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

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 12-13/03/2018

Andr´ e Gr¨ uning, EITN and University of Surrey 16th March 2018

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- 13/03/2018 Andr´ e Gr¨ uning, EITN and University of Surrey Introduction Neural Information Processing This workshop

1 Introduction 2 Neural Information Processing 3 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

What is SP4?

Subproject 4 “Theoretical Neuroscience”: Computational Neuroscience Reach out to experimentalists in SP1, SP2, SP3

get data to inform plasticity rules high-level analyses of data cast data into plasticity models.

Reach out to platforms in SP9, SP10, SP5, SP6

get plasticity rules into shape to implement on neuromorphic platforms.

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

What is CDP5?

Co-design project CDP5 “Plasticity and learning in large-scale systems” SGA1 (2016-2018): fact finding mission.explore implementation of learning rules on platforms, SGA2 (2018-2020) : deep learning on neuromorphic platforms.

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

Why this workshop?

What? Bring together researchers from

Computational Neuroscience and Machine Learning, and all those in between and around these topics.

Stimulate exchange and collaboration between researchers in these two fields. Why is this important? What does it mean to have “understood the brain”?

build machines that mechanistically simulate the brain?

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

Computation Neuroscience

Computational Neuroscience has made great progress eg identifying and modelling neural-, synapse- and system-levels of plasticity. However: the functional and computational roles of these plasticity mechanisms on a behavioural or performance level are less clear Why does the brain use a specific plasticity mechanism to support a computational function? And how exactly is computation (as opposed to classification) implemented on top of low-level neuroscientific plasticity processes?

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

Machine Learning

Machine Learning has made great progress for example with the recent paradigm of deep learning: Simulations and cognitive learning models that were abandoned in the nineties due to do lack of hardware computational power can now be modelled and even implemented in a competitive way. Data analysis possible, but does this also give us mechanistic understanding?

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

Interdisciplinarity

Improve Information flow between these two fields Exaggerating a bit: The classical stance of a computational neuroscientist is that any machine learning approaches to learning and behaviour are not relevant:

because they are not biologically plausible;

Likewise a researcher in machine learning may dismiss computational neuroscience approaches

as just not competitive and performant.

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

Join Forces

Neuroscience and Machine Learning need to get together and learn from each other: How low-level plasticity might support high-level behavioural and/or technical learning behaviour; How high-level task-driven optimisation approaches realise themselves in low-level biological constraints. What machine learning could learn from neuroscience,

for example the sparse and energy-efficient encoding using spikes.

What computational neuroscience can learn from machine learning,

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- 13/03/2018 Andr´ e Gr¨ uning, EITN and University of Surrey Introduction Neural Information Processing This workshop

1 Introduction 2 Neural Information Processing 3 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

Why neurally inspired information processing?

Brain still does better Intelligence at the power of a light bulb. Poorly understood ⇒ understand it better. Glean principles from it for next-generation computing. Physical (quantum) limits of integration Miniaturisation comes to an end. HPC simulating brain consumes as much power as a town. HPC too slow for extensive simulations

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

Biological Information Processing: Spiking Neurons

(a)

(c)

• utput spike

input spikes

• utput spike

input spikes u

(b)

from Gruning A, Bohte S, Proceedings of the 22nd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning – ESANN, Brugge, 2014.

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

Biological Information Processing: Spiking Neurons

(a)

(c)

• utput spike

input spikes

• utput spike

input spikes u

(b)

from Gruning A, Bohte S, Proceedings of the 22nd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning – ESANN, Brugge, 2014.

Forward signal transmission is completely(?) understood: No conceptual scientific challenge. Abstracted into artificial neural networks.

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

Neurally inspired information processing

Synapses between neurons change with “experience” / past activities: Hebbian “Neurons that fire together, wire together.” Generalisations: anything correlation based on pre- and postsynaptic activity goes:

∆w = f (APre, APost)

In Machine Learning: eg associative maps, SOM, . . . Essentially understood, no challenge.

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

Beyond Hebb – Nature

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.

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

Neurally inspired information processing

Beyond Hebb – Nature Pre- and post synaptic activity and third signals:

∆w = f (APre, APost, r). r: third signal, eg reward, error, attention.

Lots of evidence for third signals:

three-partite synapses / astrocytes, neuromodulators: Dopa, ACh, Ser, . . . computations in local compartments (dendritic spikes), retrograde signal transmission: endocannabinoids, NO, . . .

What is the functional / computational role of third signals?

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

Neurally inspired information processing

Beyond Hebb - Artificial Neural Networks Abstraction to learning rules for classical ANN: reinforcement learning, supervised learning, eg gradient descent / backpropagation. Challenges Performance not as good as human. Technical abstraction is not plausible in nature. no KT from nature to technology: lose out on nature’s

• pportunities

no KT from technology to nature: lose out on understanding nature’s computational ways.

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

Challenges – Example

Take eg Back-Prop

Back-prop Nature Requires feedback 1-1. Connections reciprocal only in bulk, not 1-1. Requires immediate feedback. Retrograde activity 10-100ms. Derivatives of smooth quantities. Spikes are discrete events.

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

Challenges

Conclusion One hand: unrealistic assumptions about error feedback. Other hand: not utilising what nature provides. Why is nature the way it is? Better ways of doing computation?

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

1 Introduction 2 Neural Information Processing 3 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

Standard Conferences/Seminar/Talk/Paper

Speaker wants to give best impression possible, wants to present their research in the best light possible, leaves out all points of doubt and (funding, competition, publication etc pp)

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

Standard Conferences/Seminar/Talk/Paper

Speaker wants to give best impression possible, wants to present their research in the best light possible, leaves out all points of doubt and (funding, competition, publication etc pp) My experience: this is not how science works.

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

This Workshop

Open exchange of ideas: Questions allowed anytime. Do talk about the 99 routes that did not work: Do talk about your brilliant results, but also points where

you are stuck, in doubt, went wrong, need input/help/support/motivation want to reach out to others would like to input to other research, shape community focus

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

Presentations

I would like you to upload your presentations to Zenodo: Zenodo An open access repository: supported by the EU and the H2020 programme. can automatically reference your grant / reporting to EU. takes all kinds of artefacts: presentations, data, video clip etc (and pre-prints) coins a DOI. I created a “Collection” for the workshop. In line with EU and many journals’ open access policy, putting material on Zenodo does not pre-empt publication, check http://www.sherpa.ac.uk/romeo/index.php Please discuss with me if you have concerns regarding uploading (part of) your presentation.

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

Presentation Upload

https://zenodo.org/deposit/new?c=from_ neuroscience_to_machine_learning