexity of the attention mechanism in transformer models has driven the research community to explore alternative architectures with sub-quadratic complexity, such as state-space models. Mamba has established itself as a leading model within this emerging paradigm, achieving state-of-the-art results in various language modeling benchmarks. However, despite its impressive performance, Mamba's effectiveness is limited by its pre-training context length, resulting in a pronounced degradation when the model is tasked with handling longer contexts. Our investigation reveals that Mamba's inability to generalize effectively to long contexts is primarily due to the out-of-distribution (OOD) discretization steps. To address this critical limitation, we introduce _**MambaExtend**_, a novel framework designed to significantly enhance the context extension capabilities of Mamba. Specifically, MambaExtend leverages a _**training-free**_ approach to calibrate _only_ the scaling factors of discretization modules for different layers. We demonstrate both gradient-based and gradient-free zeroth-order optimization to learn the optimal scaling factors for each Mamba layer, requiring orders of magnitude fewer updates as opposed to the parameter fine-tuning-based alternatives. Using this approach, we achieve a training-free context extension of up to 32x, expanding the context from 2k to 64k tokens with minimal increases in perplexity. In contrast to existing fine-tuning methods, MambaExtend selectively calibrates the scaling factors, requiring up to $\mathbf{5.42 * 10^6} \times$ fewer parameter updates and incurring up to $\mathbf{3.87} \times$ lower peak memory usage, while delivering comparable or superior long-context performance across multiple tasks. Codes and checkpoints are available here$^1$.

Large language models (LLMs), despite their impressive performance across a wide range of tasks, often struggle to balance two competing objectives in open-ended text generation: fostering diversity and creativity while preserving logical coherence. Existing truncated sampling techniques, including temperature scaling, top-*p* (nucleus) sampling, and min-*p* sampling, aim to manage this trade-off. However, they exhibit limitations, particularly in the effective incorporation of the confidence of the model into the corresponding sampling strategy. For example, min-*p* sampling relies on a single top token as a heuristic for confidence, eventually underutilizing the information of the probability distribution. To effectively incorporate the model confidence, this paper presents **_top-H_ decoding**. We first establish the theoretical foundation of the interplay between creativity and coherence in truncated sampling by formulating an **entropy-constrained minimum divergence** problem. We then prove this minimization problem to be equivalent to an **entropy-constrained mass maximization (ECMM)** problem, which is NP-hard. Finally, we present top-H decoding, a computationally efficient greedy algorithm to solve the ECMM problem. Extensive empirical evaluations demonstrate that top-H outperforms the state-of-the-art (SoTA) alternative of min-*p* sampling by up to **25.63%** on creative writing benchmarks, while maintaining robustness on question-answering datasets such as GPQA, GSM8K, and MT-Bench. Additionally, an *LLM-as-judge* evaluation confirms that top-H indeed produces coherent outputs even at higher temperatures, where creativity is especially critical. In summary, top-H advances SoTA in open-ended text generation and can be *easily integrated* into creative writing applications. The code is available at [https://github.com/ErfanBaghaei/Top-H-Decoding](https://github.com/ErfanBaghaei/Top-H-Decoding).

Video tasks are compute-heavy and thus pose a challenge when deploying in real-time applications, particularly for tasks that require state-of-the-art Vision Trans-formers (ViTs). Several research efforts have tried to address this challenge by leveraging the fact that large portions of the video undergo very little change across frames leading to redundant computations in frame-based video processing. In particular, some works leverage pixel or semantic differences across frames, however, this yields limited latency benefits with significantly increased memory overhead. This paper, in contrast, presents a strategy for masking regions in video frames that leverages the semantic information in images and the temporal correlation between frames to significantly reduce FLOPs and latency with little to no penalty in performance over base-line models. In particular, we demonstrate that by lever-aging extracted features from previous frames, ViT back-bones directly benefit from region masking, skipping up to 80% of input regions, improving FLOPs and latency by 3.14x and 1.5x. We improve memory and latency over the state-of-the-art (SOTA) by 2.3x and 1.14x, while maintaining similar detection performance. Additionally, our approach demonstrates promising results on convolutional neural networks (CNNs) and provides latency improvements over the SOTA up to 1.3x using specialized computational kernels. Code is available at: https://github.com/sreetamasarkar/MaskVD

Large pre-trained language models have recently gained significant traction due to their improved performance on various down-stream tasks like text classification and question answering, requiring only few epochs of fine-tuning. However, their large model sizes often prohibit their applications on resource-constrained edge devices. Existing solutions of yielding parameter-efficient BERT models largely rely on compute-exhaustive training and fine-tuning. Moreover, they often rely on additional compute heavy models to mitigate the performance gap. In this paper, we present Sensi-BERT, a sensitivity driven efficient fine-tuning of BERT models that can take an off-the-shelf pre-trained BERT model and yield highly parameter-efficient models for downstream tasks. In specific, we perform sensitivity analysis to rank each individual parameter tensor, that then is used to trim them accordingly during fine-tuning for a given parameter or FLOPs budget. Our experiments show the efficacy of Sensi-BERT across different downstream tasks including MNLI, QQP, QNLI, SST-2, MRPC, CoLA, and SQuAD, showing better performance at similar or smaller parameter budget compared to various alternatives.

Neuromorphic computing and, in particular, spiking neural networks (SNNs) have become an attractive alternative to deep neural networks for a broad range of signal processing applications, processing static and/or temporal inputs from different sensory modalities, including audio and vision sensors. In this paper, we start with a description of recent advances in algorithmic and optimization innovations to efficiently train and scale low-latency, and energy-efficient spiking neural networks (SNNs) for complex machine learning applications. We then discuss the recent efforts in algorithm-architecture co-design that explores the inherent trade-offs between achieving high energy-efficiency and low latency while still providing high accuracy and trustworthiness. We then describe the underlying hardware that has been developed to leverage such algorithmic innovations in an efficient way. In particular, we describe a hybrid method to integrate significant portions of the model’s computation within both memory components as well as the sensor itself. Finally, we discuss the potential path forward for research in building deployable SNN systems identifying key challenges in the algorithm-hardware-application co-design space with an emphasis on trustworthiness.

Recently, private inference (PI) has addressed the rising concern over data and model privacy in machine learning inference as a service. However, existing PI frameworks suffer from high computational and communication costs due to the expensive multi-party computation (MPC) protocols. Existing literature has developed lighter MPC protocols to yield more efficient PI schemes. We, in contrast, propose to lighten them by introducing an empirically-defined privacy evaluation. To that end, we reformulate the threat model of PI and use inference data privacy attacks (IDPAs) to evaluate data privacy. We then present an enhanced IDPA, named distillation-based inverse-network attack (DINA), for improved privacy evaluation. Finally, we leverage the findings from DINA and propose C2PI, a two-party PI framework presenting an efficient partitioning of the neural network model and requiring only the initial few layers to be performed with MPC protocols. Based on our experimental evaluations, relaxing the formal data privacy guarantees C2PI can speed up existing PI frameworks, including Delphi [1] and Cheetah [2], up to 2.89× and 3.88× under LAN and WAN settings, respectively, and save up to 2.75× communication costs.

The massive amounts of data generated by camera sensors motivate data processing inside pixel arrays, i.e., at the extreme-edge. Several critical developments have fueled recent interest in the processing-in-pixel-in-memory paradigm for a wide range of visual machine intelligence tasks, including (1) advances in 3D integration technology to enable complex processing inside each pixel in a 3D integrated manner while maintaining pixel density, (2) analog processing circuit techniques for massively parallel low-energy in-pixel computations, and (3) algorithmic techniques to mitigate non-idealities associated with analog processing through hardware-aware training schemes. This article presents a comprehensive technology-circuit-algorithm landscape that connects technology capabilities, circuit design strategies, and algorithmic optimizations to power, performance, area, bandwidth reduction, and application-level accuracy metrics. We present our results using a comprehensive co-design framework incorporating hardware and algorithmic optimizations for various complex real-life visual intelligence tasks mapped onto our P2M paradigm.

Due to the high activation sparsity and use of accumulates (AC) instead of expensive multiply-and-accumulates (MAC), neuromorphic spiking neural networks (SNNs) have emerged as a promising low-power alternative to traditional DNNs for several computer vision (CV) applications. However, most existing SNNs require multiple time steps for acceptable inference accuracy, hindering real-time deployment and increasing spiking activity and, consequently, energy consumption. Recent works proposed direct encoding that directly feeds the analog pixel values in the first layer of the SNN in order to significantly reduce the number of time steps. Although the overhead for the first layer MACs with direct encoding is negligible for deep SNNs and the CV processing is efficient using SNNs, the data transfer between the image sensors and the downstream processing costs significant bandwidth and may dominate the total energy. To mitigate this concern, we propose an in-sensor computing hardware-software co-design framework for SNNs targeting image recognition tasks. Our approach reduces the bandwidth between sensing and processing by 12−96× and the resulting total energy by 2.32× compared to traditional CV processing, with a 3.8% reduction in accuracy on ImageNet.

Existing deep neural networks (DNNs) that achieve state-of-the-art (SOTA) performance on both clean and adversarially-perturbed images rely on either activation or weight conditioned convolution operations. However, such conditional learning costs additional multiply-accumulate (MAC) or addition operations, increasing inference memory and compute costs. To that end, we present a sparse mixture once for all adversarial training (SMART), that allows a model to train once and then in-situ trade-off between accuracy and robustness, that too at a reduced compute and parameter overhead. In particular, SMART develops two expert paths, for clean and adversarial images, respectively, that are then conditionally trained via respective dedicated sets of binary sparsity masks. Extensive evaluations on multiple image classification datasets across different models show SMART to have up to 2.72× fewer non-zero parameters costing proportional reduction in compute overhead, while yielding SOTA accuracy-robustness trade-off. Additionally, we present insightful observations in designing sparse masks to successfully condition on both clean and perturbed images.

Pipeline parallelism has achieved great success in deploying large-scale transformer models in cloud environments, but has received less attention in edge environments. Unlike in cloud scenarios with high-speed and stable network inter-connects, dynamic bandwidth in edge systems can degrade distributed pipeline performance. We address this issue with QuantPipe, a communication-efficient distributed edge system that introduces post-training quantization (PTQ) to compress the communicated tensors. QuantPipe uses adaptive PTQ to change bitwidths in response to bandwidth dynamics, maintaining transformer pipeline performance while incurring limited inference accuracy loss. We further improve the accuracy with a directed-search analytical clipping for integer quantization method (DS-ACIQ), which bridges the gap between estimated and real data distributions. Experimental results show that QuantPipe adapts to dynamic bandwidth to maintain pipeline performance while achieving a practical model accuracy using a wide range of quantization bitwidths, e.g., improving accuracy under 2-bit quantization by 15.85% on ImageNet compared to naive quantization.

The concern over data and model privacy in machine learning inference as a service (MLaaS) has led to the development of private inference (PI) techniques. However, existing PI frameworks, especially those designed for large models such as vision transformers (ViT), suffer from high computational and communication overheads caused by the expensive multi-party computation (MPC) protocols. The encrypted attention module that involves the softmax operation contributes significantly to this overhead. In this work, we present a family of models dubbed RNA-ViT, that leverage a novel attention module called reduced-dimension approximate normalized attention and a latency efficient GeLU-alternative layer. In particular, RNA-ViT uses two novel techniques to improve PI efficiency in ViTs: a reduced-dimension normalized attention (RNA) architecture and a high order polynomial (HOP) softmax approximation for latency efficient normalization. We also propose a novel metric, accuracy-to-latency ratio (A2L), to evaluate modules in terms of their accuracy and PI latency. Based on this metric, we perform an analysis to identify a nonlinearity module with improved PI efficiency. Our extensive experiments show that RNA-ViT can achieve average 3.53×, 3.54×, 1.66× lower PI latency with an average accuracy improvement of 0.93%, 2.04%, and 2.73% compared to the state-of-the-art scheme MPCViT [1], on CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively.

Recently, private inference (PI) has addressed the rising concern over data and model privacy in machine learning inference as a service. However, existing PI frameworks suffer from high computational and communication overheads due to the expensive multi-party computation (MPC) protocols, particularly for large models such as vision transformers (ViT). The majority of this overhead is due to the encrypted softmax operation in each self-attention layer. In this work, we present SAL-ViT with two novel techniques to boost PI efficiency on ViTs. Our first technique is a learnable PI-efficient approximation to softmax, namely, learnable 2Quad (L2Q), that introduces learnable scaling and shifting parameters to the prior 2Quad softmax approximation, enabling improvement in accuracy. Then, given our observation that external attention (EA) presents lower PI latency than widely-adopted self-attention (SA) at the cost of accuracy, we present a selective attention search (SAS) method to integrate the strength of EA and SA. Specifically, for a given lightweight EA ViT, we leverage a constrained optimization procedure to selectively search and replace EA modules with SA alternatives to maximize the accuracy. Our extensive experiments show that our SAL-ViT can averagely achieve 1.28×, 1.28×, 1.14× lower PI latency with 1.79%, 1.41%, and 2.08% higher accuracy compared to the existing alternatives, on CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively.

Vision Transformer models, such as ViT, Swin Transformer, and Transformer-in-Transformer, have recently gained significant traction in computer vision tasks due to their ability to capture the global relation between features which leads to superior performance. However, they are compute-heavy and difficult to deploy in resource-constrained edge devices. Existing hardware accelerators, including those for the closely-related BERT transformer models, do not target highly resource-constrained environments. In this paper, we address this gap and propose ViTA - a configurable hardware accelerator for inference of vision transformer models, targeting resource-constrained edge computing devices and avoiding repeated off-chip memory accesses. We employ a head-level pipeline and inter-layer MLP optimizations, and can support several commonly used vision transformer models with changes solely in our control logic. We achieve nearly 90% hardware utilization efficiency on most vision transformer models, report a power of 0.88W when synthesised with a clock of 150 MHz, and get reasonable frame rates - all of which makes ViTA suitable for edge applications.

Efficient custom pooling techniques that can aggressively trim the dimensions of a feature map for resource-constrained computer vision applications have recently gained significant traction. However, prior pooling works extract only the local context of the activation maps, limiting their effectiveness. In contrast, we propose a novel non-local self-attentive pooling method that can be used as a drop-in replacement to the standard pooling layers, such as max/average pooling or strided convolution. The proposed self-attention module uses patch embedding, multihead self-attention, and spatial-channel restoration, followed by sigmoid activation and exponential soft-max. This self-attention mechanism efficiently aggregates dependencies between non-local activation patches during downsampling. Extensive experiments on standard object classification and detection tasks with various convolutional neural network (CNN) architectures demonstrate the superiority of our proposed mechanism over the state-of-the-art (SOTA) pooling techniques. In particular, we surpass the test accuracy of existing pooling techniques on different variants of MobileNet-V2 on ImageNet by an average of ~1.2%. With the aggressive down-sampling of the activation maps in the initial layers (providing up to 22x reduction in memory consumption), our approach achieves 1.43% higher test accuracy compared to SOTA techniques with iso-memory footprints. This enables the deployment of our models in memory-constrained devices, such as micro-controllers (without losing significant accuracy), because the initial activation maps consume a significant amount of on-chip memory for high-resolution images required for complex vision tasks. Our pooling method also leverages channel pruning to further reduce memory footprints. Codes are available at https://github.com/CFun/Non-Local-Pooling.

Existing models that achieve state-of-the-art (SOTA) performance on both clean and adversarially-perturbed images rely on convolution operations conditioned with feature-wise linear modulation (FiLM) layers. These layers require additional parameters and are hyperparameter sensitive. They significantly increase training time, memory cost, and potential latency which can be costly for resource-limited or real-time applications. In this paper, we present a fast learnable once-for-all adversarial training (FLOAT) algorithm, which instead of the existing FiLM-based conditioning, presents a unique weight conditioned learning that requires no additional layer, thereby incurring no significant increase in parameter count, training time, or network latency compared to standard adversarial training. In particular, we add configurable scaled noise to the weight tensors that enables a trade-off between clean and adversarial performance. Extensive experiments show that FLOAT can yield SOTA performance improving both clean and perturbed image classification by up to ~6% and ~10%, respectively. Moreover, real hardware measurement shows that FLOAT can reduce the training time by up to 1.43× with fewer model parameters of up to 1.47× on iso-hyperparameter settings compared to the FiLM-based alternatives. Additionally, to further improve memory effi ciency we introduce FLOAT sparse (FLOATS), a form of non-iterative model pruning and provide detailed empirical analysis in yielding a three-way accuracy-robustness-complexity trade-off for these new class of pruned conditionally trained models.

Large DNNs with mixed-precision quantization can achieve ultra-high compression while retaining high classification performance. However, because of the challenges in finding an accurate metric that can guide the optimization process, these methods either sacrifice significant performance compared to the 32-bit floating-point (FP-32) baseline or rely on a compute-expensive, iterative training policy that requires the availability of a pre-trained baseline. To address this issue, this paper presents BMPQ, a training method that uses bit gradients to analyze layer sensitivities and yield mixed-precision quantized models. BMPQ requires a single training iteration but does not need a pre-trained baseline. It uses an integer linear program (ILP) to dynamically adjust the precision of layers during training, subject to a fixed hardware budget. To evaluate the efficacy of BMPQ, we conduct extensive experiments with VGG16 and ResNet18 on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets. Compared to the baseline FP-32 models, BMPQ can yield models that have 15.4x fewer parameter bits with negligible drop in accuracy. Compared to the SOTA “during training”, mixed-precision training scheme, our models are 2.1 x, 2.2x, and 2.9x smaller, on CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively, with an improved accuracy of up to 14.54%.

Deep neural networks with large model sizes achieve state-of-the-art results for tasks in computer vision and natural language processing. However, such models are too compute- or memory-intensive for resource-constrained edge devices. Prior works on parallel and distributed execution primarily focus on training-rather than inference-using homogeneous accelerators in data centers. We propose PipeEdge, a distributed framework for edge systems that uses pipeline parallelism to both speed up inference and enable running larger, more accurate models that otherwise cannot fit on single edge devices. PipeEdge uses an optimal partition strategy that considers heterogeneity in compute, memory, and network bandwidth. Our empirical evaluation demonstrates that PipeEdge achieves 11.88× and 12.78× speedup using 16 edge devices for the ViT-Huge and BERT-Large models, respectively, with no accuracy loss. Similarly, PipeEdge improves throughput for ViT-Huge (which cannot fit in a single device) by 3.93× over a 4-device baseline using 16 edge devices. Finally, we show up to 4.16× throughput improvement over the state-of-the-art PipeDream when using a heterogeneous set of devices.

Today’s high resolution, high frame rate cameras in autonomous vehicles generate a large volume of data that needs to be transferred and processed by a downstream processor or machine learning (ML) accelerator to enable intelligent computing tasks, such as multi-object detection and tracking. The massive amount of data transfer incurs significant energy, latency, and bandwidth bottlenecks, which hinders real-time processing. To mitigate this problem, we propose an algorithm-hardware co-design framework called Processing-in-Pixel-in-Memory-based object Detection and Tracking (P2M-DeTrack). P2M-DeTrack is based on a custom faster R-CNN-based model that is distributed partly inside the pixel array (front-end) and partly in a separate FPGA/ASIC (back-end). The proposed front-end in-pixel processing down-samples the input feature maps significantly with judiciously optimized strided convolution and pooling. Compared to a conventional baseline design that transfers frames of RGB pixels to the back-end, the resulting P2M-DeTrack designs reduce the data bandwidth between sensor and back-end by up to 24×. The designs also reduce the sensor and total energy (obtained from in-house circuit simulations at Globalfoundries 22nm technology node) per frame by 5.7× and 1.14×, respectively. Lastly, they reduce the sensing and total frame latency by an estimated 1.7× and 3×, respectively. We evaluate our approach on the multi-object object detection (tracking) task of the large-scale BDD100K dataset and observe only a 0.5% reduction in the mean average precision (0.8% reduction in the identification F1 score) compared to the state-of-the-art.

We present a dynamic network rewiring (DNR) method to generate pruned deep neural network (DNN) models that both are robust against adversarially generated images and maintain high accuracy on clean images. In particular, the disclosed DNR training method is based on a unified constrained optimization formulation using a novel hybrid loss function that merges sparse learning with robust adversarial training. This training strategy dynamically adjusts inter-layer connectivity based on per-layer normalized momentum computed from the hybrid loss function. To further improve the robustness of the pruned models, we propose DNR++, an extension of the DNR method where we introduce the idea of sparse parametric Gaussian noise tensor that is added to the weight tensors to yield robust regularization. In contrast to existing robust pruning frameworks that require multiple training iterations, the proposed DNR and DNR++ achieve an overall target pruning ratio with only a single training iteration and can be tuned to support both irregular and structured channel pruning. To demonstrate the efficacy of the proposed method under the no-increased-training-time “free” adversarial training scenario, we finally present FDNR++, a simple yet effective training modification that can yield robust yet compressed models requiring training time comparable to that of an unpruned non-adversarial training. To evaluate the merits of our disclosed training methods, experiments were performed with two widely accepted models, namely VGG16 and ResNet18, on CIFAR-10 and CIFAR-100 as well as with VGG16 on Tiny-ImageNet. Compared to the baseline uncompressed models, our methods provide over 20× compression on all the datasets without any significant drop of either clean or adversarial classification performance. Moreover, extensive experiments show that our methods consistently find compressed models with better clean and adversarial image classification performance than what is achievable through state-of-the-art alternatives. We provide insightful observations to help make various model, parameter density, and prune-type selection choices and have open-sourced our saved models and test codes to ensure reproducibility of our results.

This paper presents a dynamic network rewiring (DNR) method to generate pruned deep neural network (DNN) models that are robust against adversarial attacks yet maintain high accuracy on clean images. In particular, the disclosed DNR method is based on a unified constrained optimization formulation using a hybrid loss function that merges ultra-high model compression with robust adversarial training. This training strategy dynamically adjusts inter-layer connectivity based on per-layer normalized momentum computed from the hybrid loss function. In contrast to existing robust pruning frameworks that require multiple training iterations, the proposed learning strategy achieves an overall target pruning ratio with only a single training iteration and can be tuned to support both irregular and structured channel pruning. To evaluate the merits of DNR, experiments were performed with two widely accepted models, namely VGG16 and ResNet-18, on CIFAR-10, CIFAR-100 as well as with VGG16 on Tiny-ImageNet. Compared to the baseline uncompressed models, DNR provides over 20x compression on all the datasets with no significant drop in either clean or adversarial classification accuracy. Moreover, our experiments show that DNR consistently finds compressed models with better clean and adversarial image classification performance than what is achievable through state-of-the-art alternatives. Our models and test codes are available at https://github.com/ksouvik52/DNR_ASP_DAC2021.

We propose a novel framework for producing a class of parameter and compute efficient models called AttentionLite suitable for resource constrained applications. Prior work has primarily focused on optimizing models either via knowledge distillation or pruning. In addition to fusing these two mechanisms, our joint optimization framework also leverages recent advances in self-attention as a substitute for convolutions. We can simultaneously distill knowledge from a compute heavy teacher while also pruning the student model in a single pass of training thereby reducing training and fine tuning times considerably. We evaluate the merits of our proposed approach on the CIFAR-10, CIFAR-100 and Tiny-ImageNet datasets. Not only do our AttentionLite models significantly outperform their unoptimized counterparts in accuracy, we find that in some cases, that they perform almost as well as their compute-heavy teachers while consuming only a fraction of the parameters and FLOPs. Concretely, AttentionLite models can achieve up to 30× parameter efficiency and 2× computation efficiency with no significant accuracy drop compared to their teacher.

Low-latency deep spiking neural networks (SNNs) have become a promising alternative to conventional artificial neural networks (ANNs) because of their potential for increased energy efficiency on event-driven neuromorphic hardware. Neural networks, including SNNs, however, are subject to various adversarial attacks and must be trained to remain resilient against such attacks for many applications. Nevertheless, due to prohibitively high training costs associated with SNNs, an analysis and optimization of deep SNNs under various adversarial attacks have been largely overlooked. In this paper, we first present a detailed analysis of the inherent robustness of low-latency SNNs against popular gradient-based attacks, namely fast gradient sign method (FGSM) and projected gradient descent (PGD). Motivated by this analysis, to harness the model’s robustness against these attacks we present an SNN training algorithm that uses crafted input noise and incurs no additional training time. To evaluate the merits of our algorithm, we conducted extensive experiments with variants of VGG and ResNet on both CIFAR-10 and CIFAR-100 dataset. Compared to standard trained direct-input SNNs, our trained models yield improved classification accuracy of up to 13.7% and 10.1% on FGSM and PGD attack generated images, respectively, with negligible loss in clean image accuracy. Our models also outperform inherently-robust SNNs trained on rate-coded inputs with improved or similar classification performance on attack-generated images while having up to 25× and ∼4.6× lower latency and computation energy, respectively. For reproducibility, we have open-sourced the code at github.com/ksouvik52/hiresnn2021.

Spiking Neural Networks (SNNs) have emerged as an attractive alternative to traditional deep learning frameworks, since they provide higher computational efficiency in event driven neuromorphic hardware. However, the state-of-the-art (SOTA) SNNs suffer from high inference latency, resulting from inefficient input encoding and training techniques. The most widely used input coding schemes, such as Poisson based rate-coding, do not leverage the temporal learning capabilities of SNNs. This paper presents a training framework for low-latency energy-efficient SNNs that uses a hybrid encoding scheme at the input layer in which the analog pixel values of an image are directly applied during the first timestep and a novel variant of spike temporal coding is used during subsequent timesteps. In particular, neurons in every hidden layer are restricted to fire at most once per image which increases activation sparsity. To train these hybrid-encoded SNNs, we propose a variant of the gradient descent based spike timing dependent backpropagation (STDB) mechanism using a novel cross entropy loss function based on both the output neurons' spike time and membrane potential. The resulting SNNs have reduced latency and high activation sparsity, yielding significant improvements in computational efficiency. In particular, we evaluate our proposed training scheme on image classification tasks from CIFAR-10 and CIFAR-100 datasets on several VGG architectures. We achieve top-l accuracy of 66.46% with 5 timesteps on the CIFAR-100 dataset with ~125x less compute energy than an equivalent standard ANN. Additionally, our proposed SNN performs 5–300 x faster inference compared to other state-of-the-art rate or temporally coded SNN models.

Knowledge distillation (KD) has recently been identified as a method that can unintentionally leak private information regarding the details of a teacher model to an unauthorized student. Recent research in developing undistillable nasty teachers that can protect model confidentiality has gained significant attention. However, the level of protection these nasty models offer has been largely untested. In this paper, we show that transferring knowledge to a shallow sub-section of a student can largely reduce a teacher’s influence. By exploring the depth of the shallow subsection, we then present a distillation technique that enables a skeptical student model to learn even from a nasty teacher. To evaluate the efficacy of our skeptical students, we conducted experiments with several models with KD on both training data-available and data-free scenarios for various datasets. While distilling from nasty teachers, compared to the normal student models, skeptical students consistently provide superior classification performance of up to ∼59.5%. Moreover, similar to normal students, skeptical students maintain high classification accuracy when distilled from a normal teacher, showing their efficacy irrespective of the teacher being nasty or not. We believe the ability of skeptical students to largely diminish the KD-immunity of potentially nasty teachers will motivate the research community to create more robust mechanisms for model confidentiality. We have open-sourced the code at https://github.com/ksouvik52/Skeptical2021

The increasing demand for on-chip edge intelligence has motivated the exploration of algorithmic techniques and specialized hardware to reduce the computation energy of current machine learning models. In particular, deep spiking neural networks (SNNs) have gained interest because their event-driven hardware implementations can consume very low energy. However, minimizing average spiking activity and thus energy consumption while preserving accuracy in deep SNNs remains a significant challenge and opportunity. This paper proposes a novel two-step SNN compression technique to reduce their spiking activity while maintaining accuracy that involves compressing specifically-designed artificial neural networks (ANNs) that are then converted into the target SNNs. Our approach uses an ultra-high ANN compression technique that is guided by the attention-maps of an uncompressed meta-model. We then evaluate the firing threshold of each ANN layer and start with the trained ANN weights to perform a sparse-learning-based supervised SNN training to minimize the number of time steps required while retaining compression. To evaluate the merits of the proposed approach, we performed experiments with variants of VGG and ResNet, on both CIFAR-10 and CIFAR-100, and VGG16 on Tiny-ImageNet. SNN models generated through the proposed technique yield state-of-the-art compression ratios of up to 33.4× with no significant drop in accuracy compared to baseline unpruned counterparts. As opposed to the existing SNN pruning methods we achieve up to 8.3× better compression with no drop in accuracy. Moreover, compressed SNN models generated by our methods can have up to 12.2× better compute energy-efficiency compared to ANNs that have a similar number of parameters.

The high energy cost of processing deep convolutional neural networks impedes their ubiquitous deployment in energy-constrained platforms such as embedded systems and IoT devices. This article introduces convolutional layers with pre-defined sparse 2D kernels that have support sets that repeat periodically within and across filters. Due to the efficient storage of our periodic sparse kernels, the parameter savings can translate into considerable improvements in energy efficiency due to reduced DRAM accesses, thus promising significant improvements in the trade-off between energy consumption and accuracy for both training and inference. To evaluate this approach, we performed experiments with two widely accepted datasets, CIFAR-10 and Tiny ImageNet in sparse variants of the ResNet18 and VGG16 architectures. Compared to baseline models, our proposed sparse variants require up to ~82% fewer model parameters with 5.6× fewer FLOPs with negligible loss in accuracy for ResNet18 on CIFAR-10. For VGG16 trained on Tiny ImageNet, our approach requires 5.8× fewer FLOPs and up to ~83.3% fewer model parameters with a drop in top-5 (top-1) accuracy of only 1.2% (~2.1%). We also compared the performance of our proposed architectures with that of ShuffleNet and MobileNetV2. Using similar hyperparameters and FLOPs, our ResNet18 variants yield an average accuracy improvement of ~2.8%.