Jörg Henkel

Giorgos Armeniakos, Georgios Zervakis, Dimitrios Soudris et al.

Deep Neural Networks (DNNs) are very popular because of their high performance in various cognitive tasks in Machine Learning (ML). Recent advancements in DNNs have brought beyond human accuracy in many tasks, but at the cost of high computational complexity. To enable efficient execution of DNN inference, more and more research works, therefore, exploit the inherent error resilience of DNNs and employ Approximate Computing (AC) principles to address the elevated energy demands of DNN accelerators. This article provides a comprehensive survey and analysis of hardware approximation techniques for DNN accelerators. First, we analyze the state of the art and by identifying approximation families, we cluster the respective works with respect to the approximation type. Next, we analyze the complexity of the performed evaluations (with respect to the dataset and DNN size) to assess the efficiency, the potential, and limitations of approximate DNN accelerators. Moreover, a broad discussion is provided, regarding error metrics that are more suitable for designing approximate units for DNN accelerators as well as accuracy recovery approaches that are tailored to DNN inference. Finally, we present how Approximate Computing for DNN accelerators can go beyond energy efficiency and address reliability and security issues, as well.

Kilian Pfeiffer, Martin Rapp, Ramin Khalili et al.

With an increasing number of smart devices like internet of things (IoT) devices deployed in the field, offloadingtraining of neural networks (NNs) to a central server becomes more and more infeasible. Recent efforts toimprove users' privacy have led to on-device learning emerging as an alternative. However, a model trainedonly on a single device, using only local data, is unlikely to reach a high accuracy. Federated learning (FL)has been introduced as a solution, offering a privacy-preserving trade-off between communication overheadand model accuracy by sharing knowledge between devices but disclosing the devices' private data. Theapplicability and the benefit of applying baseline FL are, however, limited in many relevant use cases dueto the heterogeneity present in such environments. In this survey, we outline the heterogeneity challengesFL has to overcome to be widely applicable in real-world applications. We especially focus on the aspect ofcomputation heterogeneity among the participating devices and provide a comprehensive overview of recentworks on heterogeneity-aware FL. We discuss two groups: works that adapt the NN architecture and worksthat approach heterogeneity on a system level, covering Federated Averaging (FedAvg), distillation, and splitlearning-based approaches, as well as synchronous and asynchronous aggregation schemes.

Dimitrios Danopoulos, Georgios Zervakis, Kostas Siozios et al.

Current state-of-the-art employs approximate multipliers to address the highly increased power demands of DNN accelerators. However, evaluating the accuracy of approximate DNNs is cumbersome due to the lack of adequate support for approximate arithmetic in DNN frameworks. We address this inefficiency by presenting AdaPT, a fast emulation framework that extends PyTorch to support approximate inference as well as approximation-aware retraining. AdaPT can be seamlessly deployed and is compatible with the most DNNs. We evaluate the framework on several DNN models and application fields including CNNs, LSTMs, and GANs for a number of approximate multipliers with distinct bitwidth values. The results show substantial error recovery from approximate re-training and reduced inference time up to 53.9x with respect to the baseline approximate implementation.

Giorgos Armeniakos, Georgios Zervakis, Dimitrios Soudris et al.

Printed electronics (PE) feature low non-recurring engineering costs and low per unit-area fabrication costs, enabling thus extremely low-cost and on-demand hardware. Such low-cost fabrication allows for high customization that would be infeasible in silicon, and bespoke architectures prevail to improve the efficiency of emerging PE machine learning (ML) applications. However, even with bespoke architectures, the large feature sizes in PE constraint the complexity of the ML models that can be implemented. In this work, we bring together, for the first time, approximate computing and PE design targeting to enable complex ML models, such as Multi-Layer Perceptrons (MLPs) and Support Vector Machines (SVMs), in PE. To this end, we propose and implement a cross-layer approximation, tailored for bespoke ML architectures. At the algorithmic level we apply a hardware-driven coefficient approximation of the ML model and at the circuit level we apply a netlist pruning through a full search exploration. In our extensive experimental evaluation we consider 14 MLPs and SVMs and evaluate more than 4300 approximate and exact designs. Our results demonstrate that our cross approximation delivers Pareto optimal designs that, compared to the state-of-the-art exact designs, feature 47% and 44% average area and power reduction, respectively, and less than 1% accuracy loss.

Giorgos Armeniakos, Georgios Zervakis, Dimitrios Soudris et al.

Printed Electronics (PE) exhibits on-demand, extremely low-cost hardware due to its additive manufacturing process, enabling machine learning (ML) applications for domains that feature ultra-low cost, conformity, and non-toxicity requirements that silicon-based systems cannot deliver. Nevertheless, large feature sizes in PE prohibit the realization of complex printed ML circuits. In this work, we present, for the first time, an automated printed-aware software/hardware co-design framework that exploits approximate computing principles to enable ultra-resource constrained printed multilayer perceptrons (MLPs). Our evaluation demonstrates that, compared to the state-of-the-art baseline, our circuits feature on average 6x (5.7x) lower area (power) and less than 1% accuracy loss.

Kilian Pfeiffer, Martin Rapp, Ramin Khalili et al.

Devices participating in federated learning (FL) typically have heterogeneous communication, computation, and memory resources. However, in synchronous FL, all devices need to finish training by the same deadline dictated by the server. Our results show that training a smaller subset of the neural network (NN) at constrained devices, i.e., dropping neurons/filters as proposed by state of the art, is inefficient, preventing these devices to make an effective contribution to the model. This causes unfairness w.r.t the achievable accuracies of constrained devices, especially in cases with a skewed distribution of class labels across devices. We present a novel FL technique, CoCoFL, which maintains the full NN structure on all devices. To adapt to the devices' heterogeneous resources, CoCoFL freezes and quantizes selected layers, reducing communication, computation, and memory requirements, whereas other layers are still trained in full precision, enabling to reach a high accuracy. Thereby, CoCoFL efficiently utilizes the available resources on devices and allows constrained devices to make a significant contribution to the FL system, increasing fairness among participants (accuracy parity) and significantly improving the final accuracy of the model.

Giorgos Armeniakos, Georgios Zervakis, Dimitrios Soudris et al.

Printed electronics (PE) promises on-demand fabrication, low non-recurring engineering costs, and sub-cent fabrication costs. It also allows for high customization that would be infeasible in silicon, and bespoke architectures prevail to improve the efficiency of emerging PE machine learning (ML) applications. Nevertheless, large feature sizes in PE prohibit the realization of complex ML models in PE, even with bespoke architectures. In this work, we present an automated, cross-layer approximation framework tailored to bespoke architectures that enable complex ML models, such as Multi-Layer Perceptrons (MLPs) and Support Vector Machines (SVMs), in PE. Our framework adopts cooperatively a hardware-driven coefficient approximation of the ML model at algorithmic level, a netlist pruning at logic level, and a voltage over-scaling at the circuit level. Extensive experimental evaluation on 12 MLPs and 12 SVMs and more than 6000 approximate and exact designs demonstrates that our model-to-circuit cross-approximation delivers power and area optimal designs that, compared to the state-of-the-art exact designs, feature on average 51% and 66% area and power reduction, respectively, for less than 5% accuracy loss. Finally, we demonstrate that our framework enables 80% of the examined classifiers to be battery-powered with almost identical accuracy with the exact designs, paving thus the way towards smart complex printed applications.

Ourania Spantidi, Georgios Zervakis, Iraklis Anagnostopoulos et al.

Deep Neural Networks (DNNs) are being heavily utilized in modern applications and are putting energy-constraint devices to the test. To bypass high energy consumption issues, approximate computing has been employed in DNN accelerators to balance out the accuracy-energy reduction trade-off. However, the approximation-induced accuracy loss can be very high and drastically degrade the performance of the DNN. Therefore, there is a need for a fine-grain mechanism that would assign specific DNN operations to approximation in order to maintain acceptable DNN accuracy, while also achieving low energy consumption. In this paper, we present an automated framework for weight-to-approximation mapping enabling formal property exploration for approximate DNN accelerators. At the MAC unit level, our experimental evaluation surpassed already energy-efficient mappings by more than $\times2$ in terms of energy gains, while also supporting significantly more fine-grain control over the introduced approximation.

Abolfazl Younesi, Mohsen Ansari, MohammadAmin Fazli et al.

In today's digital age, Convolutional Neural Networks (CNNs), a subset of Deep Learning (DL), are widely used for various computer vision tasks such as image classification, object detection, and image segmentation. There are numerous types of CNNs designed to meet specific needs and requirements, including 1D, 2D, and 3D CNNs, as well as dilated, grouped, attention, depthwise convolutions, and NAS, among others. Each type of CNN has its unique structure and characteristics, making it suitable for specific tasks. It's crucial to gain a thorough understanding and perform a comparative analysis of these different CNN types to understand their strengths and weaknesses. Furthermore, studying the performance, limitations, and practical applications of each type of CNN can aid in the development of new and improved architectures in the future. We also dive into the platforms and frameworks that researchers utilize for their research or development from various perspectives. Additionally, we explore the main research fields of CNN like 6D vision, generative models, and meta-learning. This survey paper provides a comprehensive examination and comparison of various CNN architectures, highlighting their architectural differences and emphasizing their respective advantages, disadvantages, applications, challenges, and future trends.

Dimitrios Danopoulos, Georgios Zervakis, Dimitrios Soudris et al.

Vision Transformer (ViT) models which were recently introduced by the transformer architecture have shown to be very competitive and often become a popular alternative to Convolutional Neural Networks (CNNs). However, the high computational requirements of these models limit their practical applicability especially on low-power devices. Current state-of-the-art employs approximate multipliers to address the highly increased compute demands of DNN accelerators but no prior research has explored their use on ViT models. In this work we propose TransAxx, a framework based on the popular PyTorch library that enables fast inherent support for approximate arithmetic to seamlessly evaluate the impact of approximate computing on DNNs such as ViT models. Using TransAxx we analyze the sensitivity of transformer models on the ImageNet dataset to approximate multiplications and perform approximate-aware finetuning to regain accuracy. Furthermore, we propose a methodology to generate approximate accelerators for ViT models. Our approach uses a Monte Carlo Tree Search (MCTS) algorithm to efficiently search the space of possible configurations using a hardware-driven hand-crafted policy. Our evaluation demonstrates the efficacy of our methodology in achieving significant trade-offs between accuracy and power, resulting in substantial gains without compromising on performance.

Kilian Pfeiffer, Mohamed Aboelenien Ahmed, Ramin Khalili et al.

In recent years, Large Language Models (LLMs) through Transformer structures have dominated many machine learning tasks, especially text processing. However, these models require massive amounts of data for training and induce high resource requirements, particularly in terms of the large number of Floating Point Operations (FLOPs) and the high amounts of memory needed. To fine-tune such a model in a parameter-efficient way, techniques like Adapter or LoRA have been developed. However, we observe that the application of LoRA, when used in federated learning (FL), while still being parameter-efficient, is memory and FLOP inefficient. Based on that observation, we develop a novel layer finetuning scheme that allows devices in cross-device FL to make use of pretrained neural networks (NNs) while adhering to given resource constraints. We show that our presented scheme outperforms the current state of the art when dealing with homogeneous or heterogeneous computation and memory constraints and is on par with LoRA regarding limited communication, thereby achieving significantly higher accuracies in FL training.

Mohamed Aboelenien Ahmed, Kilian Pfeiffer, Heba Khdr et al.

Training on edge devices poses several challenges as these devices are generally resource-constrained, especially in terms of power. State-of-the-art techniques at the device level reduce the GPU frequency to enforce power constraints, leading to a significant increase in training time. To accelerate training, we propose to jointly adjust the system and application parameters (in our case, the GPU frequency and the batch size of the training task) while adhering to the power constraints on devices. We introduce a novel cross-layer methodology that combines predictions of batch size efficiency and device profiling to achieve the desired optimization. Our evaluation on real hardware shows that our method outperforms the current baselines that depend on state of the art techniques, reducing the training time by $2.4\times$ with results very close to optimal. Our measurements also indicate a substantial reduction in the overall energy used for the training process. These gains are achieved without reduction in the performance of the trained model.

Mohamed Aboelenien Ahmed, Kilian Pfeiffer, Ramin Khalili et al.

Federated fine-tuning offers a promising approach for tuning Large Language Models (LLMs) on edge devices while preserving data privacy. However, fine-tuning these models on edge devices remains challenging due to high memory, communication, and computational demands. Zero-order optimization with task alignment provides a potential solution, enabling fine-tuning with inference-level memory requirements but requires a longer convergence time. In this paper, we propose Federated Split-Perturbation Zero-order Optimization (FedSPZO) that divides the network into two blocks, applying a different number of perturbations per block in a computationally effective way, achieving faster convergence. Our evaluation shows a $2.5 - 7\times $ reduction in computation overhead compared to zero-order state of the art techniques in federated learning.

Konstantinos Balaskas, Andreas Karatzas, Christos Sad et al.

Deep Neural Networks (DNNs) have shown significant advantages in a wide variety of domains. However, DNNs are becoming computationally intensive and energy hungry at an exponential pace, while at the same time, there is a vast demand for running sophisticated DNN-based services on resource constrained embedded devices. In this paper, we target energy-efficient inference on embedded DNN accelerators. To that end, we propose an automated framework to compress DNNs in a hardware-aware manner by jointly employing pruning and quantization. We explore, for the first time, per-layer fine- and coarse-grained pruning, in the same DNN architecture, in addition to low bit-width mixed-precision quantization for weights and activations. Reinforcement Learning (RL) is used to explore the associated design space and identify the pruning-quantization configuration so that the energy consumption is minimized whilst the prediction accuracy loss is retained at acceptable levels. Using our novel composite RL agent we are able to extract energy-efficient solutions without requiring retraining and/or fine tuning. Our extensive experimental evaluation over widely used DNNs and the CIFAR-10/100 and ImageNet datasets demonstrates that our framework achieves $39\%$ average energy reduction for $1.7\%$ average accuracy loss and outperforms significantly the state-of-the-art approaches.

Kilian Pfeiffer, Ramin Khalili, Jörg Henkel

Federated learning (FL) is usually performed on resource-constrained edge devices, e.g., with limited memory for the computation. If the required memory to train a model exceeds this limit, the device will be excluded from the training. This can lead to a lower accuracy as valuable data and computation resources are excluded from training, also causing bias and unfairness. The FL training process should be adjusted to such constraints. The state-of-the-art techniques propose training subsets of the FL model at constrained devices, reducing their resource requirements for training. But these techniques largely limit the co-adaptation among parameters of the model and are highly inefficient, as we show: it is actually better to train a smaller (less accurate) model by the system where all the devices can train the model end-to-end, than applying such techniques. We propose a new method that enables successive freezing and training of the parameters of the FL model at devices, reducing the training's resource requirements at the devices, while still allowing enough co-adaptation between parameters. We show through extensive experimental evaluation that our technique greatly improves the accuracy of the trained model (by 52.4 p.p.) compared with the state of the art, efficiently aggregating the computation capacity available on distributed devices.

Martin Rapp, Ramin Khalili, Kilian Pfeiffer et al.

We study the problem of distributed training of neural networks (NNs) on devices with heterogeneous, limited, and time-varying availability of computational resources. We present an adaptive, resource-aware, on-device learning mechanism, DISTREAL, which is able to fully and efficiently utilize the available resources on devices in a distributed manner, increasing the convergence speed. This is achieved with a dropout mechanism that dynamically adjusts the computational complexity of training an NN by randomly dropping filters of convolutional layers of the model. Our main contribution is the introduction of a design space exploration (DSE) technique, which finds Pareto-optimal per-layer dropout vectors with respect to resource requirements and convergence speed of the training. Applying this technique, each device is able to dynamically select the dropout vector that fits its available resource without requiring any assistance from the server. We implement our solution in a federated learning (FL) system, where the availability of computational resources varies both between devices and over time, and show through extensive evaluation that we are able to significantly increase the convergence speed over the state of the art without compromising on the final accuracy.

Ourania Spantidi, Georgios Zervakis, Iraklis Anagnostopoulos et al.

Recent Deep Neural Networks (DNNs) managed to deliver superhuman accuracy levels on many AI tasks. Several applications rely more and more on DNNs to deliver sophisticated services and DNN accelerators are becoming integral components of modern systems-on-chips. DNNs perform millions of arithmetic operations per inference and DNN accelerators integrate thousands of multiply-accumulate units leading to increased energy requirements. Approximate computing principles are employed to significantly lower the energy consumption of DNN accelerators at the cost of some accuracy loss. Nevertheless, recent research demonstrated that complex DNNs are increasingly sensitive to approximation. Hence, the obtained energy savings are often limited when targeting tight accuracy constraints. In this work, we present a dynamically configurable approximate multiplier that supports three operation modes, i.e., exact, positive error, and negative error. In addition, we propose a filter-oriented approximation method to map the weights to the appropriate modes of the approximate multiplier. Our mapping algorithm balances the positive with the negative errors due to the approximate multiplications, aiming at maximizing the energy reduction while minimizing the overall convolution error. We evaluate our approach on multiple DNNs and datasets against state-of-the-art approaches, where our method achieves 18.33% energy gains on average across 7 NNs on 4 different datasets for a maximum accuracy drop of only 1%.

Sami Salamin, Georgios Zervakis, Ourania Spantidi et al.

Transistor aging is one of the major concerns that challenges designers in advanced technologies. It profoundly degrades the reliability of circuits during its lifetime as it slows down transistors resulting in errors due to timing violations unless large guardbands are included, which leads to considerable performance losses. When it comes to Neural Processing Units (NPUs), where increasing the inference speed is the primary goal, such performance losses cannot be tolerated. In this work, we are the first to propose a reliability-aware quantization to eliminate aging effects in NPUs while completely removing guardbands. Our technique delivers a graceful inference accuracy degradation over time while compensating for the aging-induced delay increase of the NPU. Our evaluation, over ten state-of-the-art neural network architectures trained on the ImageNet dataset, demonstrates that for an entire lifetime of 10 years, the average accuracy loss is merely 3%. In the meantime, our technique achieves 23% higher performance due to the elimination of the aging guardband.

Georgios Zervakis, Ourania Spantidi, Iraklis Anagnostopoulos et al.

In this work, we introduce a control variate approximation technique for low error approximate Deep Neural Network (DNN) accelerators. The control variate technique is used in Monte Carlo methods to achieve variance reduction. Our approach significantly decreases the induced error due to approximate multiplications in DNN inference, without requiring time-exhaustive retraining compared to state-of-the-art. Leveraging our control variate method, we use highly approximated multipliers to generate power-optimized DNN accelerators. Our experimental evaluation on six DNNs, for Cifar-10 and Cifar-100 datasets, demonstrates that, compared to the accurate design, our control variate approximation achieves same performance and 24% power reduction for a merely 0.16% accuracy loss.

Martin Rapp, Ramin Khalili, Jörg Henkel

We consider a distributed system, consisting of a heterogeneous set of devices, ranging from low-end to high-end. These devices have different profiles, e.g., different energy budgets, or different hardware specifications, determining their capabilities on performing certain learning tasks. We propose the first approach that enables distributed learning in such a heterogeneous system. Applying our approach, each device employs a neural network (NN) with a topology that fits its capabilities; however, part of these NNs share the same topology, so that their parameters can be jointly learned. This differs from current approaches, such as federated learning, which require all devices to employ the same NN, enforcing a trade-off between achievable accuracy and computational overhead of training. We evaluate heterogeneous distributed learning for reinforcement learning (RL) and observe that it greatly improves the achievable reward on more powerful devices, compared to current approaches, while still maintaining a high reward on the weaker devices. We also explore supervised learning, observing similar gains.

Virinchi Roy Surabhi, Prashanth Krishnamurthy, Hussam Amrouch et al.

This paper reports a novel approach that uses transistor aging in an integrated circuit (IC) to detect hardware Trojans. When a transistor is aged, it results in delays along several paths of the IC. This increase in delay results in timing violations that reveal as timing errors at the output of the IC during its operation. We present experiments using aging-aware standard cell libraries to illustrate the usefulness of the technique in detecting hardware Trojans. Combining IC aging with over-clocking produces a pattern of bit errors at the IC output by the induced timing violations. We use machine learning to learn the bit error distribution at the output of a clean IC. We differentiate the divergence in the pattern of bit errors because of a Trojan in the IC from this baseline distribution. We simulate the golden IC and show robustness to IC-to-IC manufacturing variations. The approach is effective and can detect a Trojan even if we place it far off the critical paths. Results on benchmarks from the Trust-hub show a detection accuracy of $\geq$99%.