Zero-Shot Model Search via Text-to-Logit Matching

With the increasing number of publicly available models, there are pre-trained, online models for many tasks that users require. In practice, users cannot find the relevant models as current search methods are text-based using the documentation which most models lack of. This paper presents ProbeLog, a method for retrieving classification models that can recognize a target concept, such as "Dog", without access to model metadata or training data. Specifically, ProbeLog computes a descriptor for each output dimension (logit) of each model, by observing its responses to a fixed set of inputs (probes). Similarly, we compute how the target concept is related to each probe. By measuring the distance between the probe responses of logits and concepts, we can identify logits that recognize the target concept. This enables zero-shot, text-based model retrieval ("find all logits corresponding to dogs"). To prevent hubbing, we calibrate the distances of each logit, according to other closely related concepts. We demonstrate that ProbeLog achieves high retrieval accuracy, both in ImageNet and real-world fine-grained search tasks, while being scalable to full-size repositories. Importantly, further analysis reveals that the retrieval order is highly correlated with model and logit accuracies, thus allowing ProbeLog to find suitable and accurate models for users tasks in a zero-shot manner.