Kyle Roth

This was a paper I presented about in Bang Liu’s research group meeting on 2022-10-21. You can view the slides I used here.

This was a paper I presented about in Bang Liu’s research group meeting on 2022-09-30. You can view the slides I used here.

Selective annotation chooses a pool of samples to annotate from a large set of unlabeled data. The main result of the paper is that when this is combined with item-specific prompt retrieval the performance drastically improves (>10% relative gain and lower performance variance). Interestingly, selective annotation does not help for finetuning, or when the prompts are randomly selected. They call their selective annotation method “vote-\(k\)”. selective annotation method permalink Vote-\(k\) essentially creates a network of similaraccording to Sentence-BERT unlabeled instances, and then selects from them with a network importance score that is discounted to promote diversityThe discounting is performed by iteratively adding to the selection set, each time penalizing new nodes for being close to nodes that are already in the selection set. .

We observe that 48.2% [of] images [in ImageNet] are learned by all models regardless of their inductive bias; 14.3% [of] images are consistently misclassified by all models; only roughly a third (37.5%) of images are responsible for the differences between two models’ decisions. We call this phenomenon dichotomous data difficulty (DDD). The authors varied hyperparameters, optimizers, architectures, supervision modes, and sampling methods, finding that models only varied in performance on about a third of the images in the dataset. And this isn’t specific to ImageNet; they found similar results for CIFAR-100 and a synthetic Gaussian dataset. They use this measure to divide the dataset into “trivials”, “impossibles”, and “in-betweens”.

In this paper they show that we can achieve exponential performance scaling over dataset size, when the samples added are pruned to be only the best examples. This beats power law scaling in a big way. There is still no free lunch, in some sense, because in most cases it will become progressively harder to add new useful samples as the dataset gets bigger. But this is a big deal for computation, because it means that the number of samples in the dataset is not nearly as important as the coverage and quality that the dataset provides.This also means that scaling laws for compute (usually expressed as a function of dataset and model size) are dataset-specific and not generalizable, because of how much sample quality affects data scaling.