• Auto-Encoder is an unsupervised learning for data reconstruction by encoding-decoding.
Specifically, its bottleneck network forces a compressed information representation of an input.
• Dimensionality reduction: compressing an input into a lower dimensional vector (latent vector)
and then reconstructing the input from the latent vector as an output.
• Data-specific: meaningfully compressing data (1)with strong correlation between input features,
and/or (2)similar to what they have been trained on.
Weak correlation between input features results in poor performance.
• Lossy: the output is of lower quality than the input due to the bottleneck network design.