Vojtech Mrazek

Alberto Marchisio, Vojtech Mrazek, Andrea Massa et al.

Neural Architecture Search (NAS) algorithms aim at finding efficient Deep Neural Network (DNN) architectures for a given application under given system constraints. DNNs are computationally-complex as well as vulnerable to adversarial attacks. In order to address multiple design objectives, we propose RoHNAS, a novel NAS framework that jointly optimizes for adversarial-robustness and hardware-efficiency of DNNs executed on specialized hardware accelerators. Besides the traditional convolutional DNNs, RoHNAS additionally accounts for complex types of DNNs such as Capsule Networks. For reducing the exploration time, RoHNAS analyzes and selects appropriate values of adversarial perturbation for each dataset to employ in the NAS flow. Extensive evaluations on multi - Graphics Processing Unit (GPU) - High Performance Computing (HPC) nodes provide a set of Pareto-optimal solutions, leveraging the tradeoff between the above-discussed design objectives. For example, a Pareto-optimal DNN for the CIFAR-10 dataset exhibits 86.07% accuracy, while having an energy of 38.63 mJ, a memory footprint of 11.85 MiB, and a latency of 4.47 ms.

Ondrej Vlcek, Vojtech Mrazek

Approximate computing offers promising energy efficiency benefits for error-tolerant applications, but discovering optimal approximations requires extensive design space exploration (DSE). Predicting the accuracy of circuits composed of approximate components without performing complete synthesis remains a challenging problem. Current machine learning approaches used to automate this task require retraining for each new circuit configuration, making them computationally expensive and time-consuming. This paper presents ApproxGNN, a construction methodology for a pre-trained graph neural network model predicting QoR and HW cost of approximate accelerators employing approximate adders from a library. This approach is applicable in DSE for assignment of approximate components to operations in accelerator. Our approach introduces novel component feature extraction based on learned embeddings rather than traditional error metrics, enabling improved transferability to unseen circuits. ApproxGNN models can be trained with a small number of approximate components, supports transfer to multiple prediction tasks, utilizes precomputed embeddings for efficiency, and significantly improves accuracy of the prediction of approximation error. On a set of image convolutional filters, our experimental results demonstrate that the proposed embeddings improve prediction accuracy (mean square error) by 50% compared to conventional methods. Furthermore, the overall prediction accuracy is 30% better than statistical machine learning approaches without fine-tuning and 54% better with fast finetuning.

Lukas Sekanina, Vojtech Mrazek

Approximate circuits often achieve exceptional trade-offs between computational accuracy and hardware efficiency, making them attractive for deployment as reusable Intellectual Property (IP) cores. However, safeguarding such circuits against piracy is critical for enabling sustainable commercialization of approximate computing. This work addresses the emerging challenge of IP protection and piracy detection in the context of approximate hardware. We introduce a novel adversarial threat model, approximate obfuscation, in which an attacker not only conceals the design through structural obfuscation but also introduces functional modifications to ensure that the resulting circuit exhibits nearly identical error characteristics and hardware metrics as the original IP. To counter this threat, we propose an automated framework that extracts and compares statistical error profiles of protected IP cores and suspicious circuits, enabling systematic detection of potential IP theft. Through extensive experiments on a diverse set of approximate multipliers, we analyze the resilience of different approximate multipliers against approximate obfuscation. Our results provide new insights into the interplay between obfuscation, approximation, and IP protection.

Alberto Marchisio, Andrea Massa, Vojtech Mrazek et al.

Deep Neural Networks (DNNs) have made significant improvements to reach the desired accuracy to be employed in a wide variety of Machine Learning (ML) applications. Recently the Google Brain's team demonstrated the ability of Capsule Networks (CapsNets) to encode and learn spatial correlations between different input features, thereby obtaining superior learning capabilities compared to traditional (i.e., non-capsule based) DNNs. However, designing CapsNets using conventional methods is a tedious job and incurs significant training effort. Recent studies have shown that powerful methods to automatically select the best/optimal DNN model configuration for a given set of applications and a training dataset are based on the Neural Architecture Search (NAS) algorithms. Moreover, due to their extreme computational and memory requirements, DNNs are employed using the specialized hardware accelerators in IoT-Edge/CPS devices. In this paper, we propose NASCaps, an automated framework for the hardware-aware NAS of different types of DNNs, covering both traditional convolutional DNNs and CapsNets. We study the efficacy of deploying a multi-objective Genetic Algorithm (e.g., based on the NSGA-II algorithm). The proposed framework can jointly optimize the network accuracy and the corresponding hardware efficiency, expressed in terms of energy, memory, and latency of a given hardware accelerator executing the DNN inference. Besides supporting the traditional DNN layers, our framework is the first to model and supports the specialized capsule layers and dynamic routing in the NAS-flow. We evaluate our framework on different datasets, generating different network configurations, and demonstrate the tradeoffs between the different output metrics. We will open-source the complete framework and configurations of the Pareto-optimal architectures at https://github.com/ehw-fit/nascaps.

Filip Vaverka, Vojtech Mrazek, Zdenek Vasicek et al.

Energy efficiency of hardware accelerators of deep neural networks (DNN) can be improved by introducing approximate arithmetic circuits. In order to quantify the error introduced by using these circuits and avoid the expensive hardware prototyping, a software emulator of the DNN accelerator is usually executed on CPU or GPU. However, this emulation is typically two or three orders of magnitude slower than a software DNN implementation running on CPU or GPU and operating with standard floating point arithmetic instructions and common DNN libraries. The reason is that there is no hardware support for approximate arithmetic operations on common CPUs and GPUs and these operations have to be expensively emulated. In order to address this issue, we propose an efficient emulation method for approximate circuits utilized in a given DNN accelerator which is emulated on GPU. All relevant approximate circuits are implemented as look-up tables and accessed through a texture memory mechanism of CUDA capable GPUs. We exploit the fact that the texture memory is optimized for irregular read-only access and in some GPU architectures is even implemented as a dedicated cache. This technique allowed us to reduce the inference time of the emulated DNN accelerator approximately 200 times with respect to an optimized CPU version on complex DNNs such as ResNet. The proposed approach extends the TensorFlow library and is available online at https://github.com/ehw-fit/tf-approximate.

Vojtech Mrazek, Zdenek Vasicek, Lukas Sekanina et al.

The state-of-the-art approaches employ approximate computing to reduce the energy consumption of DNN hardware. Approximate DNNs then require extensive retraining afterwards to recover from the accuracy loss caused by the use of approximate operations. However, retraining of complex DNNs does not scale well. In this paper, we demonstrate that efficient approximations can be introduced into the computational path of DNN accelerators while retraining can completely be avoided. ALWANN provides highly optimized implementations of DNNs for custom low-power accelerators in which the number of computing units is lower than the number of DNN layers. First, a fully trained DNN is converted to operate with 8-bit weights and 8-bit multipliers in convolutional layers. A suitable approximate multiplier is then selected for each computing element from a library of approximate multipliers in such a way that (i) one approximate multiplier serves several layers, and (ii) the overall classification error and energy consumption are minimized. The optimizations including the multiplier selection problem are solved by means of a multiobjective optimization NSGA-II algorithm. In order to completely avoid the computationally expensive retraining of DNN, which is usually employed to improve the classification accuracy, we propose a simple weight updating scheme that compensates the inaccuracy introduced by employing approximate multipliers. The proposed approach is evaluated for two architectures of DNN accelerators with approximate multipliers from the open-source "EvoApprox" library. We report that the proposed approach saves 30% of energy needed for multiplication in convolutional layers of ResNet-50 while the accuracy is degraded by only 0.6%. The proposed technique and approximate layers are available as an open-source extension of TensorFlow at https://github.com/ehw-fit/tf-approximate.

Michal Pinos, Lukas Sekanina, Vojtech Mrazek

Integrating the principles of approximate computing into the design of hardware-aware deep neural networks (DNN) has led to DNNs implementations showing good output quality and highly optimized hardware parameters such as low latency or inference energy. In this work, we present ApproxDARTS, a neural architecture search (NAS) method enabling the popular differentiable neural architecture search method called DARTS to exploit approximate multipliers and thus reduce the power consumption of generated neural networks. We showed on the CIFAR-10 data set that the ApproxDARTS is able to perform a complete architecture search within less than $10$ GPU hours and produce competitive convolutional neural networks (CNN) containing approximate multipliers in convolutional layers. For example, ApproxDARTS created a CNN showing an energy consumption reduction of (a) $53.84\%$ in the arithmetic operations of the inference phase compared to the CNN utilizing the native $32$-bit floating-point multipliers and (b) $5.97\%$ compared to the CNN utilizing the exact $8$-bit fixed-point multipliers, in both cases with a negligible accuracy drop. Moreover, the ApproxDARTS is $2.3\times$ faster than a similar but evolutionary algorithm-based method called EvoApproxNAS.

Jan Klhufek, Alberto Marchisio, Vojtech Mrazek et al.

Transformer neural networks achieve state-of-the-art accuracy across language and vision tasks, but their deployment on embedded hardware is hindered by stringent area, latency, and energy constraints. During inference, performance and efficiency are increasingly dominated by the Key--Value (KV) cache, whose memory footprint grows with sequence length, straining on-chip memory utilization. Although existing mechanisms such as Grouped-Query Attention (GQA) reduce KV cache requirements compared to Multi-Head Attention (MHA), effectively exploiting this reduction requires understanding how on-chip memory demand evolves over time. This work presents TRAPTI, a two-stage methodology that combines cycle-level inference simulation with time-resolved analysis of on-chip memory occupancy to guide design decisions. In the first stage, the framework obtains memory occupancy traces and memory access statistics from simulation. In the second stage, the framework leverages the traces to explore banked memory organizations and power-gating configurations in an offline optimization flow. We apply this methodology to GPT-2 XL and DeepSeek-R1-Distill-Qwen-1.5B under the same accelerator configuration, enabling a direct comparison of MHA and GQA memory profiles. The analysis shows that DeepSeek-R1-Distill-Qwen-1.5B exhibits a 2.72x reduction in peak on-chip memory utilization in this setting compared to GPT-2 XL, unlocking further opportunities for power-gating optimization.

Jan Klhufek, Miroslav Safar, Vojtech Mrazek et al.

Energy efficiency and memory footprint of a convolutional neural network (CNN) implemented on a CNN inference accelerator depend on many factors, including a weight quantization strategy (i.e., data types and bit-widths) and mapping (i.e., placement and scheduling of DNN elementary operations on hardware units of the accelerator). We show that enabling rich mixed quantization schemes during the implementation can open a previously hidden space of mappings that utilize the hardware resources more effectively. CNNs utilizing quantized weights and activations and suitable mappings can significantly improve trade-offs among the accuracy, energy, and memory requirements compared to less carefully optimized CNN implementations. To find, analyze, and exploit these mappings, we: (i) extend a general-purpose state-of-the-art mapping tool (Timeloop) to support mixed quantization, which is not currently available; (ii) propose an efficient multi-objective optimization algorithm to find the most suitable bit-widths and mapping for each DNN layer executed on the accelerator; and (iii) conduct a detailed experimental evaluation to validate the proposed method. On two CNNs (MobileNetV1 and MobileNetV2) and two accelerators (Eyeriss and Simba) we show that for a given quality metric (such as the accuracy on ImageNet), energy savings are up to 37% without any accuracy drop.

Vojtech Mrazek, Konstantinos Balaskas, Paula Carolina Lozano Duarte et al.

Printed electronics offer a promising alternative for applications beyond silicon-based systems, requiring properties like flexibility, stretchability, conformality, and ultra-low fabrication costs. Despite the large feature sizes in printed electronics, printed neural networks have attracted attention for meeting target application requirements, though realizing complex circuits remains challenging. This work bridges the gap between classification accuracy and area efficiency in printed neural networks, covering the entire processing-near-sensor system design and co-optimization from the analog-to-digital interface-a major area and power bottleneck-to the digital classifier. We propose an automated framework for designing printed Ternary Neural Networks with arbitrary input precision, utilizing multi-objective optimization and holistic approximation. Our circuits outperform existing approximate printed neural networks by 17x in area and 59x in power on average, being the first to enable printed-battery-powered operation with under 5% accuracy loss while accounting for analog-to-digital interfacing costs.

Michal Pinos, Vojtech Mrazek, Lukas Sekanina

There is a growing interest in automated neural architecture search (NAS) methods. They are employed to routinely deliver high-quality neural network architectures for various challenging data sets and reduce the designer's effort. The NAS methods utilizing multi-objective evolutionary algorithms are especially useful when the objective is not only to minimize the network error but also to minimize the number of parameters (weights) or power consumption of the inference phase. We propose a multi-objective NAS method based on Cartesian genetic programming for evolving convolutional neural networks (CNN). The method allows approximate operations to be used in CNNs to reduce the power consumption of a target hardware implementation. During the NAS process, a suitable CNN architecture is evolved together with approximate multipliers to deliver the best trade-offs between the accuracy, network size, and power consumption. The most suitable approximate multipliers are automatically selected from a library of approximate multipliers. Evolved CNNs are compared with common human-created CNNs of a similar complexity on the CIFAR-10 benchmark problem.

Alberto Marchisio, Vojtech Mrazek, Muhammad Abdullah Hanif et al.

Deep Neural Networks (DNNs) have been established as the state-of-the-art algorithm for advanced machine learning applications. Recently proposed by the Google Brain's team, the Capsule Networks (CapsNets) have improved the generalization ability, as compared to DNNs, due to their multi-dimensional capsules and preserving the spatial relationship between different objects. However, they pose significantly high computational and memory requirements, making their energy-efficient inference a challenging task. This paper provides, for the first time, an in-depth analysis to highlight the design and management related challenges for the (on-chip) memories deployed in hardware accelerators executing fast CapsNets inference. To enable an efficient design, we propose an application-specific memory hierarchy, which minimizes the off-chip memory accesses, while efficiently feeding the data to the hardware accelerator. We analyze the corresponding on-chip memory requirements and leverage it to propose a novel methodology to explore different scratchpad memory designs and their energy/area trade-offs. Afterwards, an application-specific power-gating technique is proposed to further reduce the energy consumption, depending upon the utilization across different operations of the CapsNets. Our results for a selected Pareto-optimal solution demonstrate no performance loss and an energy reduction of 79% for the complete accelerator, including computational units and memories, when compared to a state-of-the-art design executing Google's CapsNet model for the MNIST dataset.

David Hodan, Vojtech Mrazek, Zdenek Vasicek

Despite many successful applications, Cartesian Genetic Programming (CGP) suffers from limited scalability, especially when used for evolutionary circuit design. Considering the multiplier design problem, for example, the 5x5-bit multiplier represents the most complex circuit evolved from a randomly generated initial population. The efficiency of CGP highly depends on the performance of the point mutation operator, however, this operator is purely stochastic. This contrasts with the recent developments in Genetic Programming (GP), where advanced informed approaches such as semantic-aware operators are incorporated to improve the search space exploration capability of GP. In this paper, we propose a semantically-oriented mutation operator (SOMO) suitable for the evolutionary design of combinational circuits. SOMO uses semantics to determine the best value for each mutated gene. Compared to the common CGP and its variants as well as the recent versions of Semantic GP, the proposed method converges on common Boolean benchmarks substantially faster while keeping the phenotype size relatively small. The successfully evolved instances presented in this paper include 10-bit parity, 10+10-bit adder and 5x5-bit multiplier. The most complex circuits were evolved in less than one hour with a single-thread implementation running on a common CPU.

Milan Ceska, Jiri Matyas, Vojtech Mrazek et al.

We present a novel approach for designing complex approximate arithmetic circuits that trade correctness for power consumption and play important role in many energy-aware applications. Our approach integrates in a unique way formal methods providing formal guarantees on the approximation error into an evolutionary circuit optimisation algorithm. The key idea is to employ a novel adaptive search strategy that drives the evolution towards promptly verifiable approximate circuits. As demonstrated in an extensive experimental evaluation including several structurally different arithmetic circuits and target precisions, the search strategy provides superior scalability and versatility with respect to various approximation scenarios. Our approach significantly improves capabilities of the existing methods and paves a way towards an automated design process of provably-correct circuit approximations.

Alberto Marchisio, Vojtech Mrazek, Muhammad Abudllah Hanif et al.

Recent advances in Capsule Networks (CapsNets) have shown their superior learning capability, compared to the traditional Convolutional Neural Networks (CNNs). However, the extremely high complexity of CapsNets limits their fast deployment in real-world applications. Moreover, while the resilience of CNNs have been extensively investigated to enable their energy-efficient implementations, the analysis of CapsNets' resilience is a largely unexplored area, that can provide a strong foundation to investigate techniques to overcome the CapsNets' complexity challenge. Following the trend of Approximate Computing to enable energy-efficient designs, we perform an extensive resilience analysis of the CapsNets inference subjected to the approximation errors. Our methodology models the errors arising from the approximate components (like multipliers), and analyze their impact on the classification accuracy of CapsNets. This enables the selection of approximate components based on the resilience of each operation of the CapsNet inference. We modify the TensorFlow framework to simulate the injection of approximation noise (based on the models of the approximate components) at different computational operations of the CapsNet inference. Our results show that the CapsNets are more resilient to the errors injected in the computations that occur during the dynamic routing (the softmax and the update of the coefficients), rather than other stages like convolutions and activation functions. Our analysis is extremely useful towards designing efficient CapsNet hardware accelerators with approximate components. To the best of our knowledge, this is the first proof-of-concept for employing approximations on the specialized CapsNet hardware.

Zdenek Vasicek, Vojtech Mrazek, Lukas Sekanina

We propose an application-tailored data-driven fully automated method for functional approximation of combinational circuits. We demonstrate how an application-level error metric such as the classification accuracy can be translated to a component-level error metric needed for an efficient and fast search in the space of approximate low-level components that are used in the application. This is possible by employing a weighted mean error distance (WMED) metric for steering the circuit approximation process which is conducted by means of genetic programming. WMED introduces a set of weights (calculated from the data distribution measured on a selected signal in a given application) determining the importance of each input vector for the approximation process. The method is evaluated using synthetic benchmarks and application-specific approximate MAC (multiply-and-accumulate) units that are designed to provide the best trade-offs between the classification accuracy and power consumption of two image classifiers based on neural networks.

Vojtech Mrazek, Muhammad Abdullah Hanif, Zdenek Vasicek et al.

Approximate computing is an emerging paradigm for developing highly energy-efficient computing systems such as various accelerators. In the literature, many libraries of elementary approximate circuits have already been proposed to simplify the design process of approximate accelerators. Because these libraries contain from tens to thousands of approximate implementations for a single arithmetic operation it is intractable to find an optimal combination of approximate circuits in the library even for an application consisting of a few operations. An open problem is "how to effectively combine circuits from these libraries to construct complex approximate accelerators". This paper proposes a novel methodology for searching, selecting and combining the most suitable approximate circuits from a set of available libraries to generate an approximate accelerator for a given application. To enable fast design space generation and exploration, the methodology utilizes machine learning techniques to create computational models estimating the overall quality of processing and hardware cost without performing full synthesis at the accelerator level. Using the methodology, we construct hundreds of approximate accelerators (for a Sobel edge detector) showing different but relevant tradeoffs between the quality of processing and hardware cost and identify a corresponding Pareto-frontier. Furthermore, when searching for approximate implementations of a generic Gaussian filter consisting of 17 arithmetic operations, the proposed approach allows us to identify approximately $10^3$ highly important implementations from $10^{23}$ possible solutions in a few hours, while the exhaustive search would take four months on a high-end processor.