Matrix multiplication (matmul) multiplies matrices and serves as the linear part of geometric transformations. A 2 × 2 matrix product can rotate or scale a point in 2D space. section 2.3.5
In a Dense layer, the matrix product contributes to an affine transformation; adding a vector supplies its translation. Repeated affine layers without activations still reduce to one affine transformation. section 2.3.5
Int8 quantization
To approximate a float32 matrix product with int8 arithmetic, scale each input to fit the int8 range, multiply, then unscale the output. Rounding during conversion can cause some accuracy loss. section 18.3.6
x <- op_array(rbind(c(0.1, 0.9), c(1.2, -0.8))) kernel <- op_array(rbind(c(-0.1, -2.2), c(1.1, 0.7)))
The example provides two small matrices used as inputs to the quantized product. section 18.3.6
