Efficient Neural Networks Accelerators Using Error-Tolerant Embedded DRAMs (2 Semester Projects / 1 Master Project)
Description: The growing size of convolutional neural networks (CNNs) requires large amounts of on-chip storage. The limited on-chip memory capacity often causes massive off-chip memory accesses and leads to very high system energy consumption. On-chip memories are usually implemented with Static Random Access Memory (SRAM), which often consume over 50% of the total chip area due to their inefficient size, and dominates system power due to its limited voltage scaling capabilitits. Gain-Cell Embedded DRAM (GC-eDRAM) is a logic-compatible alternative to SRAM, offering up-to 2X higher density, two-ported operation and low power consumption. However, GC-eDRAM requires periodic refresh cycles to retain its data, which is set according to the worst-case data retention time under process, voltage, and temperature (PVT) variations.
In this project, we will evaluate the benefits of using GC-eDRAM as the on-chip memory for CNNs, including analysis of refresh rate relaxation and its impact on CNNs accuracy, power, and bandwidth. The project can focus on two parts of the development stack; first on a Python-based CNN simulator for which an GC-eDRAM behavioral model has to be implemented, second on an FPGA/C-based emulation platform of GC-eDRAM for which a CNN needs to be implemented.
Project 1: Python/C
Project 2: FPGA/C
Data added: 07.06.2019
Neural Networks for Self-Interference Cancellation
Description: Full-duplex radios can double throughput but they require cancellation of complex transceiver non-linearities. A neural network can learn to cancel the non-linearities through examples.
The goals of this project is to improve the existing feed-forward neural network canceller, and to apply the neural-network canceller to a testbed with higher transmit powers.
Areas: Python/Keras/Tensorflow, hands-on with full-duplex radio testbed
Supervisor: Alexios Balatsoukas-Stimming
Data added: 15.12.2017
Blind Detection using Neural Networks
Description: A wireless receiver does not always know what’s coming it’s way. There’s a need to quickly and efficiently detect if a message is present, and when it’s the case, to detect which channel code parameters were used. Neural networks are generally good for classification/detection applications.
Hence, for this project, the main goal is to build a neural network that can detect if a message has been encoded using a polar code, the family of channel code that has been selected to protect the control-channel data in the 5G standard. The secondary goal is to compare the performance of the proposed neural network against that of the state of the art techniques.
Areas: Communication, neural networks, channel coding, Tensorflow
Supervisors: Alexios Balatsoukas-Stimming, Pascal Giard, and Christoph Müller
Date added: 15.12.2017
FPGA Implementation of a Neural-Network Canceller
Description: Neural network based cancellation works well, but high-speed and low-area implementations of neural-network-based cancellers are essential to prove the concept.
Hence the goals of this project is to implement a real-time neural-network-based canceller on an FPGA. It can be seen as a continuation of the Neural Networks for Self-Interference Cancellation project.
Areas: FPGA, VHDL
Supervisors: Alexios Balatsoukas-Stimming and Pascal Giard
Date added: 15.12.2017