Over the last few years, we have seen extraordinary transformations in the AI/ML landscape. In particular, we are witnessing the rise of AI foundational models (FM), and services underpinned by these foundational models (also known as, neural platforms), such as ChatGPT, BLOOM, Chinchilla, PaLM, and LLaMA, taking the center stage of mainstream AI. The foundational AI models, which are primarily large language models (LLMs) or multimodal models trained on a vast quantities of data at unprecedented scales, with billions to trillions of parameters, are applied to various downstream tasks to achieve state-of-the-art (SOTA) or near-SOTA results. The scientific community is yet able to adopt these advances, as has been shown by limitations of the recent Galactica model. We are facing the perennial and challenging questions of “Can we, and if possible, how to leverage and incorporate foundational models for advancing and accelerating scientific discoveries and integrate FMs into scientific workflows?”.
Although FMs are demonstrating remarkable capabilities on a number of tasks, there are many concerns and challenges that must be overcome before adoption of such techniques in scientific workflows. For instance, for consumer space use cases, a non-subtle erroneous, or even totally wrong response from these models (e.g., ChatGPT) can easily be tolerated, perhaps with a regeneration process or lightly appreciating the error as a form of creativity. However, in scientific workflows, such errors can not only be very catastrophic, but also potentially erode the confidence and trust between the scientists and AI.
Relying on foundational models for understanding, extracting and reasoning with knowledge from scientific datasets, making decisions such as AI-assisted planning and scientific experiment steering, has to come not only with responsibility but also with unprecedented level of verification. This is only possible, if the models are capable of (1) learning from multimodal data streams (2) being adaptive for a variety of downstream tasks, (3) potentially incorporating first-principle based domain knowledge, (4) being rigorously validated using the state-of-the-art benchmarks and beyond performance metrics e.g., reliability, robustness, and trust, (5) being prompted in a way to reduce post-training errors and hallucinations.
In this article collection, we intend to publish a collection of high-quality articles from broader intersecting areas of high-performance computing, foundational AI models, large-scale domain-adaptation of AI models, and training, fine-tuning, evaluation and prompting methods to address challenges with scientific workflows. The topics of interests we seek in this Research Topic include but not limited to:
• Unimodal and multimodal foundation AI models for science with experimental evidence of current limitations and techniques to overcome them e.g., prompting.
• Scaling, tuning, and performance optimization for facilitating AI foundational models for scientific workflows.
• Beyond industry benchmarks (e.g., BigBench) for unimodal and multimodal foundation models to ensure trusted and responsible Fm development and deployment.
• Example applications and use cases for FMs for science workflows, including model interaction and downstream task adaptation in zero and few-shot setting.
• Adaptation of domain or first-principle domain knowledge in foundation models.
• Machine- and deep-learning workflows integration for HPC, especially in the context of multimodal FMs for science applications.
Keywords:
artificial intelligence, workflows, high performance computing, foundational models, scientific workflows
Important Note:
All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.
Over the last few years, we have seen extraordinary transformations in the AI/ML landscape. In particular, we are witnessing the rise of AI foundational models (FM), and services underpinned by these foundational models (also known as, neural platforms), such as ChatGPT, BLOOM, Chinchilla, PaLM, and LLaMA, taking the center stage of mainstream AI. The foundational AI models, which are primarily large language models (LLMs) or multimodal models trained on a vast quantities of data at unprecedented scales, with billions to trillions of parameters, are applied to various downstream tasks to achieve state-of-the-art (SOTA) or near-SOTA results. The scientific community is yet able to adopt these advances, as has been shown by limitations of the recent Galactica model. We are facing the perennial and challenging questions of “Can we, and if possible, how to leverage and incorporate foundational models for advancing and accelerating scientific discoveries and integrate FMs into scientific workflows?”.
Although FMs are demonstrating remarkable capabilities on a number of tasks, there are many concerns and challenges that must be overcome before adoption of such techniques in scientific workflows. For instance, for consumer space use cases, a non-subtle erroneous, or even totally wrong response from these models (e.g., ChatGPT) can easily be tolerated, perhaps with a regeneration process or lightly appreciating the error as a form of creativity. However, in scientific workflows, such errors can not only be very catastrophic, but also potentially erode the confidence and trust between the scientists and AI.
Relying on foundational models for understanding, extracting and reasoning with knowledge from scientific datasets, making decisions such as AI-assisted planning and scientific experiment steering, has to come not only with responsibility but also with unprecedented level of verification. This is only possible, if the models are capable of (1) learning from multimodal data streams (2) being adaptive for a variety of downstream tasks, (3) potentially incorporating first-principle based domain knowledge, (4) being rigorously validated using the state-of-the-art benchmarks and beyond performance metrics e.g., reliability, robustness, and trust, (5) being prompted in a way to reduce post-training errors and hallucinations.
In this article collection, we intend to publish a collection of high-quality articles from broader intersecting areas of high-performance computing, foundational AI models, large-scale domain-adaptation of AI models, and training, fine-tuning, evaluation and prompting methods to address challenges with scientific workflows. The topics of interests we seek in this Research Topic include but not limited to:
• Unimodal and multimodal foundation AI models for science with experimental evidence of current limitations and techniques to overcome them e.g., prompting.
• Scaling, tuning, and performance optimization for facilitating AI foundational models for scientific workflows.
• Beyond industry benchmarks (e.g., BigBench) for unimodal and multimodal foundation models to ensure trusted and responsible Fm development and deployment.
• Example applications and use cases for FMs for science workflows, including model interaction and downstream task adaptation in zero and few-shot setting.
• Adaptation of domain or first-principle domain knowledge in foundation models.
• Machine- and deep-learning workflows integration for HPC, especially in the context of multimodal FMs for science applications.
Keywords:
artificial intelligence, workflows, high performance computing, foundational models, scientific workflows
Important Note:
All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.