From Neuroscience to Machine Learning Learning CPD5/SP4 Workshop - PowerPoint PPT Presentation

From Neuroscience to Machine From Neuroscience to Machine Learning Learning CPD5/SP4 Workshop at CPD5/SP4 Workshop at EITN 12-13/03/2018 EITN 12- 13/03/2018 Andr´ e Gr¨ uning, EITN and University of Andr´ e Gr¨ uning, EITN and University of Surrey Surrey Introduction Neural Information Processing 16th March 2018 This workshop

From Neuroscience to Machine Learning CPD5/SP4 Workshop at EITN 12- 13/03/2018 Andr´ e Gr¨ uning, EITN and University of Surrey Introduction Neural Information Processing This workshop

From Neuroscience to Machine Learning CPD5/SP4 Workshop at EITN 12- 1 Introduction 13/03/2018 Andr´ e Gr¨ uning, EITN and University of 2 Neural Information Processing Surrey Introduction Neural Information 3 This workshop Processing This workshop

What is SP4? From Neuroscience to Machine Learning Subproject 4 “Theoretical Neuroscience”: CPD5/SP4 Workshop at EITN 12- 13/03/2018 Computational Neuroscience Andr´ e Gr¨ uning, Reach out to experimentalists in SP1, SP2, SP3 EITN and University of get data to inform plasticity rules Surrey high-level analyses of data Introduction cast data into plasticity models. Neural Information Reach out to platforms in SP9, SP10, SP5, SP6 Processing get plasticity rules into shape to implement on This workshop neuromorphic platforms.

What is CDP5? From Neuroscience to Machine Learning CPD5/SP4 Workshop at EITN 12- Co-design project CDP5 13/03/2018 “Plasticity and learning in large-scale systems” Andr´ e Gr¨ uning, EITN and University of SGA1 (2016-2018): fact finding mission.explore Surrey implementation of learning rules on platforms, Introduction SGA2 (2018-2020) : deep learning on neuromorphic Neural Information platforms. Processing This workshop

Why this workshop? From What? Neuroscience to Machine Learning CPD5/SP4 Bring together researchers from Workshop at EITN 12- 13/03/2018 Computational Neuroscience and Andr´ e Machine Learning , and Gr¨ uning, all those in between and around these topics. EITN and University of Surrey Stimulate exchange and collaboration between researchers in these two fields. Introduction Neural Information Processing Why is this important? This workshop What does it mean to have “understood the brain”? build machines that mechanistically simulate the brain?

Computation Neuroscience From Neuroscience to Machine Computational Neuroscience has made great progress Learning CPD5/SP4 Workshop at eg identifying and modelling neural-, synapse- and EITN 12- 13/03/2018 system-levels of plasticity. However: Andr´ e Gr¨ uning, EITN and the functional and computational roles of these plasticity University of Surrey mechanisms on a behavioural or performance level are less clear Introduction Why does the brain use a specific plasticity mechanism to Neural Information support a computational function? Processing And how exactly is computation (as opposed to This workshop classification) implemented on top of low-level neuroscientific plasticity processes?

Machine Learning From Neuroscience to Machine Learning CPD5/SP4 Machine Learning has made great progress Workshop at EITN 12- 13/03/2018 for example with the recent paradigm of deep learning: Andr´ e Gr¨ uning, EITN and Simulations and cognitive learning models that were University of Surrey abandoned in the nineties due to do lack of hardware computational power can now be modelled and even Introduction Neural implemented in a competitive way. Information Processing Data analysis possible, but does this also give us This workshop mechanistic understanding?

Interdisciplinarity From Neuroscience to Machine Learning Improve Information flow between these two fields CPD5/SP4 Workshop at EITN 12- Exaggerating a bit: 13/03/2018 Andr´ e Gr¨ uning, The classical stance of a computational neuroscientist is EITN and University of that any machine learning approaches to learning and Surrey behaviour are not relevant: Introduction because they are not biologically plausible; Neural Information Likewise a researcher in machine learning may dismiss Processing computational neuroscience approaches This workshop as just not competitive and performant.

Join Forces Neuroscience and Machine Learning need to get together and From Neuroscience learn from each other: to Machine Learning CPD5/SP4 How low-level plasticity might support high-level behavioural Workshop at EITN 12- and/or technical learning behaviour; 13/03/2018 How high-level task-driven optimisation approaches realise Andr´ e themselves in low-level biological constraints. Gr¨ uning, EITN and What machine learning could learn from neuroscience, University of Surrey for example the sparse and energy-efficient encoding using spikes. Introduction Neural What computational neuroscience can learn from machine Information learning, Processing This workshop for example what computational feature representations can be expected to evolve in a learning system. Our insights complement each other to understand what intelligent behaviour is and how it can be achieved in natural and artificial systems.

From Neuroscience to Machine Learning CPD5/SP4 Workshop at EITN 12- 1 Introduction 13/03/2018 Andr´ e Gr¨ uning, EITN and University of 2 Neural Information Processing Surrey Introduction Neural Information 3 This workshop Processing This workshop

Why neurally inspired information processing? From Brain still does better Neuroscience to Machine Learning CPD5/SP4 Intelligence at the power of a light bulb. Workshop at EITN 12- Poorly understood ⇒ understand it better. 13/03/2018 Glean principles from it for next-generation computing. Andr´ e Gr¨ uning, EITN and University of Surrey Physical (quantum) limits of integration Introduction Neural Miniaturisation comes to an end. Information Processing HPC simulating brain consumes as much power as a town. This workshop HPC too slow for extensive simulations

Biological Information Processing: Spiking Neurons From (a) Neuroscience input spikes to Machine Learning output spike CPD5/SP4 Workshop at EITN 12- (c) 13/03/2018 u output spike Andr´ e Gr¨ uning, (b) EITN and University of Surrey input spikes Introduction Neural from Gruning A, Bohte S, Proceedings of the 22nd European Symposium on Artificial Neural Networks, Information Computational Intelligence and Machine Learning – ESANN , Brugge, 2014. Processing This workshop

Biological Information Processing: Spiking Neurons From (a) Neuroscience input spikes to Machine Learning output spike CPD5/SP4 Workshop at EITN 12- (c) 13/03/2018 u output spike Andr´ e Gr¨ uning, (b) EITN and University of Surrey input spikes Introduction Neural from Gruning A, Bohte S, Proceedings of the 22nd European Symposium on Artificial Neural Networks, Information Computational Intelligence and Machine Learning – ESANN , Brugge, 2014. Processing This workshop Forward signal transmission is completely(?) understood: No conceptual scientific challenge. Abstracted into artificial neural networks.

Neurally inspired information processing From Neuroscience to Machine Synapses between neurons change with “experience” / past Learning CPD5/SP4 activities: Workshop at EITN 12- 13/03/2018 Hebbian Andr´ e Gr¨ uning, EITN and University of “Neurons that fire together, wire together.” Surrey Generalisations: anything correlation based on pre- and Introduction postsynaptic activity goes: Neural Information ∆ w = f ( A Pre , A Post ) Processing In Machine Learning: eg associative maps, SOM, . . . This workshop Essentially understood, no challenge.

Beyond Hebb – Nature From Neuroscience to Machine Learning CPD5/SP4 Workshop at EITN 12- 13/03/2018 Andr´ e Gr¨ uning, EITN and University of Surrey Introduction Neural Information Processing This workshop from Takata N, Mishima T, Hisatsune C, Nagai T, Ebisui E, Mikoshiba K, et al., Journal of Neuroscience , 31:18155, 2011, doi: 10.1523/JNEUROSCI.5289-11.2011 .

Neurally inspired information processing From Neuroscience to Machine Beyond Hebb – Nature Learning CPD5/SP4 Workshop at EITN 12- Pre- and post synaptic activity and third signals : 13/03/2018 ∆ w = f ( A Pre , A Post , r ). Andr´ e Gr¨ uning, r : third signal, eg reward, error, attention . EITN and University of Surrey Lots of evidence for third signals: Introduction three-partite synapses / astrocytes, Neural neuromodulators: Dopa, ACh, Ser, . . . Information computations in local compartments (dendritic spikes), Processing retrograde signal transmission: endocannabinoids, NO, . . . This workshop What is the functional / computational role of third signals?

Neurally inspired information processing From Beyond Hebb - Artificial Neural Networks Neuroscience to Machine Learning Abstraction to learning rules for classical ANN: CPD5/SP4 Workshop at EITN 12- 13/03/2018 reinforcement learning, Andr´ e supervised learning, eg gradient descent / backpropagation. Gr¨ uning, EITN and University of Surrey Challenges Introduction Neural Information Performance not as good as human. Processing Technical abstraction is not plausible in nature. This workshop no KT from nature to technology: lose out on nature’s opportunities no KT from technology to nature: lose out on understanding nature’s computational ways.