Autoencoders learn by making mistakes and fixing them. The learning signal is the

Autoencoders learn by making mistakes and fixing them. The learning signal is the

• A. Classification accuracy
• B. Reconstruction error
• C. Cross-entropy loss
• D. Regularization penalty

Answer: (B)

Autoencoders learn by trying to reproduce their input, so the signal guiding learning is how far the reconstruction is from the original data. This reconstruction error is what you minimize during training, using a loss function (like mean squared error for continuous data or cross-entropy for binary-like data). The gradients of this error with respect to the network’s weights tell the model how to adjust so the output better matches the input, which in turn forces the encoder to capture meaningful, compact representations. Regularization terms can shape learning but they don’t measure reconstruction quality themselves, and classification accuracy isn’t relevant to reconstructing inputs. Cross-entropy can be used as a reconstruction loss in some variants, but reconstruction error is the general learning signal driving standard autoencoders.

Autoencoders learn by trying to reproduce their input, so the signal guiding learning is how far the reconstruction is from the original data. This reconstruction error is what you minimize during training, using a loss function (like mean squared error for continuous data or cross-entropy for binary-like data). The gradients of this error with respect to the network’s weights tell the model how to adjust so the output better matches the input, which in turn forces the encoder to capture meaningful, compact representations. Regularization terms can shape learning but they don’t measure reconstruction quality themselves, and classification accuracy isn’t relevant to reconstructing inputs. Cross-entropy can be used as a reconstruction loss in some variants, but reconstruction error is the general learning signal driving standard autoencoders.