Image reconstruction and raster plots in eth_mnist.py #581
Replies: 2 comments 2 replies
-
The image reconstructs the MNIST from the Poisson conversion, as the the sum of spikes for every 28x28 pixel occuring during the model running time, that is passed to the input of the model: Note: you can try to lower the running time param to look at coarser reconstructions of the original MNIST images, it will also slightly affect the model performance: The bottom graph ('X') shows the 784 (28x28) Poisson converted pixels, during the 250 time steps. The denser the dots in time, the higher the brightness of the pixel. The 'Ae' graph shows which excitatory output neurons reacts the most to the input Poisson stream. These are the neurons used to predict the actual MNIST class, through the supervised
and here is the class prediction, estimated by the assignments of the output neurons activity:
( The 'Ai' shows the inhibitory neurons spikes. These spikes prevent other excitatory output neurons from spiking, and are part of the ETH model design (from models.py):
|
Beta Was this translation helpful? Give feedback.
-
Yes
All the details is in the accompany papers, please read Diehl and Cook paper or Unsupervised learning of digit recognition using spike-timing-dependent plasticity and for more details Lattice map spiking neural networks (LM-SNNs) for clustering and classifying image data or arXiv. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Hi, so just looking at the samples below I have a few questions...
Referring to image reconstruction, what is the point of the image reconstruction? Is this the Poisson reconstructed image before it is passed through the network or is this after it has been passed through the network?
The other question is, how do the spikes in the raster plot represent the digits? The spikes are very sparse, and the learning recognition is not easily conceptualised from the raster plot of spikes, due to so little information.
Any help would be greatly appreciated. Thank you.
Beta Was this translation helpful? Give feedback.
All reactions