https://github.com/zimmerrol/spiking-bayesian-networks
Implementation of the paper Homeostatic plasticity in Bayesian spiking networks as Expectation Maximization with posterior constraints by Habenschuss et al.
https://github.com/zimmerrol/spiking-bayesian-networks
bayesian-inference bayesian-spiking-neural-network neural-network nips spiking-neural-networks
Last synced: about 1 year ago
JSON representation
Implementation of the paper Homeostatic plasticity in Bayesian spiking networks as Expectation Maximization with posterior constraints by Habenschuss et al.
- Host: GitHub
- URL: https://github.com/zimmerrol/spiking-bayesian-networks
- Owner: zimmerrol
- License: mit
- Created: 2019-02-18T13:33:42.000Z (over 7 years ago)
- Default Branch: master
- Last Pushed: 2019-04-21T19:58:55.000Z (about 7 years ago)
- Last Synced: 2025-02-26T12:38:01.635Z (over 1 year ago)
- Topics: bayesian-inference, bayesian-spiking-neural-network, neural-network, nips, spiking-neural-networks
- Language: Python
- Homepage:
- Size: 16.6 MB
- Stars: 6
- Watchers: 3
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
Implementation of the paper [Homeostatic plasticity in Bayesian spiking networks as Expectation Maximization with posterior constraints](https://papers.nips.cc/paper/4593-homeostatic-plasticity-in-bayesian-spiking-networks-as-expectation-maximization-with-posterior-constraints) by Habenschuss et al. This paper gives learning rules for a spiking neural network just based on Bayesian reasoning; therefore, the method can be used for unsupervised training of networks.
Contains code to runs different experiments on the proposed model and also on a model that is based not on a Binomial but Gaussian input distribution.
The code was written and the experiments conducted during a one week lasting seminar at the Max-Planck Institute for Dynamics and Self-Organization in 2019.
# Key insights
* `eta_b` has to be sufficiently large, otherwise homeostasis is not strong enough to keep `r` similar for all output neurons
* Even though the paper claims that a factor of 10 between the learning rates is sufficient, we find out that `A_k(V)` contributes exponentially while `b_k` contributes only linearly. Therefore, a factor of ten between `eta_V` and `eta_b` is not always optimal.
* Too few neurons for causes lead to learning of superposition states
* Network can reconstruct images is was not trained on
When images of digits between zero and five with the same ratio are shown to a network with 12 output neurons, for each class two neurons that are class-receptive arise. The neurons slowly learn to react to one of the input types.
