Synaptic Failure Rao Prnpuu, Oliver Hrmson Introduction The Aim: - - - PowerPoint PPT Presentation

Synaptic Failure

Rao Pärnpuu, Oliver Härmson

Introduction

The Aim:

• Understanding the role of synaptic failure (SF) in the

brain Two parts:

• Overview of previous studies, both computational and

biological

• Creating a model and running simulations under

different circumstances

Synaptic failure and literature

• SF reported to be 0.1 … 0.9
• Some energy losing

transformations

• Channel failure
• Quantal failure
• Quantal amplitude fluctuation
• Dendrosomatic summation

Benefit? (a) Energy expenditure (b) Information processing

The Model

• Integrate-and-fire (2 input

neurons, 1 output neuron)

• Different input spike trains:
• a. Totally regular
• b. Poissonian, 100% correlated
• c. Poissonian, 80% correlated
• d. Poissonian, uncorrelated
• e. Poissonian, 100% correlated, independent fail
• Input rate normalization

Figure Input rate normalization Figure SVM classifier decision boundrary

Methods

• Data analysis and implementation: MATLAB
• Classification used SVM classifier:

Training set: millisecond averages of membrane potentials. Training set: 80% and test set:20% of outputs

• Simulated failure rates of 0.0 to 0.9 , with 0.1 steps
• Used two combinations of input neuron weights: 5-5 and

7-3

Aims

(i) To see whether changing the membrane potential dynamics in response to a spike event in the postsynaptic neuron will change the amount of spikes fired. (ii) To see whether different probabilities of synaptic failure will affect the neuron’s ability to distinguish one input from the other, using machine learning. (iii) To see how well the postsynaptic neuron distinguishes its inputs when input spike train types are altered, using machine learning.

Results

Changing membrane potential parameters

What should the membrane potential do after a spike event occurs in the soma?

• Vreset = always -65 mV
• Vrestchanging = -2mV after every spike, constantly changing

Experiment 1

Regular spiketrains, 100% correlated and independent failure events Simulations with normalised input of 10Hz and 30Hz, over 101 seconds

10 Hz input

Figure 1.

Figure 2.

30 Hz input

Experiment 2

Poissonian fully correlated spike trains

• Probability of failure was simulated by

removing spikes from both spiketrains at simultaneous time moments.

Figure 3 10 Hz Input

Figure 4 30 Hz input

Experiment 3

Poissonian 80% correlated spike trains

Figure 5 10 Hz input

Figure 6 30 Hz input

Experiment 4

Poissonian, totally uncorrelated spiketrains Simulations with normalised input of 10Hz and 30Hz, over 101 seconds

Figure 7.

10 Hz input

Figure 8.

30 Hz input

Experiment 5

Poissonian spiketrains with independent failure events Simulations with normalised input of 10Hz and 30Hz, over 101 seconds

Figure 9.

10 Hz input

Figure 10.

30 Hz input

Conclusions

• In the model used, change in resting membrane potential did not influence
• utput rates
• In some cases, failure rates influenced the accuracy of classification.
• In some cases higher failure rates increased classification accuracy (the

effect of decorrelation)

Future Directions

• Running longer simulations over several runs for

reduction of classification errors

• Implementing models with higher number of

neurons as input (4,10,20,50,...)

• Creating complex experiments that are more

sensitive to different correlation levels between inputs

• Measuring plasticity in different synapses after

application of synaptic failure

• Trying different classifiers

Thank You!

Discuss...