On the Opportunities and Risks of Foundation Models

Closely connected to model’s expressivity is the notion of scalability. As rich data from varied sources becomes more readily available, and computational resources get stronger and more efficient (§4.5: systems), we should look for ways to match this rate of progress and harness it to improve AI competency and versatility. For foundation models to effectively fit the complex and high-dimensional distribution of images or text, they should thereby be scalable across all dimensions: including both models’ depth and width as well as their training time, number of parameters, and amount of data they could process.

Optimization

Specifically, foundation models should both be: (1) easy-to-train (§4.2: training), by being resilient to noise or imperfections in the data, and robust against instabilities like vanishing (Helfrich et al., 2018; Glorot and Bengio, 2010) or exploding gradients (Hochreiter and Schmidhuber, 1997; Nair and Hinton, 2010), but also (2) easy-to-adapt (§4.3: adaptation), by overcoming phenomena of catastrophic forgetting (Kirkpatrick et al., 2017) and supporting few-shot learning (Sung et al., 2018). We are still in the early days of understanding the principles that drive the scalability of learning algorithms, and while recent works have started to shed some light on these themes (Liu et al., 2020c; Kuditipudi et al., 2019; Nakkiran et al., 2019), much work remains to be done.

Hardware Compatibility

Moving beyond aspects of robustness and optimization, foundation models should also be practically efficient (§4.5: systems), and take advantage of contemporary and future hardware (Hooker, 2020). One example of that is parallelizablity, an important property that characterizes the computation supported by GPUs. Indeed, much of the transformers’ great success over the previously dominating recurrent approach was driven by their higher degree of parallelism.

Looking forward, given the fast-pace progress of systems development, we should further ensure that models are designed to co-adapt to future hardware advances. Consequently, foundation models should ideally be amenable to schemes such as distributed training, which is gaining popularity, as is the case for e.g., Mixture-of-Experts, and possibly leverage properties such as sparsity of the computation or representation, as is the case for the Longformer (Beltagy et al., 2020), BigBird (Zaheer et al., 2020), and Sparse Transformer (Child et al., 2019) approaches, and which likely will become more central in future hardware and processors.

Multimodality

Traditionally, the fields of computer vision, robotics, and NLP have made progress in an independent manner, with separate communities developing specific approaches suitable for each modality. A conducive consequence the rise of deep learning has brought about was the bridges it helped forming among the various communities and research areas within AI, as seemingly different problems could now be tackled by closely-related approaches, and studies of originally remote topics began converging to a common ground. This breakthrough opened up a new range of possibilities, fostering pioneering exploration into the theme of multimodality, encompassing areas as varied as language grounding (Lynch and Sermanet, 2020), visual semantics (Conser et al., 2019), embodied environments (Savva et al., 2019a) and interactive agents (Gray et al., 2019).

Essentially, multimodality serves as a key component of intelligence, and is a crucial factor for the development of both thorough and broad comprehension of the world. Concretely, language learning is more effective when occurring in a grounded environment rather than in a vacuum. And inversely, from the vision perspective, language encourages the emergence of abstractions that link between low-level perceptual signals and statistics to semantic concepts of objects, properties, agents and motivations, thereby enriching and elevating visual representations.

In light of these observations, we argue that foundation models should ideally connect together the different modalities, distill their embodied information into a shared multifaceted representation, and capture the full range of inter-connections and relations among them so as to furnish a wide range of capabilities (see §2.1: language, §2.2: vision,§2.3: robotics, §2.4: reasoning).

Generality and Specialization

An important design choice for multimodal foundation models is the degree of specialization, or the structural sharing between the modules responsible for each modality. Naturally, data of different domains exhibits diverse kinds of structures and properties — where, for instance, language is discrete while vision is continuous. At first sight, this variation hints that specialized inductive biases tailored for each modality could be of aid. Yet, as training scales upwards and models are provided with the opportunity to base their learning less on structural priors and more on the data itself, general approaches that maintain only a handful of broad general assumptions prove in fact a lot more successful than task-specific alternatives. And so, as corroborated by recent success of general-purpose models like transformers across different modalities — both linguistic (Liu et al., 2019; Lan et al., 2019) and visual (Dosovitskiy et al., 2020; Hudson and Zitnick, 2021), we see that generality is critical for improving AI capabilities.

Multimodal Interactions

Another key consideration for multimodal models relates to weight sharing: do the various modalities benefit from using the same or different parameters for their respective components? Prior works have shown that fruitful transfer could certainly occur across modalities, but the ideal degree of sharing remains unclear, so is the existence of principled ways for discovering it.

Finally, a major design question concerns with the forms of the multimodal interactions supported by the model, which vary widely between concrete cases and examples: Cross-modal or late-fusion models such as ConVIRT (Zhang et al., 2020a) and CLIP (Radford et al., 2021) maintain fully separate encoders for each data source, and compare their spaces only at the ultimate computation stage, using e.g., a simple dot product. Meanwhile, early-fusion models, such as ViLBERT (Lu et al., 2019a; Cho et al., 2021), jointly reason over multiple modalities necessary for tasks of visual reasoning and question answering. Identifying the optimal stage and form for merging the respective vector spaces (Nagrani et al., 2021) remains an open research question.

Overall, while there seems to be a consensus within the community about the importance of multimodality, models that go beyond shallow alignment of vision and language are yet to exist, and the theme of grounded language learning in embodied environments still has much room for exploration.

Memory

So far, we have discussed the foundation models’ goal to gather and accumulate information from varied modalities at large scales. This knowledge encompasses both broad understanding of the world as well as specific mastery of niche subjects or particular facts. Representing such a large body of learned information is by no means trivial, and is leading to interesting questions about effective mechanisms for access, storage, retrieval and manipulation of particular items or memories.

Explicit Storage

An important design principle that could achieve these desiderata is to separate out computation from memory (Weston et al., 2014; Graves et al., 2016; Hudson and Manning, 2018; Hudson and Manning, 2019a) in order to enhance models’ ability to transfer knowledge by applying previously acquired abstract skills to new concrete settings.

In this context, it is important to distinguish between explicit facts — that can be stored in an external memory storage, and implicit knowledge — that is reflected through the networks’ trainable weights. Such decoupling of explicit and implicit knowledge enjoys multiple advantages compared to the alternative of implicitly encoding all information together through the network weights. The separation mitigates the inflation in models’ size and number of parameters needed to store the growing quantities of knowledge (Guu et al., 2020), improves models’ trust and reliability by increasing their knowledge provenance (Cheney et al., 2009), and most notably, is key for memory update, manipulation or adaptation (Lewis et al., 2020b) (§4.3: adaptation), which could in turn enable generalization to novel contexts and downstream tasks.

Indeed, disentanglement between memory and computation has been a recurring goal in deep learning and NLP research over the last years, including models such as Memory Networks (Weston et al., 2014; Sukhbaatar et al., 2015), the Neural Turing Machine (Graves et al., 2014; Graves et al., 2016), the Neural State Machine (Hudson and Manning, 2019a), and MAC (Hudson and Manning, 2018). Furthermore, using key-value structures (Miller et al., 2016) for accessing external memories has been shown to be very effective for modeling long-term dependencies (Henaff et al., 2016; Bosselut et al., 2018; Lample et al., 2019). Transformers, the celebrated architecture underlying most foundation models to date, likewise exhibits operations that involve key-value memory-access and computation among the contextual word representations they gradually build (Geva et al., 2020).