A Subspace Network That Determines Its Own Output Dimension

@inproceedings{Plumbley1994ASN,
  title={A Subspace Network That Determines Its Own Output Dimension},
  author={Mark D. Plumbley},
  year={1994}
}
Neural networks which extract the principal components or principal subspace of their input data optimize information extraction from noisy inputs. With noise on both input and output, a rst-order approximation is to extract the principal sub-space of the input, suppressing those input components with variance below a certain threshold value. We describe a 2-stage linear network which uses interneurons connected between output units, and which uses local Hebbian and anti-Hebbian learning to… CONTINUE READING