[2106.05258] Generative Models as a Data Source for Multiview Representation Learning

Generative models are now capable of producing highly realistic images that look nearly indistinguishable from the data on which they are trained. This raises the question: if we have good enough generative models, do we still need datasets? We investigate this question in the setting of learning general-purpose visual representations from a black-box generative model rather than directly from data. Given an off-the-shelf image generator without any access to its training data, we train representations from the samples output by this generator. We compare several representation learning methods that can be applied to this setting, using the latent space of the generator to generate multiple "views" of the same semantic content. We show that for contrastive methods, this multiview data can naturally be used to identify positive pairs (nearby in latent space) and negative pairs (far apart in latent space). We find that the resulting representations rival those learned directly from real

3 mentions: @ak92501@hillbig@hillbig
Date: 2021/06/11 00:17

Referring Tweets

@hillbig データセットから生成モデルを学習し、生成したデータのみから表現学習する。潜在変数上での近傍を正例としさらに画像上でオーグメンテーションを適用した方が良い表現が得られる。元のデータ上で学習した場合に近い性能が出るが超えず、サンプル数増で精度改善するがサチる t.co/x7urGJcndz
@hillbig Investigation of image representation learning using generated dataset. The best result is obtained when the positive pairs are nearby in latent space, and data augmentation is applied. Achieved close but not better performance than real-data training. t.co/x7urGJcndz
@ak92501 Generative Models as a Data Source for Multiview Representation Learning pdf: t.co/VGQRmL0BS1 abs: t.co/50gFZE92QQ project page: t.co/FfqqQKTWCY t.co/rRo5HrZD3i

Related Entries

Read more Meta-Sim2
0 users, 5 mentions 2020/08/25 23:21
Read more [2010.10505] SDF-SRN: Learning Signed Distance 3D Object Reconstruction from Static Imagesopen searc...
0 users, 6 mentions 2020/10/22 17:22
Read more [2011.13775] Image Generators with Conditionally-Independent Pixel Synthesisopen searchopen navigati...
0 users, 7 mentions 2020/12/01 23:21
Read more NeRV
0 users, 4 mentions 2020/12/09 00:53
Read more LoFTR: Detector-Free Local Feature Matching with Transformers
1 users, 4 mentions 2021/04/04 23:23



安定したサイト運営のためにGitHub sponsorを募集しています。