In this video we look at some enhancements to the CNN architecture, running multiple convolution operations and then pooling gives a significant pick up in accuracy. We then look at techniques to explain the performance of our neural networks.
A legitimate criticism of Neural networks specifically and AI more broadly is that these techniques are a 'black box'. While many solutions have impressive results it is often difficult to explain why the AI is successful or how it generates its results.
We counter this by building an 'Artificial EEG' (Electroencephalogram) . This shows which artificial neural pathways 'light up' when presented with specific data. We can look at the intensities in our convolutional filters as well as the 'feature maps' formed when we apply these filters to our images.
The intuition here is that by repeatedly applying convolutions and downsampling we will start to generate abstract patterns that are characteristic of specific types of object in our image. Simple stated Cruise Ships will produce very different patterns of pattern activation that Cats.
We build some visualization tools to allow us to see these filters and feature maps.