LaFTer: Label-Free Tuning of Zero-shot Classifier using Language and Unlabeled Image Collections Supplementary Material

Christian Doppler Laboratory for Embedded Machine Learning. In real-world applications, the unlabeled image collection, such as the one we use in LaFTer (in conjunction with the text-only training) can also contain unrelated images, e.g., images of other These results are provided in Table 1. Results for CIFAR-100 in this scenario, also follow a similar trend. We follow [5] and query GPT -3 with different prompts in order to obtain descriptions for each class. How can you identify a category?