Autoencoders are neural networks where the outputs are its own inputs. They are split into an encoder part which maps the input \( \boldsymbol{x} \) via a function \( f(\boldsymbol{x},\boldsymbol{W}) \) (this is the encoder part) to a so-called code part (or intermediate part) with the result \( \boldsymbol{h} \)
$$ \boldsymbol{h} = f(\boldsymbol{x},\boldsymbol{W})), $$where \( \boldsymbol{W} \) are the weights to be determined. The decoder parts maps, via its own parameters (weights given by the matrix \( \boldsymbol{V} \) and its own biases) to the final ouput
$$ \tilde{\boldsymbol{x}} = g(\boldsymbol{h},\boldsymbol{V})). $$The goal is to minimize the construction error.