Active Projects

Green Deep Learning for Edge Devices

Despite the continuous improvement of deep learning (DL) design and deployment frameworks, an energy-efficient design process guaranteeing user constraints (accuracy, latency, and energy consumption) is still missing from the energy saving perspective.

GreenDL aims to develop theoretical foundations and practical algorithms that (i) enable designing scalable and energy-efficient DL models with low energy footprint and (ii) facilitate fast deployment of complicated DL models for a diverse set of Edge devices satisfying given hardware constraints. To address research challenges, we will design the greenDL framework for energy-efficient design and deployment of DLs on Edge devices. 

Dependable AI in Safe Autonomous Systems

Data-driven development methods show great promise in producing accurate models for perception functions such as object detection and semantic segmentation, however, most of them lack a holistic view for being implemented in dependable systems. This project proposal aims at producing Machine Learning (ML) models of robust nature to meet and stay ahead of emerging certification requirements. A large part of the accuracy and robustness of a trained model is due to the data it was trained on, yet most research today focuses on model architecture development. It is the intention of this project to emphasize the dataset side of the problem, including novel methods of data augmentation e.g. neural augmentation. The expected outputs of the project would be to set the basis of a safety-conscious ML system and provide the methodology to iterate and refine such systems.

Dependable Deep Learning for Safety-Critical Airborne Embedded Systems

Deep neural networks (DNNs) have shown to be very successful in several areas, e.g. for object detection in autonomous cars. DNNs may also be successful in airborne systems. One such possible application is guided landing. The enabling of safe landing in adverse weather conditions without full ground support from the instrument landing system, decreases aerospace greenhouse gas emissions as multiple landing attempts and aerospace congestion are mitigated. To land autonomously without support from ground infrastructure requires advanced airborne systems including algorithms for detecting the runway. These systems are safety-critical.
This project addresses design methods for the use of DNNs in airborne safety-critical systems. DNNs cannot rely on traditional design assurance techniques described in documents from certification authorities or standardization bodies. In this project, the research focus is on mitigation techniques for design errors in both hardware and software and for adversarial effects which can lead to system failures. The expected results are design methodologies and fault tolerant architectures for airborne safety-critical applications using neural networks.

Developing Predictable Vehicle Software Utilizing Time Sensitive Networking

Recent advancement in the functionality and new customer features in modern vehicles, especially autonomous vehicles, requires massive computational power and high-bandwidth on-board real-time communication. While there is a lot of research done to meet the challenge of computational requirements, relatively small efforts have been spent to deal with the challenge of supporting the high-bandwidth onboard communication requirements. In a recent effort to support high-bandwidth low-latency onboard real-time communication in modern vehicles, the IEEE Time-Sensitive Networking (TSN) task group has developed a set of standards targeting different classes of real-time traffic, support for time-triggered traffic at the same time as non-real-time traffic, support for resource reservation for different classes of traffic, support for clock synchronization, and providing several traffic shapers. However, a complete development support including modelling, timing analysis, configuration, deployment and execution for vehicular applications that use TSN is still missing from the state of the art. Since TSN is a complex technology with many options and configuration possibilities; taking full advantage of TSN in execution of complex vehicular functions is daunting task. This project aims at developing innovative techniques to provide a full-fledged development environment for vehicular applications that use TSN as the backbone for on-board communication. One unique characteristic of the project consortium is that it offers a clear value chain from academia (MDH), through the tool developer/vendor (Arcticus), and to the end user of the technology (Volvo), who will use the techniques and tools to develop prototype vehicles.

Heterogeneous systems - software-hardware integration

The need for high-performance computing is increasing at a daunting pace and computational heterogeneity is the answer. High-performance computing platforms are increasingly becoming heterogeneous, meaning that they contain a combination of different computational units such as CPUs, GPUs, FPGAs, and AI accelerators. This computational power is needed both in hyped products like autonomous vehicles, but also in (maybe) less obvious cases like industrial automation where future intelligent production will be based on smart, autonomous, and collaborative industrial robots.

When this diverse range of computing architectures are put together on a single board (or a single chip even); the main challenge is to maximize the use of the huge computational power and at the same time to meet several criteria like performance, energy efficiency, real-time response, and dependability. To overcome these challenges, programmers of heterogeneous systems are expected to write parallel software, explicitly describe potential parallelism in their code, and identify which computations should be executed by which type of computational units. Currently, these activities are mostly manual, thereby difficult, slow, and error-prone.

The overall goal of HERO is to provide a framework that enables development of optimized parallel software, automatic mapping of software to heterogeneous hardware platforms, and provision of automatic hardware acceleration for the developed software.

Through HERO, Mälardalen University and five companies will develop deep competence to bridge the syntactic and semantic gap between modeling and programming languages, as well as automatically manipulating artifacts for analysis and synthesis of software for multiple heterogeneous targets. We will be able to drastically enhance the current practices for the design, analysis, and synthesis of parallel software for heterogeneous platforms. We will advance the knowledge on how to design and implement efficient functions for next-generation advanced hardware platforms and develop support for hardware programming, thanks to automatic synthesis of accelerators for heterogeneous parallel platforms.

HERO represents a substantial step towards an innovative solution for systematic and efficient development of complex heterogeneous systems. The research conducted in HERO is expected to provide substantial advances to the current state of the art in (i) model-based development and resource analysis of parallel software, (ii) pre-runtime code-level resource analysis, and (iii) automatic hardware acceleration.

The HERO team is composed of a strong group of researchers covering all aspects of the Synergy, with proven research records, and a group of companies strategically important for Swedish industry. Moreover, the Embedded Systems research environment at Mälardalen University represents the ideal soil for HERO, where we draw from, and contribute, to the rich and deep competence in embedded systems.

Automatic Design of Safe, High-Performance and Compact Deep Learning Models for Autonomous Vehicles

Deep Neural Networks (DNN) are increasingly being used to support decision-making in autonomous vehicles. While DNN holds the promise of delivering valuable results in safety-critical applications, broad adoption of DNN systems will rely heavily on how the computation-intensive DNN could be customized and deployed on the resource-limited vehicle embedded hardware platform and also how much to trust their outputs. In this project, we will develop the AutoDeep framework to design performance-efficient DNNs suitable for deployment on embedded resources-limited computing platforms while enhancing the robustness of DNN models. The mission is to strengthen Swedish industrial competence and competitiveness in the area of deep learning in the context of autonomous systems through close collaboration between academia and industry. AutoDeep can have a tangible impact on designing DL architectures for safety-critical applications and thus a successful demonstration of AutoDeep can increase Swedish industry’s market shares in ICT sectors that produce safe and high-performance embedded computing platforms for autonomous systems. The project consortium consists of three partners, including the main applicant Mälardalen University (MDH), ZenseAct and Volvo Construction Equipment (VolvoCE). An outstanding characteristic of this consortium is that it provides a value chain from academia (MDH), who will develop the framework and offers customized DNNs for safety-critical applications and the end users of the technology (ZenseAct and VolvoCE) will use the framework results and the customized DNNs on their prototype vehicles.

Integrated Time Sensitive Networking and Legacy Communications in Predictable Vehicle-platforms

The functionality advancements, innovation and novel features in modern vehicles require numerous high data-rate sensors (e.g. lidars). These sensors generate hundreds of megabytes of data per second that needs to be communicated among onboard computers with predictable low latencies. The emerging IEEE Time-Sensitive Networking (TSN) standards provide a promising solution to address these requirements. TSN provides a high-speed and low-latency backbone onboard network that can connect to legacy communication devices and subsystems. This project embraces the novel possibility of seamlessly integrating TSN and low-cost legacy onboard communications to create an attractive solution for next-generation vehicles. Alas, the state-of-the-art lacks a holistic model-based software development and timing verification framework for these integrated onboard communications. The vehicle industry also lacks a model-based software development tool-chain to utilize these integrated onboard communications. The aim of INTERCONNECT is to develop novel techniques, frameworks, industrial tool prototypes and demonstrators for holistic model-based development of vehicular software functions on predictable platforms that use seamlessly integrated TSN and legacy onboard communications. The benefits for the vehicle industry include cost-efficient system development, better quality of developed functions to lower costs, better management and use of expensive and scarce computing and communication resources. A major trait of the project consortium is that it offers a clear value chain initiating from academia (MDH); through a tools developer (Arcticus Systems); and finally, to an end user of the technology (HIAB).

Predictable Software Development in Connected Vehicles Utilising Blended TSN-5G Networks

Modern vehicles in many segments of the vehicular domain need to communicate and collaborate to achieve a joint functionality in, e.g., an autonomous quarry, mine or a recycling site. To provide such functionality, these vehicles need to be equipped with high data-rate sensors (e.g., cameras and lidars). The large amount of data acquired from these sensors needs to be communicated within as well as among the vehicles with predictable low latencies. The traditional intra-vehicle communication (based on field buses) and inter-vehicle communication (based on WiFi and 4G) are becoming a bottleneck in meeting the high-bandwidth and low-latency communication requirements. The recently introduced IEEE Time-Sensitive Networking (TSN) standards and 5G communication offer promising solutions to address these requirements within and among the vehicles respectively. Alas, there is a lack of a holistic software development framework and execution environment for predictable vehicular systems that utilise blended TSN-5G communication. This lack hinders the vehicle industry from taking full advantage of these ground- breaking technologies. The aim of PROVIDENT is to develop novel techniques to provide a full- fledged holistic software development environment for vehicular systems that utilise blended TSN- 5G communication. The benefits for the vehicle industry include cost-efficient system development, better quality of developed functions to lower costs, and better use of expensive and scarce computing and communication resources. A major trait of the project consortium is that it offers a clear value chain initiating from academia (MDH); through a tools developer (Arcticus Systems); and finally to an end user of the technology (HIAB).

Finished Projects

Deep Learning Accelerator on Commercial Programmable Devices

DeepMaker aims to provide a framework to generate synthesizable accelerators of Deep Neural Networks (DNNs) that can be used for different FPGA fabrics. DeepMaker enables effective use of DNN acceleration in commercially available devices that can accelerate a wide range of applications without a need of costly FPGA reconfigurations.

Energy-Efficient Hardware Accelerator for Embedded Deep Learning

In this joint project, we aim at decreasing the power consumption and computation load of the current image processing platform by employing the concept of computation reuse. Computation reuse suggests temporarily storing and reusing the result of a recent arithmetic operation for anticipated subsequent operations with the same operands. Our proposal is motivated by the high degree of redundancy that we observed in arithmetic operations of neural networks where we show that an approximate computation reuse can eliminate up to 94% of arithmetic operation of simple neural networks. This leads to up to 80% reduction in power consumption, which directly translates to a considerable increase in battery life time. We further presented a mechanism to make a large neural network by connecting basic units in two UT-MDH joint works.