Intriguing properties of generative classifiers
This work addresses the problem of improving object recognition models to be more human-like, which is incremental as it builds on existing generative modeling techniques.
The paper investigated whether generative models, when used as classifiers, better approximate human object recognition compared to discriminative models, finding that generative classifiers achieve a 99% human-like shape bias, near human-level out-of-distribution accuracy, and state-of-the-art alignment with human errors.
What is the best paradigm to recognize objects -- discriminative inference (fast but potentially prone to shortcut learning) or using a generative model (slow but potentially more robust)? We build on recent advances in generative modeling that turn text-to-image models into classifiers. This allows us to study their behavior and to compare them against discriminative models and human psychophysical data. We report four intriguing emergent properties of generative classifiers: they show a record-breaking human-like shape bias (99% for Imagen), near human-level out-of-distribution accuracy, state-of-the-art alignment with human classification errors, and they understand certain perceptual illusions. Our results indicate that while the current dominant paradigm for modeling human object recognition is discriminative inference, zero-shot generative models approximate human object recognition data surprisingly well.