From Fake to Real: Pretraining on Balanced Synthetic Images to Prevent Spurious Correlations in Image Recognition
The episode discusses a paper proposing 'From Fake to Real' (FFR), a two-step training process using balanced synthetic images to prevent spurious correlations in image recognition models. The authors argue that mixing fake and real data introduces new biases, which FFR avoids by pre-training on balanced synthetic data separately from the real data. This method improves performance significantly over existing techniques.