Today was my last working day as a postdoctoral researcher at TU Chemnitz, where I had been researching in the area of neuromorphic computing for the past six months. It was a very interesting endeavor, and I am grateful to have been part of it.
I worked on the „Print the Brain“ project (which was funded for just one year and recently officially ended), where my research focused on the development of 3D neuromorphic circuits with printed memristors. In my first five months here, I successfully advanced the field of neuromorphic computing through three significant research works:
– First, I developed SNNtrainer3D, a novel software application that addresses the need for more user-friendly tools in the field of SNNs. This tool provides an intuitive interface for the design, training and visualization of SNN architectures. By integrating Three.js, the application provides a three-dimensional visualization of network structures and thus enables a deeper understanding of model behavior. It supports training on the modified National Institute of Standards and Technology (MNIST) dataset and lays the foundation for future integration with physical memristor technology, which is an important step in making SNNs more accessible and understandable to AI researchers and practitioners. This work also led to a publication in an international journal just five months after I started working here.
– Secondly, I presented a novel method for the design and simulation of neuromorphic circuits using ex-situ trained SNNs. This approach shows high efficiency and accuracy in solving XOR and MNIST problems using specialized 2D crossbar circuit architectures integrated with operational amplifiers and Lapicque neuron models. The method showed robust inference capabilities and achieved 100% accuracy on XOR problems and over 89% on MNIST datasets. This work confirms the practicality of ex-situ trained SNNs in neuromorphic hardware and opens avenues for more efficient and powerful AI systems. I will try to publish the results of this work in a research paper soon.
– Finally, I developed an innovative approach to in-situ memristor learning by creating a memristor emulator in the PySpice library. This method allows SNNs to learn and adapt directly to the memristive hardware, eliminating the need for external training. The approach was tested on several datasets, including XPUE, a custom 3×3 image dataset, a 3×5 digit dataset called 012345, and a resized 10×10 MNIST dataset. The neuromorphic circuit showed successful pattern learning and outperformed other in-situ training simulations on SPICE. This progress paves the way for autonomous, adaptive neuromorphic circuits for pattern classification. I will try to publish the results of this work also in a research paper soon.
These works highlight my contributions to the advancement of neuromorphic computing and emphasize practical applications and innovative methods to improve the design, training, and implementation of SNNs.
In addition, while working here at pmTUC, I also satisfied my curiosity regarding the electrical characterization of printed silver ink through stacked paper layers and the modeling of memristors. I also found interesting things about the ERC Starting Grant and what matters when applying for it. By the way, did you know that there is no research paper in the literature at the moment that shows the use of Physics-Informed Neural Networks (PINNs) to model a printed memristor?
If you are interested in working in the area of neuromorphic computing, Dresden seems to have many interesting projects. As for myself, I will look into something new.
Neueste Kommentare