On Memorization in Probabilistic Deep Generative Models
By Gerrit J. J. van den Burg, Christopher K. I. Williams. 2021
90 citations recorded by the source corpus.
Recent advances in deep generative models have led to impressive results in a variety of application domains. Motivated by the possibility that deep learning models might memorize part of the input data, there have been increased efforts to understand how memorization arises. In this work, we extend a recently proposed measure of memorization for supervised learning (Feldman, 2019) to the unsupervised density estimation problem and adapt it to be more computationally efficient. Next, we present a study that demonstrates how memorization can occur in probabilistic deep generative models such as variational autoencoders. This reveals that the form of memorization to which these models are susceptible differs fundamentally from mode collapse and overfitting. Furthermore, we show that the proposed memorization score measures a phenomenon that is not captured by commonly-used nearest neighbor tests. Finally, we discuss several strategies that can be used to limit memorization in practice. Our work thus provides a framework for understanding problematic memorization in probabilistic generative models.
Subfields: Computer Vision and Pattern Recognition, Generative Models, Safety, Robustness & Interpretability
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Computer Vision and Pattern Recognition
- Florian Tramèr
- Nicholas Carlini
- Reza Shokri
- Ali Borji
- Jamie Hayes
- Eric Wallace
- Colin Raffel
- Luc Van Gool
In Generative Models
- Reza Shokri
- Nikhil Kandpal
- Eric Wallace
- Borja Balle
- Colin Raffel
- Florian Tramèr
- Nicholas Carlini
- Ali Borji
In Safety, Robustness & Interpretability
- Florian Tramèr
- Nicholas Carlini
- Eric Wallace
- Reza Shokri
- Jamie Hayes
- Matthew Jagielski
- Daphne Ippolito
- Milad Nasr
Citations in the corpus
- Deduplicating Training Data Mitigates Privacy Risks in Language Models (2022, methodology): “Models We focus on Transformerbased (Vaswani et al., 2017) language models that range in scale from millions to billions of parameters. Specifically, we use the 117M and 345M parameter models from the Mistral project(1) and the 1.5B parameter forward language model from West et al. (2021), all of wh”
- Enhanced Membership Inference Attacks against Machine Learning Models (2021, background): “, the leave-one-out setting studied in Attack L and memorization literature [10, 35]).”
- Extracting Training Data from Diffusion Models (2023, background): “For instance, numerous metrics exist to measure similarity with the training data [32, 3], the extent of mode collapse [61, 15], and the impact of individual training samples [4, 75].”
- Membership Inference Attacks against Generative Models with Probabilistic Fluctuation (unknown): “In generative models with mem-orization, member records tend to strike higher generative probabilities than neighbor records [ van den Burg and Williams, 2021 ] , i.e., it usually corresponds to a local maximum.”
- Pros and Cons of GAN Evaluation Measures: New Developments (2021, methodology): “Motivated by works that study memorization in supervised learning works, van den Burg and Williams (2021) proposed new methods to understand and quantify memorization in generative models.”