AI for science: a moving target

Many AI tools are also rapidly being developed to support scientists' research. AI-powered search engines function as digital assistants. They can provide yes-no answers to research questions, tidy up bibliographies, suggest new  papers, and generate summaries.

And, of course, the advent of large-language models (LLMs) such as ChatGPT has already had a big impact on scientists' productivity. LLMs are now used to conduct literature reviews and brainstorm research ideas, and many researchers used them to write code, manuscripts, grant applications, and summaries.

But Wang et al. point to many challenges as well. For example, training data for AI algorithms must be annotated reliably, requiring time-consuming and resource-intensive experimentation and simulations. Algorithmic approaches such as pseudo-labeling and label propagation have been suggested to automate the annotation of vast unlabeled datasets based on only a small set of accurate annotations. In pseudo-labeling, surrogate models - trained on manually labelled data - annotate unlabeled samples. In label propagation, labels are diffused to unlabelled samples via similarity graphs constructed based on so-called feature embeddings. In addition to these strategies, active learning could also be used to identify the most informative data points to be labelled by humans or the most important experiments to be done.

AI techniques are also useful for refining data, for example, to significantly increase resolution (through deep convolutional methods) and decrease noise (through so-called denoising autoencoders) in measurements.

And Wang et al. point out that deep learning can extract meaningful representations of scientific data at various levels of abstraction and optimize them to guide research. Three such strategies are geometric deep learning (based on graph neural networks that model relational patterns in the data), self-supervised learning (a technique that enables models to learn the general features of a dataset without relying on explicit labels), and language modelling (particularly useful for biological sequences).

The flip side of AI4Science is the emergence of many difficult questions that need urgent answers. One question raised by Wang et al. relates to AI models' ability to generalize. This is currently a prospective weak point in AI models because neural networks trained on data from a specific regime may discover patterns that do not generalize to a different regime, whose underlying distribution is different. Humans are better at generalizing from data, the hypothesis being that we do so because we use causal rather than statistical models when we make inferences. Hence, the review authors argue that implementing causality in AI would be key.

Another step in the AI-development roadmap pointed out by Wang et al. includes developing strategies for accounting for multimodal input to the training set. The authors argue this could prospectively be implemented in neural networks by exploiting their modularity - distinct neural modules could transform diverse data modalities into universal vector representations.

An overarching goal should also be to explain how AI models work. Many models currently operate as black boxes, which poses challenges for their use and risks decreasing trust in AI.

AI Developments need to extend beyond algorithmic developments. As the authors point out, AI can only transform research if AI expertise is incorporated into research teams. Research teams will need to include AI specialists, software and hardware engineers, and new forms of collaboration. Resources would likely need to be shared because the computational and data requirements associated with AI models are massive.

Educational programs would need to be revamped to train scientists in designing, implementing, and applying laboratory automation and AI in scientific research. They should also teach scientists when the use of AI is appropriate and how to use AI ethically. Achieving this requires tackling many thorny ethics, privacy, security, and intellectual property issues.

The coming years will see stakeholders in the research ecosystem navigate uncharted territory. There are many visions and interests for AI developments in research. Hiroaki Kitano articulated one of these in his perspective Nobel Turing Challenge: creating the engine for scientific discovery. His Nobel Turing Challenge "...aims to develop a highly autonomous AI system that can perform top-level science, indistinguishable from the quality of that performed by the best human scientists, where some of the discoveries may be worthy of Nobel Prize level recognition and beyond."

According to Kitano, the key to this challenge would be to develop 'AI Scientists', comprised of software and hardware modules that dynamically interact to generate hypotheses, learn from data and interactions with humans and other parts of the system, reason, and autonomously make decisions.

To achieve this, he proposes that these AI Scientists would not only be capable of autonomously conducting scientific research to generate scientific discoveries at scale but also to make strategic choices on the topic of research and communicate through publications and other means to explain the value, methods, reasoning behind the discovery, and their applications and social implications.

Kitano posits that AI Scientists would be good at exploring a vast hypothesis space and that, therefore, it would enable moving away from what he calls the traditional approach to doing research. In this approach, scientists aim to maximize the probability that the discovery they make will be significant under certain criteria—he dubs this hunt for significant discoveries a value-driven approach.

The alternative approach he proposes would maximize the probability of discovery at any level of significance instead of maximizing value, a process he dubs unbiased exploration of hypotheses. He views this transition as a "logical evolution of the modality of science where a vast hypothesis space is searched in an unbiased manner rather than depending on human intuition."

With an increased level of autonomy, AI Scientists would make decisions on which topics and hypotheses to pursue. Kitano suggests that such strategies could be based on 1) goal-oriented approaches of defining very high-level goals and finding multiple paths to best achieve such goals, or 2) a bottom-up approach of exploring hypothesis search space based on discoveries already made by a specific AI Scientist.

He further states that "The real value of AI Scientist is its capability to explore hypothesis space magnitude more efficiently into seemingly low-value domains with the expectation that may eventually lead to major outcomes.", something that he claims would be infeasible to do by human scientists. He suggests that there would be two roles for AI Scientists: 'AI Scientist as a Problem Solver' (where the problem-solving would be aligned with the value of stakeholders) and 'AI Scientist as an Explorer' (which would "boldly explore hypothesis space nobody has gone before"). The common denominator for these roles, and what Kitano views as distinguishing the AI Scientist discovery process from the traditional one, is the exhaustive hypothesis generation and verification.

For implementation, Kitano envisions two possible characteristics for an architecture of AI Scientists: 1) a multiplexed agent system where multiple instances of specialized AI agents are created to explore hypotheses spaces organically whilst communicating to merge discoveries for further exploration; 2) a human-in-the-loop system, where humans enter as domain experts or in commanding and monitoring roles.  The AI Scientists may exist at the institutional, academic community, or national levels and communicate with each other in networks formed based on data ownership, privacy, and intellectual property considerations.

Like many other AI-related visions, Kitano's AI-Scientist vision focuses predominantly on opportunities, which is understandable. But many thorny questions arise in the rapidly evolving AI4Science.

Take, for instance, Kitano's claim that the transition from a value-based to exploration-based discovery is a "logical evolution of the modality of science where a vast hypothesis space is searched in an unbiased manner rather than depending on human intuition." It is neither clear why this transition is logical nor is it clear that it would be desirable - after all, why should we ignore human intuition? Why is the purportedly unbiased exploration approach desirable as a replacement for the current approach? To what extent can these approaches generate mechanistic insights and causal principles rather than correlations-driven predictions only? Why would it be desirable to move away from value-driven research in the first place?

These are only a few questions that would need addressing in the rapidly developing AI4Science realm. Other pressing questions relate to issues such as: intellectual property and attribution; evaluation of scientists' research output; inequalities arising in the research ecosystem because of the major investments needed; biases in AI algorithms and training data; reproducibility issues; dual-use and security risks; energy demands of AI systems; fraud and misinformation that could be accelerated with AI; the role of scientists and the interplay between different stakeholders in the research ecosystem as AI permeates it.

Tackling these questions requires researchers from all disciplines (science, technology, social science, humanities), funders, publishers, policy-makers, legal experts, tech companies, and key non-governmental organizations that represent everyone that will be affected by these exponential developments, to come together to discuss the various visions that exist and set up common frameworks, standards, and guidelines. 

I aim to look into some of these developments in future blog posts.