A key challenge for neural modeling is to understand how a continuous stream of multi-modal input from a rapidly changing environment can be processed.
Abstarct
A key challenge for neural modeling is to understand how a continuous stream of multi-modal input from a rapidly changing environment can be processed. This projects deals with a computational model for real-time computing on time-varying inputs, to answer this challenge. This Model is called LSM (Liquid State Machine).
Background On LSM
LSM (Liquid State Machine) is a computational model, based on principles of dynamical systems together with statistical learning theory. In contrast to FSM, the liquid state consists of analog values that may change over time. It is not constructed for a specific task.
Project’s Goals
Examining LSM and network properties on the MEA and on the simulator:
1. Separation property –
Examining whether this basic and necessary property exists in the biological network
2. Network’s parameters –
Finding the network’s best parameters for its best performance
3. Network’s responsiveness –
Examining the selective adaptation of the network to frequency
4. Network with feedback –
Investigating the feedback of the network
Results and Conclusions
The Separation Property (SP) –
This property is a basic and necessary property of a real-time computing on perturbations.
SP address the amount of separation between the trajectories of internal states of the system that are caused by two different input stream.
The study was done by giving known inputs to the neural circuit, recording its response to each input and analyze.
Conclusion:
Our results showed a good separation property in our neural circuit.
Network’s parameters-
Our study here concentrated on finding the optimal paramters of the network simulated in the simulator.
Our goal was to see whether these parameters resembles the reality (MEA)
Majors Conclusions:
-Performance increases with larger sizes
-There is only one optimal connection length. In this length, the performance increases with the size
-Right combinations of length and connectivity can give the optimal performance for a certain network
Responsivness –
A key property of neural systems is their ability to adapt selectively to stimuli invading the network.
Responsiveness is defined as the number of spikes detected in the network within 150 msec after a stimulus.
Our focus was on selective adaptation to the frequency of the stimuli
Conclusion:
Simulator’s results didn’t reflect reality (MEA).
Hence we concluded that a new simulation model is needed.
Network With Feedback
Dopamine is a neurotransmitter, manufactured in the brain. It randomly changes the synapses’ strength (the connections between the neurons), and thus influences the network’s performance.
In our study we examined two kinds of correlations:
1. NextStrength=f(CurrentStrength,error)
2. NextStrength=f(error)
Conclusion:
Our results showed a better convergence while basing on the current strength as well as the error.
Acknowledgment
We would like to thank the following people:
Karina Odinaev and Igal Raichelgauz who gave us patient and thorough guidance throughout the project.
Johanan Erez and all VISL lab staff in the Faculty of Electrical Engineering at the Technion.