«Neuroscience Crash Course - Tackling the Brain Code» · 2017. 3. 14. · Neural Networks. Neural...

transcript

«Neuroscience Crash Course - Tackling the Brain Code»Pascal Kaufmann

NeuroscienceCrash Course

Agenda

I. Tackling the Brain: Motivation- Artificial Brains, interfacing brains with machines, boosting intelligence

II. Neuroscience Crash Course: Principles- Understanding the difference between a brain and a computer

III. Interfacing with brains: From neurointerface to artificial brains- What’s going on in neuroscience

Agenda

I. Tackling the Brain: Motivation- Artificial Brains and Bird Wings, interfacing brains with machines

Prometheus – stealing the fire from the Gods

2016, Zurich Landesmuseum

Tackling the human brain: What is the brain code?

2000Neurohybrid InterfaceProf. Sandro Mussa-IvaldiNorthwestern University, USA

1999Sieving electrodeProf. Martin SchwabPeter FromherzUniversity of Zurich, CHMax Plank Inst., GER

2004 - 2007Laser stimulation of neuronsProf. A. HierlemannPhysical Electronics LabETH Zurich

2003 - 2010+Virtual Neural Tissue (VNT)Prof. Rolf PfeiferArtificial Intelligence LaboratoryUniversity of Zurich

2004 - 2005Neurochip recordingsProf. Henry BaltesInst. For Quantum ElectronicsETH Zurich

Science...

State of the art interface as realized byresearchers at MIT around 1995

Perfect virtual reality: Scene from „the Matrix“ by the Wachowskis, 1999

Approaches in Neural Interfacing: Science versus Science Fiction

... vs. Science fiction

Vision today – thinking with the power of 1000 brains

From fishbrain to Hololens: Boosting Intelligence

Malcolm Maciver, Computational Neuroscientist, Northwestern University, 03/2017

Agenda

II. Neuroscience Crash Course: Principles- Understanding the difference between a brain and a computer

Study of the Structure of a Wing. Pen and ink "Dissect the bat(…) and on this model arrange the machine”. Leonardo daVinci (1452-1519), "Codice sul volo degli uccelli", 1505, Originalat the Biblioteca Reale, Turin, Italy.

On Artificial Brains and Bird Wings: An Analogy

Many sought to realize their dream to flyby mimicking birds. Man-craftedmachines that much resembled artificialbirds were envisaged, and in spite ofnumberless attempts pursued overcenturies the mystery of flight could notbe unravelled.

How could we steer a robot with human-like bodydynamics?Can we give life to Cronos?

Example: Cronos humanoid robot in Shanghai

Example: Roboy 2013

Neuroscience Crash Course: Foundations

Neural Tissue

Neurites

Growth Cones

Synchronizity

Neural Networks

Neural Plasticity

Intelligence

Neural TissueNeurons are embedded in a chemical environment, where theinterplay between morphology, the environment and physicsdetermine circuitry and connectivity and thence neural dynamics.

Neurites

Growth Cones

Synchronizity

Neural Networks

Neural Plasticity

Intelligence

Some Neuroscientific foundations

human brain

Some Neuroscientific foundations

How does the brain structure emerge?

the neuron

a single neuron: branching astrocyte

the neuron

a single neuron: branching astrocyte (schematic)

the neuron

a single neuron: branching astrocyte (schematic)

the synapse

human braina single neuron: branching astrocytea single neuron: branching astrocyte (schematic)transmitter release at the synaptic cleft

By the way – why are there gyri and sulci?

human brain

Btw: How does the shape of the brain evolve?

human brain

Brain corals

Walnut

Btw: How does the shape of the brain evolve?

How a brain gets ist shape: 02/2016, Harvard, Mahadevan Lab/Harvard SEAS

Neural Tissue

NeuritesThe observation that axons may grow and branch dependent on the dynamics of underlying micro-elements (microtubuli) promotes the implementation of local assembling strategies that are both interesting from a conceptual and computational viewpoint.

Growth Cones

Synchronizity

Neural Networks

Neural Plasticity

Intelligence

Neural Tissue

Neurites

The observation that axons may grow and branch dependent on the dynamics of underlying micro-elements (microtubuli) promotes the implementation of local assembling strategies that are both interesting from a conceptual and computational viewpoint.

Dendrites + Axons = Neurites

Neural Tissue

Neurites

Growth ConesThe scanning of the local environment for chemical gradients performed by the growth cone occurs ceaselessly and constitutes an interesting strategy that is widely applied in nature. By varying the sensitivity of surface receptors the growth cone may select distinct pathways embedded within a chemical environment.

Dendrites

Synchronizity

Neural Networks

Neural Plasticity

Intelligence

Diploma Thesis: Growing an Intelligent Artficial Neural Network

Motivation : biological Neurons

The ant analogy: intelligent microtubuli ?

Kalil, K. et al. (2000) "Common Mechanisms Underlying Growth Cone Guidance and Axon Branching", J. Neurobio. 44:145-158.Penrose, R., (2000) "The Large, the Small and the Human Mind", UK, Cambridge University Press

Developmental Strategies: Ant Analogy

Neuron (A) and three different ant raiding patterns of ant-colonies(B-D). Note the branching structures that are a result of eitherneural development and ant raiding patterns (adapted fromCamazine et al., 2001).

- How does a neuron grow? - What are the underlying principles?- How does nervous tissue emerge?- How does it process information?- Where are the roots of Intelligence?

Neural Tissue

Neurites

Growth Cones

SynchronizityTononi and Edelman showed in simulations that fast changes in synaptic efficacy and spontaneous activity may rapidly establish a transient but globally coherent process. They then pointed out the possibility of solving the ‘binding problem’ by neuronal synchronization and thus neuronal coherence (Singer, 1995; Edelman et al, 2000).

Neural Tissue

Neurites

Growth Cones

Synchronizity

Neural NetworksNeural Plasticity

Intelligence

First Artificial Neural Network –Frank Rosenblatt’s Perceptron (1957)

"the embryo of an electronic computer that [the Navy] expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence”

New York Times (1958)

Artificial Neural Networks: Principle

The last 60 years in ANN research…

input layer

hidden layer

output layer

Artificial Neural Network 1957

“The Perceptron”110 tons

+ 60 years of research+ Trillions $ of cash+ 1’000’000’000’000 x

computing power

2017 ?

input layer

hidden layer

output layer

input layer

hidden layer 1

output layer

hidden layer 2

hidden layer X

“The Perceptron”110 tons

“Deep Learning”

The last 60 years in ANN research…

150mio b.c. 1490

19032010 1896

1670Artificial Flight: Developments in Aviation

Artificial Thinking: Development in AI “Deep Learning” (2017)

400 b.c.

Development in Aviation and AI: A comparison

Neural Tissue

Neurites

Growth Cones

Synchronizity

Neural Networks

Neural PlasticityIntelligence

Longterm Potentiation (LTP)

no stimulation (3) weak stimulation (2) tetanic stimulation (1)

Problem:

Relating plasticity to behavior is very difficult: Behavioral studies cannot beperformed while precisely recording cellular responses. Any movement preventsaccurate electrophysiological recordings.

Neurons are integrate and fire units (simplification)

…Neurons are integrate and Fire UnitsBiological foundations of perceptrons…

VNT Neurons: Simple Integrate and Fire Units

Integrate and Fire Units… …underlying STDP

Right Panel: Spike Timing Dependent Plasticity as a universal learning rule? Adapted from L. F. Abbott andSacha B. Nelson (2000), “Synaptic plasticity: taming the beast“, Nature Neuroscience 2000.

no stimulation (3) weak stimulation (2) tetanic stimulation (1)

Problem:

Relating plasticity to behavior is very difficult: Behavioral studies cannot beperformed while precisely recording cellular responses. Any movement preventsaccurate electrophysiological recordings.

Axons 1, 2 and 3 (pre) transmitting signals to a pyramidal neuron (post).

B: Current applied to axons 1, 2 and 3.

-> equivalent post-synaptic spikes.

C: 3 different stimulation protocols applied.

-> equivalent post-synaptic spikes

D: After one hour application of same currentas at the beginning.

-> Potentiated, long-lasting response.

VNT Neurons underlying Spike Timing Dependent Plasticity

VNT Integrate and Fire Units underlying STDP

Adapted from L. F. Abbott and Sacha B. Nelson (2000), “Synapticplasticity: taming the beast“, Nature Neuroscience 2000.

Neural Tissue

Neurites

Growth Cones

Synchronizity

Neural Networks

Neural Plasticity

Intelligence

What is Artificial Intelligence?

HumanoidsArtificial BrainsTerminatorMatrixCyborgsSingularity

AI Winter AI Spring

Artificial Intelligence

MachineLearning

Deep Learning

What is Artificial Intelligence?

Silver, David, et al. "Mastering the game of Go with deep neural networks and tree search" Nature, (2016)

Lee Sedol vs. AlphaGo, 03/2016

Artificial Neural Networks: Principle

….your opinion ?

What is Intelligence?What is Intelligence?

Playing chess

Professor Computer

What is Intelligence?

Playing chess

What is Intelligence?

Playing chess

- In a nutshell: “Intelligence requires a body”

Embodiment: Definition The concept that intelligence involves movement, a large variety of behaviors, adaptivity, skills andmuch more.

When Robots learn how to play soccer

Rufus T. FireflyPeople involved: 2

“i-Robot” ?People involved: ?

1990 today 2050 ?

Roboy 2013People involved: 80+

The new Artificial Intelligence: Embodiment- 25 years of research at the AI Lab in Zürich

20151995 ?2010

The Jaquet-Droz Automata, built anno 1770 inNeuchâtel, Switzerland

AI: How much have we progressed since 1770 ?

Neuroscience Crash Course: You made it !

Neural Tissue

Neurites

Growth Cones

Synchronizity

Neural Networks

Neural Plasticity

Intelligence

Agenda

III. Interfacing with brains: From neurointerface to artificial brains- What’s going on in neuroscience

recording signals in nerve cells

recording neural signals: action potentials

high-precision glass electrode for electro-physiological recording of neural activity.

recording neural signals: action potentials

triggering signals in nerve cells

triggering an action potential by depolarization.

signal encoding in nerve cells

triggering a burst of action potentials

visual input: sensory readingsconverted visual input: spikes

recorded output: neural spikes

converted output: motor commands

In collaboration with Mussa-Ivaldi, S., Alford, S., Sanguineti, V., Reger, B., Fleming. K., Kaufmann, P., 2001Northwestern University, University of Illinois, University of Genova, Swiss Federal Institute of Technology

Pascal Kaufmannkpascal@ifi.unizh.ch

Artificial Intelligence Laboratory, Sect. Neural InterfacingUniversity of Zurich, Switzerland

Neural Interfacing: Connecting Brains to Robots

The lamprey: a brain to chat with

Lamprey Brain

Cat Brain

Neurointerface: Experimental Setup

in vitro preparation converter (Matlab) Khepera

Core of the neuro-robotic interface: Stimulating electrodes and recording electrodes.The former deliver and the later sense electrical stimulation that was modified by thebrain of the lamprey sitting in the recording chamber.

abstracted network for predicting the lamprey‘s behavioral responses.

Neurinterfacing: the artificial behavior

1. The axons of the vestibular system are delivered artificial (visual) signals sensed by the robot’s eyes.

2. Instead of adjusting the body’s position in space, the lamprey controls a robot.

in vitro preparation converter (Matlab) Khepera

Khepera: star-guest from Switzerland

Neurointerface: Theory versus Experiment

computer simulated behavior experimentally recorded behavior

Neurointerface: Theory versus Experiment

positive phototaxis mixed taxis negative phototaxis

computer simulation

experimentally recorded behavior

Working at the interface of computer simulations, neuroscience and robotics to tackle the phenomenon of intelligent behavior.

Growing Artificial Neural Tissue: What brings the Future ?

How about today?

There is still a long way to go ...

...and where does that all lead to?

Employee 2.0 – tapping corporate intelligence

Employee 3.0 – incorporating technology

Employee 4.0 – a lot of leisure time for humans

Thank you !

«Neuroscience Crash Course - Tackling the Brain Code» · 2017. 3. 14. · Neural Networks. Neural...

Documents