Skip to main content
. 2020 Jun 30;14:653. doi: 10.3389/fnins.2020.00653

Figure 5.

Figure 5

AGD backpropagation chain rule update with BackRes connections. The BackRes block is essentially unrolled n times and the loss is propagated through the unrolled graph to compute the weight updates at each unrolling step n = 1, 2 as shown. Note, the Conv layer weights w of the BackRes block receive two updates with different input and output combinations giving rise to diverse gradients.