Cracking Open the Black Box: Meet KREPES

Self-supervised learning (SSL) has long been a powerhouse, capable of unearthing patterns from vast swathes of unlabeled data. But let’s face it, these models often operate like mysterious black boxes. Enter KREPES, a new framework designed to shed light on the internal workings of SSL models like SimCLR, BYOL, and VICReg.

Beyond the Black Box

KREPES does something rather fascinating. It leverages empirical neural tangent kernel approximations along with the Representer Theorem for kernels to directly represent the learned latent space. In plain English, it makes it possible to track those influential, unlabeled examples that leave the biggest footprint in the model’s training process. These are what KREPES calls 'Representer Landmarks'.

Why does this matter? Because understanding which training examples are steering the ship can reveal hidden biases. The framework's ability to audit these landmark influences means we can now pinpoint biases lurking in datasets like Adult-1M, where demographic proxies for income sneak into the learning process.

Introducing New Metrics

KREPES isn't just about exposing what’s under the hood. It offers new ways to measure transparency in learned representations. With metrics like 'Sample-Specific Influence Score', 'Concept-Conditioned Influence Score', and 'Feature Alignment Gap', KREPES quantifies how aligned a model's understanding is with reality. This can fundamentally change how we view and evaluate these models.

Ask yourself, how many times have we trusted AI without questioning its training data? This framework is a breakthrough in holding these models accountable.

Scaling to the Big Leagues

The framework doesn't shy away from scalability challenges either. To accommodate massive datasets like ImageNet-1K and Adult-1M, KREPES employs a Nyström approximation-based analytical inference method. This ensures it can handle over a million samples without breaking a sweat.

But the real question here's, who will use KREPES? Will it be embraced by industry leaders hungry for transparency, or will it be left in the hands of the few who dare to question their models? Whose data? Whose labor? Whose benefit? The potential is immense, but it demands a shift in mindset.