Analysis of Learning a Flow-based Generative Model from Limited Sample Complexity
Authors: Hugo Cui, Florent Krzakala, Eric Vanden-Eijnden, Lenka Zdeborova
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The theoretical predictions for the skip connection strength ˆct and the component mt, qξ t of the weight vector ˆwt are plotted as solid lines in Fig. 1, and display good agreement with numerical simulations, corresponding to training the DAE (9) on the risk (10) using the Pytorch (Paszke et al., 2019) implementation of the Adam optimizer (Kingma & Ba, 2014). |
| Researcher Affiliation | Academia | Hugo Cui Statistical Physics of Computation Laboratory École Polytechnique Fédérale de Lausanne (EPFL) Lausanne, Switzerland Florent Krzakala Information Learning and Physics Laboratory École Polytechnique Fédérale de Lausanne (EPFL) Lausanne, Switzerland Eric Vanden-Eijnden Courant Institute of Mathematical Science New York University (NYU) New York, USA Lenka Zdeborová Statistical Physics of Computation Laboratory École Polytechnique Fédérale de Lausanne (EPFL) Lausanne, Switzerland |
| Pseudocode | No | The paper describes mathematical derivations and processes for the models but does not present any formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code used in the present manuscript is provided in this repository. |
| Open Datasets | No | We consider the case of a target density ρ1 given by a binary isotropic and homoscedastic Gaussian mixture... and ...a finite number n of samples from the target distribution. |
| Dataset Splits | No | The paper defines a 'training set D = {xµ 1, xµ 0}n µ=1' but does not specify any train/validation/test splits, percentages, or absolute sample counts for different phases of model development (training, validation, testing). |
| Hardware Specification | No | The paper mentions running 'numerical simulations in dimension d = 5 104' and training models with 'Pytorch implementation of full-batch Adam' but does not specify any particular GPU models, CPU types, or other hardware components used. |
| Software Dependencies | No | The paper mentions using 'Pytorch (Paszke et al., 2019) implementation of the Adam optimizer (Kingma & Ba, 2014)' but does not provide specific version numbers for PyTorch or other libraries, which are required for reproducible software dependencies. |
| Experiment Setup | Yes | The paper specifies training parameters such as 'learning rate 0.0001 over 4 104 epochs and weight decay λ = 0.1' for some simulations, and 'learning rate 0.01 for 2000 epochs' for others, along with optimizer details (Adam). |