43 |
44 | This mock-up mimics the look of the in-progress article to inform a design
45 | that embeds the demo into the article. The relevant assets just need to be
46 | migrated into the final article.
47 |
48 |
49 |
We’ve developed Glow, a new type of generative model which uses
50 | invertible 1x1 convolutions to create rich, synthetic models of data,
51 | automatically discovering features we can manipulate. The model extends
52 | previous work on reversible generative models, simplifying the
53 | architecture and leading to substantially better results. We’re releasing
54 | code for the model and an online visualization tool so people can explore
55 | and build on these results.
56 |
57 |
Read Paper
58 |
View Code
59 |
60 |
61 |
62 |
63 |
64 |
65 |
66 |
67 |
Motivation
68 |
Generative modeling is about observing data, like a set of pictures of
69 | faces, then learning a model of how this data was generated. Learning to
70 | approximate the data-generating process requires learning all structure
71 | present in the data, and successful models should be able to synthesize
72 | outputs that look similar to the data. Accurate generative models have
73 | broad applications, including speech synthesis, text analysis and
74 | synthesis, semi-supervised learning and model-based control. The technique
75 | we propose can be applied to those problems as well.
76 |