Loss in Latent Space Models

1

Running the Edward implementation-example of Latent Space Models http://edwardlib.org/tutorials/latent-space-models, I get an increasing “final” model loss when I increase the number of latent dimensions. I expected to see a decrease in loss. Any idea what is going on?

  • Latent dim (K): 3 | Loss: 35638.617

  • Latent dim (K): 6 | Loss: 35677.238

  • Latent dim (K): 30 | Loss: 41935.812

  • Latent dim (K): 60 | Loss: 50123.602