English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Visual recognition models are prone to learning spurious correlations inducedby a biased training set where certain conditions $B$ (\eg, Indoors) areover-represented in certain classes $Y$ (\eg, Big Dogs). Synthetic data fromgenerative models offers a promising direction to mitigate this issue byaugmenting underrepresented conditions in the real dataset. However, thisintroduces another potential source of bias from generative model artifacts inthe synthetic data. Indeed, as we will show, prior work uses synthetic data toresolve the model's bias toward $B$, but it doesn't correct the models' biastoward the pair $(B, G)$ where $G$ denotes whether the sample is real orsynthetic. Thus, the model could simply learn signals based on the pair $(B,G)$ (\eg, Synthetic Indoors) to make predictions about $Y$ (\eg, Big Dogs). Toaddress this issue, we propose a two-step training pipeline that we call FromFake to Real (FFR). The first step of FFR pre-trains a model on balancedsynthetic data to learn robust representations across subgroups. In the secondstep, FFR fine-tunes the model on real data using ERM or common loss-based biasmitigation methods. By training on real and synthetic data separately, FFRavoids the issue of bias toward signals from the pair $(B, G)$. In other words,synthetic data in the first step provides effective unbiased representationsthat boosts performance in the second step. Indeed, our analysis of high biassetting (99.9\%) shows that FFR improves performance over the state-of-the-artby 7-14\% over three datasets (CelebA, UTK-Face, and SpuCO Animals).

From Fake to Real: Pretraining on Balanced Synthetic Images to Prevent Bias