Extra

Question 1

What is the difference between a convolutional network and a fully connected network?

Answer 1

In a convolutional network, units in a layer are connected to a spatially restricted set of units in a lower layer. In a fully connected network, units connect to all units in the lower layer
this restriction is useful because:
-upper units focus on local relationships first, spatially restricted or temporally restricted inputs, then needs lots of layers to find patterns over the whole image
-fully connected networks tend to overfit at each layer, layers find anything, not only spatially restricted patterns
-in fully connected network, all units in a layer are looking at the same inputs, can find any patterns within the first layer of the network

Question 2

How do cells detect spatial and non spatial comparisons?

Answer 2

The same ganglion cell looks at the cones and rods (photoreceptors) to determine spatial comparisons and colour comparisons

Question 3

How do predNets predict the next image?

Answer 3

Network is trained to minimise the error signal which is the difference between the prediction and the actual next input, so units in the recurrent layer can make predictions about what will happen next based on what normally happens next

Question 4

Which cells look at low spatial frequencies and which look at high spatial frequencies?

Answer 4

Low spatial frequencies - parasol cells, see light and dark of whole image
High spatial frequencies - midget cells, see fine details