Garrick Orchard: Spiking Neural Networks for Event-based Vision

[@vishalsingh-ds2zf]

really liked how he summarised many of the problems one faces with event-based cameras and SNNs.Saved a lot of reading for me

[@patrickjdarrow]

The presenter did a great job keeping things condensed yet simple. Props!

[@LammaDrama]

Amazing presentation! Btw, why does the image input to the network at 15:45 have to move?

[@puttatidam.1819]

they are using spiking neural networks for event-based data processing. Such data usually comes from DVS (dynamic visual sensor) that captures real-time events. You have to read about it since it will be quite long to explain. But basically, the sensor only captures changes in the pixel, so if the background is static, it would not generate a spike. This way it saves a lot of computational power since you dont have to process the whole image but only where there are changes happening. DVS data is mainly used for gesture and object recognition. The number you see has already been transformed into a DVS version, thus the movement. IF it's static then it would just a normal static classification and not event-based (dependent on time)

[@LammaDrama]

@@puttatidam.1819 Oh, thanks a lot! Do you know if it's possible to use a static image as an input to a SNN?

[@puttatidam.1819]

@@LammaDrama Of course! In fact, the majority of the work up till today has been on static images rather than neuromorphic datasets. However, you have to convert those images into spikes input, so you need to apply a conversion algorithm (most studies use intensity where higher intensity--> higher spiking rate). But using it that way, they aren't as effective as traditional deep learning networks. People are still trying to optimize it since SNN for computer vision/machine learning is very new. Please look for the paper "Deep learning in Spiking Neural Network". They pretty much lay out the fundamental ideas of SNN. You're welcome:)

[@LammaDrama]

@@puttatidam.1819 amazing! Thanks

[@이인섭-r2x]

wow... so u guys using error estimated tuning like back-propagation? impressive :)

[@skasrafulislam8442]

I love you

[@skasrafulislam8442]

Pablik I wanted