Unsupervised Learning of Compositional Energy Concepts
arXiv:2111.03042
Abstract
Humans are able to rapidly understand scenes by utilizing concepts extracted from prior experience. Such concepts are diverse, and include global scene descriptors, such as the weather or lighting, as well as local scene descriptors, such as the color or size of a particular object. So far, unsupervised discovery of concepts has focused on either modeling the global scene-level or the local object-level factors of variation, but not both. In this work, we propose COMET, which discovers and represents concepts as separate energy functions, enabling us to represent both global concepts as well as objects under a unified framework. COMET discovers energy functions through recomposing the input image, which we find captures independent factors without additional supervision. Sample generation in COMET is formulated as an optimization process on underlying energy functions, enabling us to generate images with permuted and composed concepts. Finally, discovered visual concepts in COMET generalize well, enabling us to compose concepts between separate modalities of images as well as with other concepts discovered by a separate instance of COMET trained on a different dataset. Code and data available at https://energy-based-model.github.io/comet/.
arXiv commentNeurIPS 2021, website and code at https://energy-based-model.github.io/comet/
References in corpus (17)
- Multi-Object Representation Learning with Iterative Variational Inference
- MONet: Unsupervised Scene Decomposition and Representation
- A Theory of Generative ConvNet
- Isolating Sources of Disentanglement in Variational Autoencoders
- GENESIS: Generative Scene Inference and Sampling with Object-Centric Latent Representations
- Relational Neural Expectation Maximization: Unsupervised Discovery of Objects and their Interactions
- Implicit Generation and Generalization in Energy-Based Models
- Weakly-Supervised Disentanglement Without Compromises
- Sequential Attend, Infer, Repeat: Generative Modelling of Moving Objects
- Your Classifier is Secretly an Energy Based Model and You Should Treat it Like One
- Deep Directed Generative Models with Energy-Based Probability Estimation
- Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning
- Investigating Object Compositionality in Generative Adversarial Networks
- Contrastive Learning Inverts the Data Generating Process
- ICE-BeeM: Identifiable Conditional Energy-Based Deep Models Based on Nonlinear ICA
- R-SQAIR: Relational Sequential Attend, Infer, Repeat
- On Linear Identifiability of Learned Representations