| [3] Ability of Models to Predict 0 Users' Behavior | [4] The graph to the left illustrates how well a set of artificial neural networks (ANNs) do at modeling an aspect of human cognition: aesthetic preference. Right now, 0 users are taking the experiment; 0 users have finished the experiment. For each user, 30 ANNs have tried to model the behavior of him or her by observing which image they choose from among a pair of images. If the ANNs failed to model the user's behavior, then their predictions about the user's choices should be no better than 50%. This is equivalent to a human predicting the outcome of a flipped coin: they would be right half of the time.
The red bars on the left show ANNs that simply make random predictions about the users: as expected, the ANNs' prediction rates tend to hover around 50%. The green bars show the prediction rates of all of the ANNs that were trained while the user made their choices. If they learned something, they should be able to predict much better than 50%. If not, the green bars should correspond more or less to the red bars. If the center of mass of the green graph is lower than the red graph, then the ANNs are generally successful. |
