Thanks for checking out my poster!
Links
👇 Learn More Below 👇
Black-box model access means we only know model inputs and model outputs. Using what we know, we let the collection of input-output pairs define the model itself.
Why black box models?
- A lot of work in ML interpretability requires access to things like internal activations or model weights, but most users don’t have this information. Even with access, memory and storage can limit scalability.
- The black box setting is more general and less computationally constrained, making it a great choice for performing inference with model collections.
To simplify analysis, we can represent each model as a vector by using model responses to a set of input queries1. Prior work shows that these vector representations are effective for inference tasks on collections of models. Selecting good queries ensures better vector representations — and better results in inference — with lower cost.
The Discriminative Factorization framework is a way to decompose distances between black box models that allows us to analyze the impact of individual queries on the model representations.
The Discriminative Factorization makes model inference better
The discriminative factorization gives us a way to find the best queries for the job.
Better queries means:
- more information in fewer computations. Inference is more efficient
- better representations. Operate on the optimistic side of existing guarantees
With this framework, we can find the good queries without needing to know what the task is. We can find good queries for any collection of models.
When task type is known, we can guarantee performance for a particular query budget.
- e.g. for classification we only need label data to guarantee error bounds wrt query count… and we don’t need to know what the classes “mean” (!)
Footnotes
-
In this setting, we then treat a collection of outputs (repsponses) as an estimate for the model itself. To make different outputs directly comparable, we choose an embedder that takes the raw model outputs (text, images, etc) and converts them into vectors. This allows us to obtain a matrix surrogate for each model. We estimate the underlying geometry (obtain a vector) by computing pairwise distances between the model matrices. ↩