Cooperative Distributed ML:
In the recent years, a paradigm shift could be observed in machine learning. Instead of training a model on a centralized server, many devices, such as IoT devices or smartphones, perform cooperative distributed training ("federated learning"). However, these devices have limited resources for communication and computations (weak processor, low communication bandwidth, etc.). This topic studies challenges and solutions that arise in federated learning on devices with limited resources.

Energy and Thermal Management in Mobile Platforms:
Power and thermal management in mobile multi-core platforms is becoming more and more challenging due to the limited available cooling options and the increasing performance demands of the market. This proseminar will be your opportunity to explore the different techniques, trends and challenges of energy and thermal management in mobile platforms.

Efficient Processing of Deep Neural Network:
Deep neural networks (DNNs) have achieved great success in challenging tasks such as image classification, object detection, and robotics. Processing of DNNs comes at the cost of high computational complexity and there are techniques that enable efficient processing of DNNs to improve efficiency without sacrificing accuracy.
  This topic will provide a tutorial and survey about enabling the efficient processing of DNNs. It will start with an overview of DNNs and the different hardware platforms and architecture that support DNNs.

When Hardware Eavesdrop:
In the modern world, smart devices have become essential in our lives. They contain sensitive personal data about their users. Hence, securing the data is crucial. That is where cryptography comes in. Encrypting the data makes it infeasible for the attackers to use it. However, as these are electronic devices monitoring the electrical properties of the device can be used to reveal the data by looking for the right patterns.
  This is called side-channel leakage. This topic will discuss the side-channel leakage and how it can be combined with analysis to attack the systems. Moreover, it will go through the different methods that exist to prevent such attack.

Run-time Resource Management for Operating Systems
As the complexity of multi/many-core architectures increases, operating systems must evolve to adapt to the diversity of computing, memory and communication on-chip resources, as well the as the emerging goals and requirements of these complex systems. In this scenario, run-time (dynamic) resource management has been established as an effective technique to improve and balance critical metrics, such as performance, reliability, efficiency and quality of service (QoS). In this seminar, students will study the background and current trends in on-chip resource management, by identifying the nature of the chip's resources, the relevant metrics on high-end systems, and the state-the-art techniques to manage those resources, varying from models and heuristics to machine learning approaches.

Neural Network Model Optimization for Fast Execution on Embedded Devices
The ever-increasing complexity of neural network models makes use of it on embedded systems challenging. Current ML models grow at a higher rate than Moore's law can keep up with. This topic deals with state-of-the-art methodologies to make ML models run fast on embedded devices, such as low-rank factorization, quantization, and pruning of models.

Embedded Systems Security
Hardware Security is one of the key issues for embedded system dependability. Over the past decade, hardware Trojans have got immense attention and proved to be the biggest threat to the integrity, security, and trust in modern digital systems. A hardware Trojan is a malicious modification to an existing design/system that serves a shadow purpose besides its intended functionality and typically implements a trigger mechanism that activates a payload mechanism such that Trojan payload is available during operation. The hidden functionality is unspecified and undocumented and skilfully inserted to evade detection during functional testing.
  Many different techniques have been provided in the literature to detect such Trojans, reverse engineering, thermal, and power side-channel also machine learning as well. However, most of the techniques have been proposed for static architectures and very little effort has been spent to detect Trojans in dynamically reconfigurable architectures where the Trojan can be inserted into the bit-stream configuration file.
  This pro-seminar is aimed at exploring basics of open source tools for reconfigurable FPGA, Hardware Trojan Insertion/Detection, Trojan detection in a bitstream using ML etc.