guest ::
| Home > Publications database > MemSpikingTM: Neuromorphic sequence learning with memristive in-memory computing - from algorithm to hardware demonstration |
| Home > Publications database > MemSpikingTM: Neuromorphic sequence learning with memristive in-memory computing - from algorithm to hardware demonstration |
| Poster (After Call) | FZJ-2024-04781 |
; ; ; ; ;
2024
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2024-04781
Abstract: Information processing in the neo-cortex happens in a sequential manner and sequence learning is considered a key functionality of the human brain. Even though there are machine learning solutions to these problems, they, unlike the biological brain, often require large amounts of training data and suffer from a high energy consumption. Therefore, in this project we take the neuromorphic approach of bringing the biological principles of sparse neural activity and in-memory computing into electronic hardware. Thereby, our goal is to achieve a robust and energy-efficient solution for sequence learning.The Hierarchical Temporal Memory[1] concept and its biologically plausible version, SpikingTM[2], describe a possible algorithm for sequence learning in the neo-cortex. We prove that this algorithm can operate with memristive synapses[3]. Memristive devices are an emerging non-volatile memory and a prominent candidate for in-memory computing substrates. In order to fully leverage the possibilities of these devices, we adapt the SpikingTM algorithm and create a complete analog / mixed signal system model around a synaptic array of memristive devices[4]. We demonstrate sequence learning by sparse neural activity and showcase that the use of memristive devices leads to a significant gain of energy efficiency. Lastly, we validate the system with real memristive synaptic arrays on a custom nanometer CMOS demonstrator chip by performing complex sequence learning tasks with our memristive algorithm (MemSpikingTM) on hardware[5].This project shows the complete neuromorphic journey from a bio-plausible algorithm for the brain functionality over a hardware-aware adaption for emerging memristive device technology and a complete system model to a successful hardware demonstration. We showcase that by combining the biological principles of sparse activity and connectivity with a memristive in-memory computing substrate, we can fulfil the promise of robust brain-like functionality and energy efficiency. [1] S. Ahmad and J. Hawkins, arXiv preprint arXiv:1503.07469,2015[2] Y. Bouhadjar et al., PLOS Computational Biology, 18.6, 2022[3] Y. Bouhadjar et al., Neuromorph. Comput. Eng., 3.3, 2023[4] S. Siegel et al., Neuromorph. Comput. Eng., 3.2, 2023[5] S. Siegel et al., Proceedings of the 2023 NICE Conference, 2023
|
The record appears in these collections: |
PDF