Stöckel, Andreas: Design space exploration of associative memories using spiking neurons with respect to neuromorphic hardware implementations. 2016
Inhalt
- Contents
- List of Figures
- List of Tables
- List of Algorithms
- Introduction
- Motivation and goals
- Neuromorphic hardware systems in the Human Brain Project
- Willshaw associative memory as a spiking neural network
- Associative memories as hardware benchmark
- Goals of this thesis
- Structure
- Notational conventions
- Background and Related Work
- History of artificial neural network models
- First generation: binary McCulloch-Pitts cells
- Second generation: firing-rate coded neural networks
- Third generation: spiking neural networks
- Biophysical neuron model
- Simplified neuron and synapse models
- Neuron model base equation
- Synapse models
- Excitatory and inhibitory synapses
- Linear integrate-and-fire neuron model
- Non-linear integrate-and-fire models
- Two-dimensional Hodgkin-Huxley approximations: the AdEx model
- Neuromorphic hardware
- The Willshaw associative memory model (BiNAM)
- Artificial associative memory models
- Formal description of the Willshaw model
- Choice of the threshold
- Storage capacity and sparsity
- Neural network implementation
- Impact of noise
- Summary and outlook
- Spiking Associative Memory Architecture and Testing
- Neural network topology and data encoding
- Input-/output spike sequences
- Data encoding and input noise parametrisation
- Neuron populations
- Required neuron behaviour
- Memory evaluation measures
- Data generation
- Dataset parametrisation
- Expected behaviour in reaction to uncorrelated random data
- Random data generation algorithm
- Balanced data
- Balanced data generation algorithm
- Conclusion
- Neuron Parameter Evaluation and Optimisation
- Design space exploration
- On the terms "design space" and "exploration"
- Full network evaluation
- Single neuron evaluation
- Parameter constraints and intra-dependencies
- Single neuron simulation
- Neuron simulation loop
- Numerical integration of the AdEx model
- Differential equation integrators
- Integrator benchmark
- Approach 1: spike train
- Approach 2: single group, single output spike
- Approach 3: single group, multiple output spikes
- General idea
- Fractional spike count
- Minimal apical voltage difference
- Minimal membrane potential perturbation
- Neuron evaluation software framework
- Evaluation method comparison
- Conclusion
- Full Network Simulation Experiments
- Methodology and software architecture
- Neuron parameter evaluation
- System parameter sweeps
- Conclusion
- Conclusion and Outlook
- Summary
- Future work
- Large scale simulations and benchmarking
- Neglected design space parameters
- Extensions of the BiNAM network
- Neuron evaluation and parameter optimisation
- Fractional spike count measure
- Conclusion
- Code Examples
- Tables
- Single Neuron Evaluation Comparison
- Acronyms
- Symbols
- Bibliography