DND21 Benchmarking

Denoising is prototyped and tested in jAER's NoiseTesterFilter.

A denoising filter subclasses AbstractNoiseFilter. By adding your filter to the list of filters in NoiseTesterFilter, it can be selected.

NoiseTesterFilter lets you choose the noise source (synthetic or real recorded noise), and how to display the testing results.

jAER installation

See jaerproject.org for jAER fork and run instructions.

Video guide to using NoiseTesterFilter

Evaluation by using ROC

Algorithms

To benchmark any of the algorithms you ca select it in NoiseTesterFitler

Denoising algorithms

Source code links

NoiseTesterFilter.java

Obtaining a ROC curve

NoiseTesterFilter works with our external python scripts in our code folder in capture_roc_values subfolder to collect and plot the ROC TPR and FPR data.

readuspmsg.py controls jAER to sweep the filter parameters. Filters implement a RemoteControlled interface to receive the messages.

extractCSV2Npy.py extracts and calculates the average TPR/FPR and other metrics from the specific fractions of the stored csv files derived from parameter sweeping, and stored them into npy files for later plot usage.

readAndPlotROCAll.py plots the ROC curve from the collected data.

Using the MLPF (Multilayer perceptron noise filter)

Add MLPNoiseFilter to list of jAER filters.
Download one of the trained models; we recommend one of the models that uses age and polarity input channels, in particular 2xMSEO1H20_linear_7.pb (a+p x 20 x 1 unit).
Expose Controls for MLPNoiseFilter by hitting the Controls button on it.
Hit the LoadNetwork button to select the .pb network file that you downloaded.
Now you can control the time window and threshold for MLPF so see the denoising effect.
MLPF is also one of the denoising filters you can compare in NoiseTesterFilter.

If you have python and matplotlib on your PATH, you can plot classification histograms directly from MLPNoiseFilter

You can also visualize the TI patch inputs to MLPNoiseFilter. Use the checkboxes to show only signal or noise classification results.

Page updated

Google Sites

Report abuse