Selection for Reinforcement-Free Learning Ability as an Organizing Factor in the Evolution of Cognition
Figure 6
Explanation of neural connectivity plots used in Figures 7 and 8. (a) This plot shows the weights of all incoming and outgoing connections of a single hidden neuron. The bottom part shows weights of incoming connections (from input neurons), arranged after the positions in the field of view that the input neuron gets its input from (e.g., we see that this example neuron has a negative connection to the input neuron that perceives the state of the cell directly to the left of the organism and a positive connection to the input neuron that perceives the state of the cell two steps ahead of the organism). These weights are innately fixed in both models. In the main model, the outgoing connections (to output neurons, top part of plot) are subjected to learning. To plot these, we ran the individual with every possible output-action assignment (24 assignments in total), repeating each assignment 25 times. The connection weights shown (numbered rows) are averages over the 25 lifetimes per assignment, and a global average over all assignments. For convenience, we arrange the weights of the outgoing connections not by output but by action of the output neuron the connection projects to (e.g., the first column shows strength of connection to the output neuron assigned to the “step” action, regardless of which output neuron that is under a given output-action assignment). In the baseline model the weights of all connections are constant over the lifetime and only a single output-action assignment exists, so in plots of baseline model networks there is only a genetic weight to show for each outgoing connection. The rightmost column of the top part shows the fitness of the individual containing the neuron. In case of the main model, fitness is shown as average per assignment and as global average, like the connection weights. For the baseline model, we show a global average over the same number of lifetime runs used for the global averages of the main model, that is, . Note that fitness columns are identical for all neurons from a single individual. (b) Flipping the sign on all incoming and outgoing connections yields a functionally identical connection pattern. This symmetry was taken into account in further analysis. (c) Colour scales for connection weight and fitness value.
