A theory of continuous generative flow networks

Salem, Lahlou,; Tristan, Deleu,; Lemos, Pablo; Zhang, Dinghuai; Alexandra, Volokhova,; Hernández-García, Álex; Ezzine, Léna Néhale; Bengio, Yoshua; Malkin, Nikolay

doi:10.48550/arxiv.2301.12594

Cited by 3 publications

(6 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Future work will seek to develop a method that is able to both select the most probable causal generative factors and measure how well they correspond to distinct latent dimensions. Recent work on causal representation learning using neural networks [ 31 , 32 ] could be an applicable approach. Our developed two-step pairwise correlation analysis is also data efficient, allowing one to probe a subset of the data instead of the entire dataset.…”

Section: Discussionmentioning

confidence: 99%

In Search of Disentanglement in Tandem Mass Spectrometry Datasets

Abram,

McCloskey

2023

Biomolecules

View full text Add to dashboard Cite

Generative modeling and representation learning of tandem mass spectrometry data aim to learn an interpretable and instrument-agnostic digital representation of metabolites directly from MS/MS spectra. Interpretable and instrument-agnostic digital representations would facilitate comparisons of MS/MS spectra between instrument vendors and enable better and more accurate queries of large MS/MS spectra databases for metabolite identification. In this study, we apply generative modeling and representation learning using variational autoencoders to understand the extent to which tandem mass spectra can be disentangled into their factors of generation (e.g., collision energy, ionization mode, instrument type, etc.) with minimal prior knowledge of the factors. We find that variational autoencoders can disentangle tandem mass spectra data with the proper choice of hyperparameters into meaningful latent representations aligned with known factors of variation. We develop a two-step approach to facilitate the selection of models that are disentangled, which could be applied to other complex and high-dimensional data sets.

show abstract

Section: Discussionmentioning

confidence: 99%

In Search of Disentanglement in Tandem Mass Spectrometry Datasets

Abram,

McCloskey

2023

Biomolecules

View full text Add to dashboard Cite

show abstract

“…Bengio et al (2021b) attempt at laying the foundation of the theory of GFlowNets, they discuss many openings for future applications or explorations of the method. The work of Lahlou et al (2023) using a similar approach to generalize GFlowNets. Finally, Li et al (2023d) made a first attempt at training GFlowNets for continuous state space.…”

Section: Related Workmentioning

confidence: 99%

“…However, none of these works attack cyclic space limitations, in particular, our stability property is new. Furthermore, our framework is somewhat less involved than that of Lahlou et al (2023) in that most of the fundamental work deals with general finite non-negative measures. Extra hypotheses enforcing acyclicity are not used nor even specified.…”

Section: Related Workmentioning

confidence: 99%

“…Worse, state space and transitions may be constrained in such a way that cycles are unavoidable: one may think of a game in which the adversary may force to return to a previous state. Recent work (Li et al 2023d;Lahlou et al 2023) made attempts at the former but the latter is still unaddressed.…”

Section: Introductionmentioning

confidence: 99%

“…Main Contributions & Outline of the Paper 1) We describe in section a property of losses on graphs leading flows to be trapped into loops and define a notion of stability in order to control this behavior which is detrimental to training and increases inference cost. 2) Section describes the elementary measure theoretical generalization of GFN slightly differently from that of Lahlou et al (2023) lifting the built-in acyclicity constraint. The mathematical framework is enriched in particular with the definition of a suitable generalization of cycles.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

A Theory of Non-acyclic Generative Flow Networks

Brunswic,

Li,

et al. 2024

AAAI

View full text Add to dashboard Cite

GFlowNets is a novel flow-based method for learning a stochastic policy to generate objects via a sequence of actions and with probability proportional to a given positive reward. We contribute to relaxing hypotheses limiting the application range of GFlowNets, in particular: acyclicity (or lack thereof). To this end, we extend the theory of GFlowNets on measurable spaces which includes continuous state spaces without cycle restrictions, and provide a generalization of cycles in this generalized context. We show that losses used so far push flows to get stuck into cycles and we define a family of losses solving this issue. Experiments on graphs and continuous tasks validate those principles.

show abstract

In search of disentanglement in tandem mass spectrometry datasets

Abram

McCloskey

2023

Preprint

View full text Add to dashboard Cite

Generative modeling and representation learning of tandem mass spectrometry data aim to learn an interpretable and instrument-agnostic digital representation of metabolites directly from MS/MS spectra. Interpretable and instrument-agnostic digital representations would facilitate comparisons of MS/MS spectra between instrument vendors and enable better and more accurate queries of large MS/MS spectra databases for metabolite identification. In this study, we apply generative modeling and representation learning using variational autoencoders to understand the extent to which tandem mass spectra can be disentangled into its factors of generation (e.g., collision energy, ionization mode, instrument type, etc.) with minimal prior knowledge of the factors. We find that variational autoencoders can disentangle tandem mass spectra data with the proper choice of hyperparameters into meaningful latent representations aligned with known factors of variation. We develop a two-step approach to facilitate the selection of models that are disentangled which could be applied to other complex and high-dimensional data sets.

show abstract

A theory of continuous generative flow networks

Cited by 3 publications

References 0 publications

In Search of Disentanglement in Tandem Mass Spectrometry Datasets

In Search of Disentanglement in Tandem Mass Spectrometry Datasets

A Theory of Non-acyclic Generative Flow Networks

In search of disentanglement in tandem mass spectrometry datasets

Contact Info

Product

Resources

About