A foundational model starts with training already behind it. Its learned weights can be loaded and used on new data instead of training a network from scratch. Running it on new data is called inference. Some pretrained models serve a particular task, such as image classification; a model described as foundational is useful across a broader range of inputs or tasks. section 2.7 section 15.3.1
Two practical starting points show the range:
- For image recognition, load a model such as AlexNet through TorchVision with pretrained weights, feed it an image, and inspect its predicted labels. Its training data included labeled ImageNet images. section 2.1 section 2.1.1
- For segmentation, Segment Anything can produce object masks from an image and accept points, boxes, or masks as prompts. Its zero-shot use does not require training on each new dataset; CT nodule segmentation is one application. section 15.3.1
The architecture and the weights are different things. TorchVision supplies model architectures and access to pretrained weights. A saved PyTorch state dictionary contains parameters, not the model’s structure: loading one requires an instance of the corresponding model class. Further training is also possible; one exercise fine-tunes a pretrained Vision Transformer on CIFAR-10. section 2.1.1 section 8.4.2 section 9.11
