Novel Brain-Inspired Learning Paradigms for Large-Scale Neuronal Networks
- 0 Kommentarer
- 2684 Antal besök
The goal of this project is to produce a set of rules for synaptic plasticity and network reorganisation that describe the actual adaptive processes that take place in the living brain during learning and to port these rules into current and next-generation neuromorphic hardware.
Current designs of neurally inspired computing systems rely on learning rules that appear to be insufficient to port the superior adaptive and computational capabilities of biological neural systems into large-scale recurrent neural hardware system. This is not surprising, since most of these learning rules had to be extrapolated from results of neurobiological experiments in vitro. New experimental techniques in neurobiology – such as 2-photon laser-scanning microscopy, optogenetic cell activation, and dynamic clamp techniques – make it now possible to record the changes that really take place in the intact brain during learning. First results indicate that the rules for synaptic plasticity have in fact to be rewritten. In particular, it appears that local synaptic plasticity is gated in multiple ways by global factors such as neuromodulators and network states. One primary goal of this project is to apply and extend new cutting-edge experimental techniques to produce a set of rules for synaptic plasticity and network reorganisation that describe the actual adaptive processes that take place in the living brain during learning.
These new rules will be analysed by computational neuroscience experts and their consequences for learning in simulated large-scale networks of neurons and neurally inspired computing systems will be ascertained. The goal of this project is to port essential aspects of learning in the intact brain into current and next-generation neuromorphic hardware. New interchangeable software tools, that have recently been developed in the FP6 project FACETS, will be employed to carry out these investigations. Open questions that arise in these modelling studies will be addressed by changes in experimental protocols of the neuroscientists, building on long standing interdisciplinary collaborations among the partners.
We will achieve the project goals by focusing on the following four main objectives:
- to examine how local rules of synaptic changes, in particular spike-timing-dependent plasticity (STDP), are regulated by global factors
- We will elucidate the phenomenology and mechanisms of neuromodulatory-state-dependent STDP. This requires the use of new optogenetic tools for light-activation of neuromodulatory afferents in combination with state-of-the-art imaging, electrophysiology and computer modelling approaches.
- We will investigate in the intact brain the neuromodulatory state dependency of STDP-based learning. We will directly control the nucleus basalis magnocellularis (NBM), which in vivo governs acetylcholine release in the neocortex, while inducing STDP. This will overcome the limitations of in vitro studies of state-dependent STDP gating.
- We will investigate in the intact brain the network state dependency of STDP-based learning. Natural state fluctuations, as observed in vivo, together with a novel dynamic clamp interface which allows us to simulate the contextual effect of changes in ongoing synaptic bombardment at the postsynaptic site, will be used to constrain neural network activity, and the impact of this on the induction and expression of synaptic plasticity, and on dynamic modulation of network behaviour, will be examined.
- to directly measure neural network activity in the intact brain and to determine how it is regulated by global factors and reconfigured during learning
- A goal is to further advance two-photon imaging of neural network activity towards imaging of the same network over long time-spans (days to weeks),and to enable optical recordings at high speed.
- Using these imaging tools, we hope to clarify how local circuit dynamics is affected by cholinergic inputs.
- We plan to investigate how local neural network activity is reconfigured over days to weeks, especially during and following learning of a sensory discrimination task.
- to design and analyse novel non-Hebbian learning rules based on the experimental findings
- The experiments for Objectives 1 and 2 will provide previously not available data and clues about the actual mechanisms behind network learning in vivo. One goal is the extraction and analysis of the rules that govern these processes.
- We will also develop in a meta-theory in which different existing and newly developed learning rules using spike-based plasticity modulated by global factors appear as special cases.
- We expect that the systematic investigation and quantification of mechanisms by which biological organisms gate local learning rules by global factors will provide completely new concepts for the design of learning algorithms in machine learning and robotics.
- to implement such novel learning rules in recurrent neural network models, analyse their functional consequences and study feasibility of hardware implementation
- The results of Objectives 1-3 will be used to extend detailed computermodels for cortical microcircuits by including rules for plasticity of recurrent synaptic connections witin the circuit.
- We are aiming at the discovery of methods that dramatically increase the adaptability, resilience, and task-dependent network reconfiguration capability of novel neuro-inspired computing systems.
- We expect that the extracted learning principles will provide an attractive target for hardware emulations of adaptive cortical networks of neurons. Experts in neuromorphic hardware design will study the chances that these new learning rules will in fact be embodied in the next generation of neuromophic hardware in Europe.
The overall long-term vision of this project is to develop new design principles for adaptive, reconfigurable very-large-scale hardware systems implementing novel learning rules inspired by biological neural networks in vivo.
Learning mechanisms implemented in the brain appear to be much more robust and flexible than those currently used in neurally inspired computing systems. To confer the superior adaptive and computational capabilities of biological neural systems to large-scale recurrent neural hardware systems and other novel massively parallel computing devices, new and more sophisticated learning rules are needed.
Our long-term vision is that the learning rules for global gating of local learning, identified and explored in this project, will become ideal candidates for implementation in hardware. Conceptually the interaction of local factors that can be monitored and stored at the site of each connection with one or a few global factors is very attractive for hardware implementation. Previous collaborations of several partners of the project have shown that networks of spiking neurons can be implemented in a truly large-scale, parallel, mixed analog-digital hardware system. The inclusion of learning rules that go beyond the classical Hebbian or STDP rules for unsupervised learning, by including a third factor representing for example information on saliency or reward, will advance the hardware into a regime where a much broader class of learning tasks can be solved by these ultra-rapid machines.
Extracted from Brain-i-Nets