Random Weight Initialization
Strong Inductive Bias of inherent model
Embedding training only models work for simple pattern matching.
SAE only single token features
Is SAE really meaningful or it is the property of sparse text dataset? (that only single-token features are discovered) I suspect the results are more due to the structure of the transformer itself rather than the superposition in token embeddings, since the comparison between random and intact token embeddings showed similar outcomes. This makes me curious about how these findings would generalize to other architectures.
Local KL Volume
This methodology defines a set of KL-neighbors (behaviorally similar region) around the trained model weights and efficiently estimates the probability (=Local KL Volume) that this region occupies under the initialization distribution using Monte Carlo Method + Importance sampling. Local KL Volume measures the "size" of the parameter region where the output distribution remains nearly unchanged (KL divergence ≤ ε) from the perspective of the initialization distribution.
KL volume is data-dependent, and the ratio of KL local volume between test and train datasets can be used to assess Overfitting. If the ratio of valid to train is less than 1, it indicates overfitting; if it is close to 1, it's optimal; and if greater than 1, it suggests Underfitting. Using second-moment information from optimizers like Adam Optimizer reduces directional variance, significantly decreasing the variance in volume estimation. The negative log of local volume can be interpreted as network information content (from an MDL perspective) and linked to generalization performance. As training progresses towards overfitting, local volume decreases (complexity increases).