Convolutional RNN

• $h_t={\sigma }_h(W_{hh}\star h_{t-1}+W_{xh}\star x_t+b_h)$
• $y_t={\sigma }_y(W_{hy}\star h_t+b_y)$
• $\star$ is spatial Conv
• 5D shapes → [samples, timesteps, width, height, channels]
• Very memory intensive
• $x^2+x$