Figure 11.

Activation patterns of the hidden units of DNNs can be visualized (a lower layer). Samples of the activation patterns of receptive fields in a lower convolution LSTM layer are presented (R-Layer0/ConvLSTM/Variable_1_3_120_160_float32_29, refer to image data numbers of TensorBoard). A network trained with 500K video frames was used for prediction. Twenty consecutive images were input to the trained networks, and visualization was then performed using TensorBoard. In this layer, there are three channels with a 120 (width) × 160 (height) pixel image. In the cw illusion, horizontal inversion pictures of the single-ring illusion (ccw) presented in Figure 2 were used as input pictures. Note that noisy random patterns are specifically observed in “Non-Illusion”.