Foundation models

The Stanford Institute for Human-Centered Artificial Intelligence's (HAI) Center for Research on Foundation Models (CRFM) coined the term "foundation model" in August 2021, tentatively referring to "any model that is trained on broad data (generally using self-supervision at scale) that can be adapted (e.g., fine-tuned) to a wide range of downstream tasks".[14] This was based on their observation that existing overlapping terms were not adequate, submitting that "'(large) language model' was too narrow given [the] focus is not only language; 'self-supervised model' was too specific to the training objective; and 'pretrained model' suggested that the noteworthy action all happened after 'pretraining."[15] After considering many terms, they settled on "foundation model" to emphasize the intended function (i.e., amenability to subsequent further development) rather than modality, architecture, or implementation.

They also note that the concept is not truly new, as it is based on deep neural networks and self-supervised learning, but asserted that the scale at which the area has developed in recent years, and the increasing potential for any given model to be used for different purposes, warranted a new term.[14]

A foundation model is a "paradigm for building AI systems" in which a model trained on a large amount of unlabeled data can be adapted to many applications.[16][17] Foundation models are "designed to be adapted (e.g., finetuned) to various downstream cognitive tasks by pre-training on broad data at scale".[18]

Key characteristics of foundation models are emergence and homogenization.[14] Because training data is not labelled by humans, the model emerges rather than being explicitly encoded. Properties that were not anticipated can appear. For example, a model trained on a large language dataset might learn to generate stories of its own, or to do arithmetic, without being explicitly programmed to do so.[19] Furthermore, these properties can sometimes be hard to predict beforehand due to breaks[20] in downstream scaling laws. Homogenization means that the same method is used in many domains, which allows for powerful advances but also the possibility of "single points of failure".[14]

Since foundation models are pre-trained on a massive dataset, they are not capable of handling specific "personal" concepts that a user may be interested in. A series of methods were designed to augment a foundation model with personal, specific items without retraining the full model. For example, for few-shot image retrieval it was shown how to adapt a vision-language foundation model (CLIP) by adding new concept to its vocabulary.[21] For text-to-image generation, an approach called textual inversion[22] can be similarly used to teach the system new concept that can later be generated in conjunction with the concepts that the foundation model is already familiar with.

A 2021 arXiv report listed foundation models' capabilities in regards to "language, vision, robotics, reasoning, and human interaction", technical principles, such as "model architectures, training procedures, data, systems, security, evaluation, and theory", their applications, for example in law, healthcare, and education and their potential impact on society, including "inequity, misuse, economic and environmental impact, legal and ethical considerations".[14]

An article about foundation models in The Economist notes that "some worry that the technology's heedless spread will further concentrate economic and political power".[19]