Figure 1: Neural net capable of tensor product representation of role-filler binding or unary relation (A) and as arithmetic example (B). A binary relation shown as a tensor product net (C) and as an arithmetic example (D). A ternary relation is chunked to a binary relation R(a,b/c) in E (with symbol vector omitted for simplicity). A circular convolution calculated from the tensor product in Figure 1B is shown in Figure 1F. The circular convolution is computed by adding along the curved lines and is:
[0.50 0.71 0.50] * [-0.5 0.71 -0.50] = [-0.25 -0.25 0.00]
The shadings in Figures A and C are to make the spatial layout clear, and do not represent levels of activation.