KTH ROYAL INSTITUTE OF TECHNOLOGY
Introduction to Neuroinformatics Importance of Modelling and Simulations
Pawel Herman
Department of Computational Science and Technology, School of Computer Science and Communication KTH Royal Institute of Technology, Sweden
Neuroscience course, May 17 Why is it important to study brains?
• Quest for knowledge
• Computational inspiration
• Brain disorders and diseases
http://www.modernmedicalguide.com/alzheimers-disease/ Marja-Leena Line, INCF
2 Where are we today?
• collecting more data, building more advanced models, performing more complex analyses
3 Where are we today?
• collecting more data, building more advanced models, performing more complex analyses • BUT: – studying individual components of neural systems with little integration/generalisation effort – focused on a limited set of spatio-temporal scales – working on disconnected data sets using different tools, procedures, protocols – disappointing reproducibility of experimental work (much improved for simulations) – poor understanding of neural mechanisms (computational primitives) at any level of the organisation -> richer computational mindset needed
4 Where are we today?
Progress is not satisfactory and the needs are immense, so are the implications of the advancement of brain science
What could we do to advance the brain science?
5 Where are we today?
Progress is not satisfactory and the needs are immense, so are the implications of the advancement of brain science
What could we do to advance the brain science?
Combine efforts, collaborate interdisciplinarily, organise and share data, integrate biological evidence, build multi-scale models etc.
6 What is neuroinformatics?
NEUROSCIENCE
NEUROINFORMATICS COMPUTER SCIENCE INFORMATION NEUROMEDICINE TECHNOLOGY (IT)
Neuroinformatics is the mean to connect neuroscience, medical science, information technology and computer science.
7 What is neuroinformatics?
NEUROSCIENCE eSCIENCE
NEUROSCIENCE
NEUROINFORMATICS COMPUTER SCIENCE INFORMATION NEUROMEDICINE TECHNOLOGY (IT)
Neuroinformatics is the mean to connect neuroscience, medical science, information technology and computer science.
8 Fundamental goals of neuroinformatics
ICT-based brain research – aims and implications
• organise and integrate neuroscience data
• accelerate our quest for understanding the brain
• support neuromedicine, understanding brain diseases
• develop brain-inspired future computing technologies, brain-like intelligent systems
• promote education, deliver multi-disciplinary training.
9
Fundamental goals of neuroinformatics
ICT-based brain research – fundamental goals
• organise and integrate neuroscience data
10 Fundamental goals of neuroinformatics
ICT-based brain research – fundamental goals
• organise and integrate neuroscience data
“Data tsunami”
“We are drowning in information but starved for knowledge” John Naisbitt
11 Data Age ̶ Multiomic Neuroscience Data
From sub-cellular resolution to whole brain resolution Sean Hill, INCF
12 Fundamental goals of neuroinformatics
ICT-based brain research – fundamental goals
• organise and integrate multi-level data
gather existing data across scales and levels, identify missing data
Sean Hill, INCF
13 Fundamental goals of neuroinformatics
ICT-based brain research – fundamental goals
• organise and integrate multi-level data
gather existing data across scales and levels, identify missing data from genes to behaviour
14 Integrative and data-oriented neuroinformatics
NEUROINFORMATICS
Sean Hill, INCF
15 Focus for neuroinformatics – data, theory
organise and integrate multi-level data
gather existing data at develop tools for build models, simulate, multiple levels, storing, visualising develop theories identify missing data and sharing information build databases, devise new brain atlases approaches to data analysis
13 Focus for neuroinformatics – data, theory
organise and integrate multi-level data
gather existing data at develop tools for build models, simulate, multiple levels, storing, visualising develop theories identify missing data and sharing information build databases, devise new brain atlases approaches to data analysis
13 Computational neuroscience and neuroinformatics
Marja-Leena Linne, INCF
18 Modelling – towards integrative neuroscience
What is the purpose of computational modelling?
adapted by A Kumar
19 Modelling – towards integrative neuroscience
What is the purpose of computational modelling?
• to integrate available data and build theories • to describe & understand the underlying mechanisms • to reveal causal relationships • to generate insights and predictions for experimental neuroscience, etc.
adapted by A Kumar
20 Modelling in neuroscience
• What is a model?
Mathematical model is a description of a system using mathematical concepts - rules, mainly in terms of formulae, e.g.
du τ m = −u(t) + R I(t) dt subthreshold activity in the integrate-and-fire (IF) model
14 An example of a single neuron model – HH formalism
• Various levels of mathematical description – • zooming in (on details) vs. zooming out (abstraction)
15 An example of a single neuron model – rate unit
• Various levels of mathematical description – • zooming in (on details) vs. zooming out (abstraction)
y y = φ(Σ w x ) φ i i
16 An example of a single neuron model – rate unit
• Various levels of mathematical description – • zooming in (on details) vs. zooming out (abstraction)
y y = φ(Σ w x ) φ i i
y
time time
16 Modelling strategy – bridging levels
EMERGING PHENOMENA, ”top-down” HIGHER-LEVEL FUNCTION / THEORY, GLOBAL / DYNAMICS FUNCTIONAL PRINCIPLES
NEURAL DETAIL, WEALTH OF NEURAL IMPLEMENTATION BIOLOGICAL DATA (DYNAMICS, ARCHITECTURE) Gerstner et al., Science 2012 ”bottom-up”
17 Modelling strategy – bridging levels
EMERGING PHENOMENA, ”top-down” HIGHER-LEVEL FUNCTION / THEORY, GLOBAL / DYNAMICS FUNCTIONAL PRINCIPLES
Synthetic LFP 60 50 40 30 20 Frequency (Hz) Frequency 10 NEURAL DETAIL, WEALTH OF NEURAL IMPLEMENTATION BIOLOGICAL DATA (DYNAMICS, ARCHITECTURE)
2 3 4 5 6 7 8 ”bottom-up” Time (seconds)
18 Modelling strategy – bridging levels
EMERGING PHENOMENA, ”top-down” HIGHER-LEVEL FUNCTION / THEORY, GLOBAL / DYNAMICS FUNCTIONAL PRINCIPLES
RESEARCH QUESTION SUITABLE STRATEGY (DATA, CONSTRAINTS) AND LEVEL OF DETAIL
NEURAL DETAIL, WEALTH OF NEURAL IMPLEMENTATION BIOLOGICAL DATA (DYNAMICS, ARCHITECTURE) ”bottom-up”
19 Computational modelling approaches
How can we go about modelling? David Marr’s theoretical approach
Phenomenological models – Statistical models – mathematical description of description of a system in phenomena without handling terms of random variables constituent parts and their distributions (they Mechanistic models – can be mechanistic or description of a system in phenomenological) terms of its constituent parts
28 Computational mindset of David Marr
Three levels of description:
1. Computational level – what does the system do?
• What logic defines the nature of resulting mental 1945-1980 representations of incoming stimuli?
2. Algorithmic level – how does the system do it? • What processes are involved in building mental representations? How is input translate to output?
3. Implementation level – how is the system physically realised – implemented? • What is the neural hardware – substrate?
24 Typical modelling workflow
1. Defining the model: the research question and the research hypothesis determine the type of model, the model components, and the approach to solving the model. 2. Parameter fitting: complex high-dimensional models (biophysical) often have a huge parameter space that cannot be fully explored, instead parameters are fitted from the data, available data influences construction of the model. 3. Simulation: model is implemented in the suitable simulator(s), the obtained simulation results are analyzed and visualized. 4. Validation: the model is confronted with more experimental data, the model behaviour should correspond to the modelled biological system (at least qualitatively). 5. Prediction: good models have predictive power, when additionally perturbed they can show the behaviour of the system under the new conditions.
30 My general modelling philosophy
abstract and conceptual functional models
Model/simulate functional aspects cognitive phenomena, behavioural Develop or build effects on a theory constrain translate
Implement neural anatomy, substrate signal recordings detailed models with neural dynamics
31 My general modelling philosophy
Model/simulate functional aspects Recurrent associative memory Develop or build (Hopfield, 1982) on a theory
Implement neural substrate
Cortical attractor networks
32 Cortical attractor memory model example
Cortical column Hopfield network
Cortical attractor model
Local RSNP
Distant pyramidal
Local basket cell
Local pyramidal Hypercolumn with columns
20 From abstract to biologically detailed implementation
• individual neurons • neural populations • cortical columns mapping (Mountcastle et al., 1955) to biology
Hopfield recurrent neural network
(the concept of a cell assembly, (horizontal connections in the cortical Hebb’s association) layer 2/3 implementing recurrency)
34 Biologically detailed cortical models
HYPERCOLUMN
MINICOLUMN
Modular structure with hypercolumns consisting of minicolumnar units (distributed patterns)
35 Biologically detailed cortical models
Attractor networks HYPERCOLUMN
MINICOLUMN ~1.5 mm
36 Cortical attractor memory model
Hopfield network
Cortical mapping attractor model to biology
Cortical patch
20 Cortical memory function
completion
0 1 2 3 4 5 6 7 seconds
bistability, competition Lundqvist, Herman et al. (2011) J Cogn Neurosci Herman, Lundqvist et al. (2013) Brain Research
21 Oscillatory dynamics in the model
Mesoscopic scales
22 Oscillatory phenomena
Synthetic LFP 60 50 40 30 20 Frequency (Hz) Frequency 10
2 3 4 5 6 7 8 Time (seconds)
Herman, Lundqvist et al. (2013) Brain Research Lundqvist, Herman et al. (2011) J Cogn Neurosci
22 Large-scale simulations
Hodgkin-Huxley
TOOLS
41 A holistic computational model of mammalian olfactory system
receptors and olfactory receptor cells
Buck and Axel, 1991
42 A holistic computational model of mammalian olfactory system
olfactory bulb
Benjaminsson, Herman et al.(2012) Buck and Axel, 1991
43 A holistic computational model of mammalian olfactory system
olfactory cortex
Benjaminsson, Herman et al.(2012) Buck and Axel, 1991
44 Abstract model vs spiking detailed model
Benjaminsson, Herman et al.(2012) Kaplan et al., 2014
45 Abstract model of mammalian olfaction
A wide spectrum of results
Benjaminsson, Herman et al.(2012)
46 How do computational models help?
• Integrate (and fit) experimental data • Describe neural systems – provide mechanistic understanding of the neural system • Make predictions about the system behavior in new conditions • Provide new ways to study brain diseases Marja-Leena • Provide principles to develop new technology (brain-like computing, neuromorphic systems, control for robots)
47 Future challenges for computational neuroscience
• Design of biologically realistic models that span over many levels of spatial organization and a wide range of temporal scales • The need for development of multi-scale interoperable simulation tools (e.g. MUSIC) • Further advancement of simulation technology allowing for interactive control with visualisation capabilities • Enforcing tighter links with biology – interactive and iterative process that deeply involves experimental work
48 Trends and future outlook – what do we need for that?
• more theory and simulations
• tighter connections with experimentalists (from genes to behaviour)
• computational power for large-scale massively parallel simulations
23 Trends and future outlook – what do we need for that?
• more theory and simulations
• tighter connections with experimentalists (from genes to behaviour)
• computational power for large-scale massively parallel simulations
• tools for simulations, analysis and visualisation
MOOSE
23 Thank you for attention
QUESTIONS ?
51