The focus of our eSPML projects lies within the scope of SAL research program areas “Advanced Signal Processing and Integrated Circuit Design for power and complexity reduction” and “Embedded SW/HW for smart and distributed systems”, not excluding other programs’ major research topics and applications where machine learning approaches and signal processing sub-problems provide substantial benefit.

We plan to integrate and combine deep learning methods into/with conventional physical models in order to improve nonlinear modeling for electronic devices like e.g. communications and radio frequency (RF) transceivers. In particular, we will enhance or replace conventional signal processing models like e.g. Volterra series, Wiener-models, Hammerstein models or memory polynomials with new techniques of machine learning. Involved research fields are deep learning, deep unfolding, model –based neural networks, few shot learning, meta learning (transfer models from one device to another), sequence analysis using LSTMs and transformers/BERT.

Reinforcement learning methods should be considered for improving models that have delayed signals, therefore also delayed error or reward signals. The goal is to enhance existing physical models by new deep learning techniques to make model simulation faster without losing precision and modeling capacity or to improve existing models using new machine learning approaches. Also online and tracking models should be investigated, that is, models that adapt to the current environment like temperature, humidity, power consumption, voltage deviations, fast changing system parameters etc. These online methods might be combined or substituted by few shot and meta-learning methods.

Furthermore, we also plan to enhance and/or substitute model-based signal processing blocks like e.g. pre-coding, digital pre-distortion, nonlinear distortion mitigation, or adaptive channel estimation/equalization, by machine learning approaches.

Due to the ever incre­a­sing comple­xity of circuits and systems, the design of System on Chips starts at high levels of abstrac­tion. Model-based design is a state-of-the-art design metho­do­logy which addresses this deve­lop­ment. Here, initial drafts of the design are defined using mode­ling languages such as UML, SysML, etc. However, already at this stage severe design choices are made which will have a substan­tial impact, e.g., at the even­tual cost of the design. But properly esti­ma­ting those effects is hard to impos­sible using conven­tional methods, which is why those problems are thus far tackled, e.g., by “trial and error”, “gut feeling”, and “expe­ri­ence of the design engi­neer”. This frequently yields situa­tions where, after months of imple­men­ta­tions, design choices turn out to not satis­fying the desired objec­tives—a severe threat for time to market which is crucial in the EDA industry.  

Machine Learning re­search in the fields of deep learning, few shot learning, meta learning, rein­force­ment learning, sequence analysis methods like LSTM or Trans­former/BERT should be performed and new methods to provide improved cost esti­ma­tions should be deve­l­oped. Onto these esti­mates, design choices can be based to obtain opti­mized SoCs, and even­tually to auto­mate further steps of the design process. Possible industry part­ners (provi­ding corre­spon­dingly needed data and use cases) are avail­able. Results gene­rated by this project will directly be applied in indus­trial prac­tice.

Interested in cooperating with us?