Denoising autoencoder with modulated lateral connections learns invariant representations of natural images