FDSL
A method that generates synthetic images from mathematical formulas such as fractals and performs pretraining with discretized-label classification. It enables pretraining of vision models without real data.
Timeline (last 90 days)
FDSL (Formula-Driven Supervised Learning) generates images from fractal formulas and discretizes labels to accomplish classification-style pretraining.
FDSL's discrete label classification approach discards the continuous relational information contained in the image generation parameters, limiting the model's ability to learn fine-grained relationships, especially affecting tasks that depend on relative ordering such as image retrieval.
All Facts (2)
FDSL (Formula-Driven Supervised Learning) generates images from fractal formulas and discretizes labels to accomplish classification-style pretraining.
50%UnverifiedFDSL's discrete label classification approach discards the continuous relational information contained in the image generation parameters, limiting the model's ability to learn fine-grained relationships, especially affecting tasks that depend on relative ordering such as image retrieval.
50%