KISS Project



On April 21, we will present our latest results of the KISS Project at the german forum for AI. Please note that the event will be held in german.

The focus of the talk will be the methods and tools needed to efficiently execute neural networks on embedded systems.

We would like to invite you to join us!

Apart from this talk, two colleagues from the Fraunhofer IIS will also present the latest news on the topic of AI.

Feel free to take a closer look at the full schedule:

KISS project goals

Deep learning allows solving problems that were considered unsolvable before. This is achieved by learning how to solve a problem from a set of training data. On the one hand, this requires solutions to efficiently create and process such training data. On the other hand, a major challenge today is how to determine efficient implementations of related multi-layer neuronal networks.

The KISS project aims to develop and offer methods and tools that help achieve these goals. More specifically, on the one hand it researches tools that generate better and more complete training data and allow for a more efficient training process. On the other hand, tools are developed to transform trained deep neuronal networks into an efficient implementation on different possible hardware architectures, including many-core processors (CPU), graphical processing units (GPUs) and field-programmable integrated circuits (FPGAs).

© Fraunhofer IIS

KISS project – what we offer

The findings of the KISS project are offered in two forms:

  • Educational and training seminars to instruct future experts in both industry and university
  • Software tools and expertise to help industrial partners create competitive products

If you are interested, please send us an e-mail.

Possible applications

The KISS project addresses signal processing applications in a broader sense. More specifically, it focuses on image and audio processing applications. To ensure wide usability of the developed tools and methods, the following example applications are considered during research and development:

  • Speaker localization to identify the source of an audio signal in the 3D space
  • Acoustic scene and event classification to determine the circumstances a sound snippet has been recorded under.
  • Depth estimation for one or multiple images to determine a per-pixel depth value
  • Image segmentation and object detection

The developed tools and methods are of course not limited to these example applications.  We are happy to offer you an application tailored to meet your specific requirements.

Developed tools and methods

© Fraunhofer IIS

Training data generation and efficient training

Robust deep neuronal networks require an extensive training data set, which is very costly to create manually. KISS is focused on the design of methods and tools to simplify the process for image processing applications. Concepts include the combination of simulated and real data, as well as data augmentation and parallel training.

System-level design tools

With the research on system-level design tools, KISS provides the missing link in design automation to automatically transform a trained deep neuronal network into an efficient implementation. Starting from a Pytorch or Tensorflow implementation, the trained deep neuronal network is first reduced in size and automatically optimized for a selected target hardware platform. The results are stored in an established interchange format. State-of-the-art compilers can be used to generate the executable that can be integrated into the desired customer solution.

Seminars for industrial education and training

Industrial KISS seminars are based on the outcome of the performed research and will start in summer 2021.

If you are a company, and have a specific need concerning the training content, the seminar can also be tailored to your specific needs for an individual training. To know more about this possibility, please contact us.

Topic Content
Training data generation and efficient training This seminar will show different techniques for limiting the efforts necessary in training data generation. Moreover, it will explain how to make efficient use of the available computation and storage resources when training large image processing deep neuronal networks.
Deep neuronal network compression This seminar explains in detail the optimization techniques that can be applied to deep neuronal networks for a more efficient implementation (during inference).
System-level design for neuronal networks This seminar gives an introduction to the latest system design tools that allow transforming a trained neuronal network into an efficient implementation. It includes compilers and profilers as well as system-level optimization tools. Various target hardware platforms (CPU, GPU, TPU) are considered.

SLOHA Workshop

n order to strengthen scientific exchange, we organized a workshop on "System-level Design Methods for Deep Learning on Heterogeneous Architectures" as part of the DATE conference on the 5th of February.

We were able to win the following interesting  lectures from renowned participants, which you can now rewatch:

“In-Sensor ML - Heterogeneous Computing in a mW”
Luca Benini, ETH Zürich, Switzerland

“Specialization in Hardware Architectures for Deep Learning”
Michaela Blott, Xilinx Research, Ireland

“Vision Dynamics for controlling autonomous vehicles”
Sorin Grigorescu, Transilvania University of Brasov and Elektrobit Automotive, Romania

“SoCs for Autonomous Vehicles: Agile Design and Programmability Challenges”
Augusto Vega, IBM T. J. Watson Research Center, USA


About the project

»KISS« is a German abbreviation which corresponds to »Laboratory for system-level design of machine learning based signal processing applications«. It is funded by the federal Ministry of Education and Research under the project number 01IS19070A.

The KISS project is scheduled to run from October 1, 2019 to September 30, 2021.



© Fraunhofer IIS

Project partners

KISS is a joint project of Fraunhofer IIS and the Friedrich-Alexander-University Erlangen-Nürnberg (Hardware-Software-Co-Design-Chair). Find further information about the project and on the topic of artificial intelligence here.





The Friedrich-Alexander-Universität Erlangen-Nürnberg is a project partner.

Back to Audio and Media Technologies

Credits Header: Fraunhofer IIS/fotomek –