Hauptinhalt
Topinformationen
Mitarbeiterverzeichnis
Study Project: Computing with Spikes (Part II)
Prof. Dr. Pascal Nieters
Dr. rer. nat. Farbod Nosrat Nezami
Veranstaltungstyp: Studienprojekt
TeilnehmerInnen:
Beschreibung:
Spiking neural networks (SNNs) are models of brain computation in which individual neurons communicate with spikes, just like the brain does. Compared to more successful artificial neural network models (think Deep Nets), SNNs still raise a number of unique and challenging research questions: (i) How do you learn in a neural network when error-backpropagation doesn’t work? (ii) And can you do this with local update (plasticity) rules that use local information only? (iii) How do you learn a neural network or neuron model that uses spikes very efficiently and still finds a good representation of a continuous input stream?. The potential upsides of spiking nets: We can build specialized, neuromorphic hardware for spike-based networks that promises low-latency real-time, low-power and adaptive computation for the next generation of A.I.
In this context, the study project will follow the neuromorphic tradition of building a system that starts with spike-based sensing: An event-based stereo-vision setup with 2 artificial retinas and a stereo audio sensor for which we will implement an artificial cochlea ourselves. Our goal is to put state-of-the-art theories and algorithms for SNN computation to rigorous testing. We will thus challenge ourselves to tackle multi-modal multi-sensor real-time integration of information – an incredibly challenging technical task that our brain is purpose-built for and handles with ease. Step-by-step, we will tackle aspects of this problem in spiking neural network models that range from simple to complex, from “this-works-but-is-not-how-the-brain-does-it” to “this-is-closer-to-the-brain-but-really-hard-to-make-work”.
The project and accompanying interdisciplinary course are primarily aimed at Master students with an interest in one or more of the following themes: modeling networks of dynamical systems (plastic neurons), simulating spiking neural networks, low-level programming and engineering of A.I. systems using unconventional technologies, staying up to date and documenting state-of-the-art research in SNNs, or supporting a potentially quite complex project organisationally.
Erstes Treffen:
Montag, 28.10.2024 14:00 - 16:00, Ort: 93/E12
Ort: 35/E23-E24: Mo. 12:00 - 14:00 (13x), 93/E12: Mo. 12:00 - 14:00 (1x) Do. 10:00 - 12:00 (13x)
Semester: WiSe 2024/25
Zeiten:Mo. 12:00 - 14:00 (wöchentlich), Ort: 35/E23-E24, 93/E12, Do. 10:00 - 12:00 (wöchentlich), Ort: 93/E12
Leistungsnachweis:
Veranstaltungsnummer:
8.3073
ECTS-Kreditpunkte:
12
Bereichseinordnung:
Veranstaltungen > Cognitive Science > Master-Programm
Prof. Dr. Pascal Nieters
Dr. rer. nat. Farbod Nosrat Nezami
Veranstaltungstyp: Studienprojekt
TeilnehmerInnen:
Beschreibung:
Spiking neural networks (SNNs) are models of brain computation in which individual neurons communicate with spikes, just like the brain does. Compared to more successful artificial neural network models (think Deep Nets), SNNs still raise a number of unique and challenging research questions: (i) How do you learn in a neural network when error-backpropagation doesn’t work? (ii) And can you do this with local update (plasticity) rules that use local information only? (iii) How do you learn a neural network or neuron model that uses spikes very efficiently and still finds a good representation of a continuous input stream?. The potential upsides of spiking nets: We can build specialized, neuromorphic hardware for spike-based networks that promises low-latency real-time, low-power and adaptive computation for the next generation of A.I.
In this context, the study project will follow the neuromorphic tradition of building a system that starts with spike-based sensing: An event-based stereo-vision setup with 2 artificial retinas and a stereo audio sensor for which we will implement an artificial cochlea ourselves. Our goal is to put state-of-the-art theories and algorithms for SNN computation to rigorous testing. We will thus challenge ourselves to tackle multi-modal multi-sensor real-time integration of information – an incredibly challenging technical task that our brain is purpose-built for and handles with ease. Step-by-step, we will tackle aspects of this problem in spiking neural network models that range from simple to complex, from “this-works-but-is-not-how-the-brain-does-it” to “this-is-closer-to-the-brain-but-really-hard-to-make-work”.
The project and accompanying interdisciplinary course are primarily aimed at Master students with an interest in one or more of the following themes: modeling networks of dynamical systems (plastic neurons), simulating spiking neural networks, low-level programming and engineering of A.I. systems using unconventional technologies, staying up to date and documenting state-of-the-art research in SNNs, or supporting a potentially quite complex project organisationally.
Erstes Treffen:
Montag, 28.10.2024 14:00 - 16:00, Ort: 93/E12
Ort: 35/E23-E24: Mo. 12:00 - 14:00 (13x), 93/E12: Mo. 12:00 - 14:00 (1x) Do. 10:00 - 12:00 (13x)
Semester: WiSe 2024/25
Zeiten:Mo. 12:00 - 14:00 (wöchentlich), Ort: 35/E23-E24, 93/E12, Do. 10:00 - 12:00 (wöchentlich), Ort: 93/E12
Leistungsnachweis:
Veranstaltungsnummer:
8.3073
ECTS-Kreditpunkte:
12
Bereichseinordnung:
Veranstaltungen > Cognitive Science > Master-Programm