Data driven model optimization [autosaved]
1 like121 views
Russell Jarvis is developing a general purpose optimizer called NeuronUnit to fit abstract neural models to the firing dynamics of specific biological neurons. As a proof of concept, he is using NeuronUnit to fit the Izhikevich model to a murine layer 5 neocortex pyramidal neuron. He discusses using virtual electrophysiology experiments in NeuronUnit along with real neuron recordings from the Allen Brain Atlas to derive error functions that guide the optimization of the model parameters to replicate the biological neuron.
![Data driven model optimization [autosaved]](https://image.slidesharecdn.com/datadrivenmodeloptimizationautosaved-180427045633/85/Data-driven-model-optimization-autosaved-2-320.jpg)
![Data driven model optimization [autosaved]](https://image.slidesharecdn.com/datadrivenmodeloptimizationautosaved-180427045633/85/Data-driven-model-optimization-autosaved-3-320.jpg)
![Data driven model optimization [autosaved]](https://image.slidesharecdn.com/datadrivenmodeloptimizationautosaved-180427045633/85/Data-driven-model-optimization-autosaved-4-320.jpg)
![Data driven model optimization [autosaved]](https://image.slidesharecdn.com/datadrivenmodeloptimizationautosaved-180427045633/85/Data-driven-model-optimization-autosaved-23-320.jpg)
![10 constants: a,b,C,c,d,k,vr,v0,vt,vpeak.
Each constant is searched over 10 samples:
vr=[-75., -72.22222222, -69.44444444, -66.66666667, -63.88888889, -61.11111111, -58.33333333, -
55.55555556, -52.77777778, -50. ]
dim(a)*dim(b)*dim(C)*dim(c)*dim(d)*dim(k)*dim(vr)*dim(v0)*dim(vt)*dim(vpeak)
= 10^10
= 10 000 000 000
10billion samples.
Each sample corresponds to a 1-3 second simulation.](https://image.slidesharecdn.com/datadrivenmodeloptimizationautosaved-180427045633/85/Data-driven-model-optimization-autosaved-33-320.jpg)
![Data driven model optimization [autosaved]](https://image.slidesharecdn.com/datadrivenmodeloptimizationautosaved-180427045633/85/Data-driven-model-optimization-autosaved-37-320.jpg)
Data driven model optimization [autosaved]
- 1. Rigorous Neuron Model Optimization using NeuronUnit Russell Jarvis PhD student Interdisciplinary neuroscience Mentor Professor Sharon Crook Mathematics and Statistics Laboratory for Informatics and Computation in Open Neuroscience https://iconlab.asu.edu Arizona State University
- 5. The Goal of My Current Research ● Make a general purpose optimizer that can fit abstract, general models to the firing dynamics of specific biological neurons • As a Proof of Concept: • Fit Izhikivech model to a murine layer 5 neocortex pyramidal neuron ● Improve existing algorithms by making them faster and scalable. http://neurolex.org/wiki/Category:Neocortex_pyramidal_cell_layer_5-6#tab=Basic
- 7. Fit Izhikevich Model to Layer 5 Pyramidal Neuron Description: http://neurolex.org/wiki/Category:Neocortex_pyramidal_cell_layer_5-6#tab=Detail Firing patterns: (in response to normal physiological input) Regular firing and bursting Spontaneous firing patterns: Regular firing and bursting Spontaneous firing rate: medium firing
- 8. Can replicate a lot of complicated biological sceanarios with simpler eq ● In a digital representation of these equations values for the parameters ● a,b,C,c,d,k,vr,v0,vt,vpeak were systematically explored. ● The effects of changing these parameters were explored by running dedicated simulations for some but not all of their obtainable values.
- 9. MCMSC Mini-workshop 2011 Approaches for Model Reproducibility Izhikevich, Dynamical Systems in Neuroscience, 2007.
- 11. Meta Modeling ● Digital Neuronal Modelling has existed since the at most the early 80s (approaching 40 years). ● Digital models are no longer obscure ● Advances in computing, data digitization, model description, machine readability and collaborative coding, has meant the burden of scientific rigour against all digital models has rightfully increased. Hines, M. Efficient computation of branched nerve equations. International Journal of Bio-Medical Computing 15:69-76, 1984. Journal of Mathematical Biology February 1986, Volume 23, Issue 2, pp 137–161 An analysis of a dendritic neuron model with an active membrane site
- 12. NeuronUnit https://github.com/scidash/neuronunit • Peripheral code includes virtual experiments and feature extraction. • Differentiated from others by emphasising scientific rigor of models • Facilitates testing and comparisons between model prediction and experimental observations. • This helped my research as •Experimental observations are the best starting point for good models •This framework was leveraged to guide model fitting with great effect.
- 14. Rheobase Current Injection The rheobase (pA) value depended on behavior resulting from changing the model parameters. In subsequent slides I will show how Vm (mV) was then compared against specific criteria.
- 17. Rheobase
- 18. 12 spikes 4 spikes Rheobase Search Distributed For Speed 116.66pA Sub-firing Sub firing CPU 0 CPU 1 CPU 2 CPU 3 CPU 4 CPU 6 CPU 5 Supra-firing Supra-firing Rheobase Rheobase 50pA 66.66pA 83.33pA 166.66pA 0 spikes 150pA 0 spikes 133.33pA 0 spikes 0 spikes 0 spikes 0 spikes 2 spikes
- 25. Conclusion ● In this talk I have described the concept and the usage of insilico neuronal modelling. ● I have discussed tests contained inside NeuronUnit to fit Izhikivech model to the layer 5 pyramidal neuron. ● I have discussed a way that I have speed up an existing feature extraction algorithm ● And I have discussed optimization generally.
- 26. Acknowledgements: Sharon Crook, Rick Gerkin, Justas Birgiolas, Reuben Haynes, Padraig Gleeson NeuronUnit: a Social Product Local History: Eugene M. Izhikevich 1996 2000 Visiting Professor of Department of Mathematics & Center for Systems Sciences Arizona State University 1998-2000, Faculty Research Associate, Systems Science & Engineering Research Center , Arizona State University.
- 28. Future Directions. Limitations: How far away am I from general purpose optimization. Porting the existing model onto NSG cluster. Running parallel NeuronUnit as a web service using a cluster the NSG as a backend, OSB as a front end. Experiment with Substitution of Ipythonparallel for scoop as would facilitate ipython notebooks.
- 29. 7 Error functions where used to guide the evolution of the GA. Each test is the result of a virtual electrophysiology experiment. Besides the rheobase test for rheobase current which has a simple scalar value, the rest of the neuronunit tests used the mean and standard deviation measures in order to place model predictions inside a normalized error range. RheobaseTest, candidate observation versus prediction RheobaseTest value 130.0 pA, InputResistanceTest mean 120672073.643411 ohm, std 77633160.8333564 ohm TimeConstantTest mean 15.7342424242424 ms, std 7.31162636832495 ms CapacitanceTest mean 1.50584166666667e-10 F, std 1.39683884626343e-10 F RestingPotentialTest mean -68.2481434599156 mV std 6.53234788156637 mV InjectedCurrentAPWidthTest mean 1.20769387755102 ms std 0.534345918375033 ms InjectedCurrentAPAmplitudeTest mean 80.4351020408164 mV std 12.7488030357545 mV InjectedCurrentAPThresholdTest mean -42.7357232704403 mV std 8.04073233409085 mV
- 30. What is Docker? Not an actual VM but similarly providing containment and insulation of the HOST. Recreating of complex build environments is automated in a script which actually . Given the complicated git file merge history depicted previously, its very difficult to keep developers on the same page. If developers are not on the same page, bug replication is difficult.
- 31. Genetic Operations Coordinates in parameter space, are represented as binary strings. New genes are derived from very simple operations on these strings.
- 33. 10 constants: a,b,C,c,d,k,vr,v0,vt,vpeak. Each constant is searched over 10 samples: vr=[-75., -72.22222222, -69.44444444, -66.66666667, -63.88888889, -61.11111111, -58.33333333, - 55.55555556, -52.77777778, -50. ] dim(a)*dim(b)*dim(C)*dim(c)*dim(d)*dim(k)*dim(vr)*dim(v0)*dim(vt)*dim(vpeak) = 10^10 = 10 000 000 000 10billion samples. Each sample corresponds to a 1-3 second simulation.
- 34. Parallel Implementation map Single processorMultiprocessor flat topology Multiprocessor Tree topology
- 36. Solution: Unit Testing Test 1 Test 2 Test 3 Test 4 Overall Model 1 0.1 -0.1 2.0 -0.3 0.6 Model 2 0.7 1.0 -0.1 -0.3 0.5 Model 3 0.3 0.3 -0.4 -0.3 0.3 Model 4 0.0 0.1 -0.1 -0.3 0.1 Model 5 -1.5 2.9 -2.4 -0.2 1.8 Overall 0.5 0.9 1.0 0.3 Model Test = Expe Obse Model Prediction
- 38. MCMSC Mini-workshop 2011 Approaches for Model Reproducibility What properties of models will improve their scientific impact across neuroscience? Reproducibility: easy to rerun and validate simulation result reported in a scientific paper. Accessibility: available to theoretical and experimental neuroscientists in an understandable format Portability: cross-simulator validation and exchange of models and components enabling reuse Transparency: exposure of internal properties and automated validation
- 40. Layer 5
- 41. Progress of the GA genes from conception to convergences on a candidate.