A new study introduces a method for estimating how much a model knows about a specific datapoint, formally separating unintended memorization from generalization. The authors measure the capacity of GPT-style models at approximately 3.6 bits per parameter and observe that models memorize until this capacity fills, triggering "grokking" as they begin to generalize.
- Unintended memorization is defined as information about a specific dataset, while generalization is information about the true data-generation process.
- Total memorization provides an estimate of model capacity, measured at roughly 3.6 bits per parameter for GPT-style models.
- Researchers trained hundreds of transformer language models ranging from 500K to 1.5B parameters on datasets of increasing size.
- The study produces scaling laws relating model capacity and data size to membership inference.
This work addresses the difficulty of disentangling memorization from generalization in prior studies, offering a clearer framework for understanding model capacity and training dynamics.