Refinement of receptive field properties for various pre-training strategies.

<b>A.</b> Spatio-temporal receptive field of a target neuron in the last 2D convolutional layer of models using a vanilla RNN architecture. Shown are the 10 input frames, each cropped to a square of 27 × 27 pixels, which maximally activate the target neuron, indicated by a cross. <b>B.</b> Receptive field area in pixels for each frame. Note that the randomly initialized and the <i>DirectedProp</i> models are the same as shown in <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1012830#pcbi.1012830.g006" target="_blank">Fig 6</a>. <b>C.</b> Receptive field strength, measured as the 95th percentile of pixel activations, for each frame. (PDF)