Take away lessons

Prescribed error sensitivity: A biologically plausible NEF-based supervised learning rule which modifies connection weights between populations of neurons to minimize an external error signal

ANN for computer vision. For image classification, a 64 × 64 × 3 image would have 12,288 features and a 100 × 100 × 3 image 3,000,000 features. A DNN with a single fully connected hidden layer comprises 1,000 neurons and entails more than 3 billion parameters.

Convolution. When we work with images which are discretized entities we can define convolution as a discrete operator in any dimension.

By padding an image with additional pixels layers, image downsizing during convolution is prevented. Padding a 3 × 3 image (p = 1, value = 0) creates a 5 × 5 image. When convolved with a 3 × 3 kernel, the result is a 3 × 3.

In a strided convolution, we jump over positions during con- volution. Convolving a 7 × 7 image with a 3 × 3 kernel and stride = 2 results in a 3 × 3 image.