What do large networks memorize?

Michal Lukasik, Aditya Krishna Menon, Ankit Singh Rawat, Vaishnavh Nagarajan, Sanjiv Kumar

Keywords: memorization, overparameterization, example difficulty

Abstract:

The success of modern neural models has prompted renewed study of the connection between memorisation and generalisation: such models typically generalise well, despite being able to perfectly fit ("memorise") completely random labels.
To more carefully study this issue, Feldman (2019); Feldman & Zhang (2020) provided a simple metric to quantify the degree of memorisation of a specific training example, and empirically quantified the corresponding memorisation profile of a ResNet model on image classification benchmarks.
While an exciting first glimpse into how real-world models memorise, these studies leave open several questions about memorisation of practical networks.
In particular, how is memorisation affected by increasing model size, and by distilling a large model into a smaller one?
We present a systematic empirical analysis of these questions.
On standard image classification benchmarks, we find that training examples exhibit a diverse set of memorisation trajectories across model sizes, with some samples having increased memorisation under larger models.
Further, we find that distillation tends to inhibit memorisation of the student model, while also improving generalisation.
Finally, we show that computationally tractable measures of memorisation do not capture the properties we identify for memorisation in the sense of Feldman (2019), despite highly correlating to the latter.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning

TL;DR: Increasing model size may increase memorisation of certain training samples, while distillation inhibits memorisation