AI Research Driven by Real-World Problems

A robotic arm in a greenhouse picking tomatoes.

In recent years, artificial intelligence (AI) research has begun to be applied in many different fields, from healthcare to climate science. However, most AI algorithms were not designed with specific problems in mind and often fall short when trying to solve them. Instead of solely developing ever-more powerful and general AI models for innovation’s sake, we need to focus more on the specific needs and constraints of problems we actually want to solve.

Amidst a surge in adoption of machine learning, many algorithms developed by machine learning researchers are now applied in use cases they were not originally designed for: For example, large language models (LLMs) - like the one behind OpenAI’s chatbot ChatGPT - were not designed to say things that are true, they were designed to say things that sound good. Indeed, there is currently a lot of hype around AI tools, which can lead people to use “flashy” AI solutions instead of solutions that could actually help solve their problems.

If we want to move beyond these blind spots and design AI algorithms that better align with societal goals, now is a good time to embrace a new research paradigm.


Complementary Approaches to AI Research

In our paper Application-Driven Innovation in Machine Learning, which was just accepted at ICML 2024, we argue that using the problem we want to solve as a starting point for research will allow us to better use AI for solving it down the line.

Methods-driven AI research usually involves widely-used and standardized benchmark tasks and data sets such as ImageNet or the OpenAI Gym. These benchmarks, and the algorithms that build on them, aim to represent abstract problem statements, but often fall short from what people are actually using machine learning for. Indeed, the most general purpose formulation of a machine learning solution is often directly useful to nobody, instead of being useful to everybody. 

In contrast, application-driven machine learning research is a form of innovation driven by data and tasks from specific problems in consultation with experts in related fields, who can best define what solving the solution entails. This could mean making the model more interpretable (according to the specific needs of the user), incorporating known constraints from physics and engineering, or better quantifying the uncertainty in the output.

Let us be clear: we are not saying that application-driven innovation needs to replace the methods-driven approach to AI research. Rather, the two are meant to complement and enrich each other. Oftentimes, application-driven work builds on methods-driven work, and the other way around: by focusing on the needs of a particular problem, we can come up with cross-cutting innovations that are useful in many other problems.

Indeed, some of the biggest innovations in machine learning in recent years - such as U-NetPhysics-informed Neural Networks and Fourier Neural Operators - have emerged from application-driven innovation.


AI Research Should Focus on Impact

For years, subcommunities of AI researchers have been building algorithms in close collaboration with the multidisciplinary experts who will be using them. For example, machine learning for healthcare has a very rich history of designing innovative methods based upon needs faced by doctors and public health experts.

But such research hasn’t necessarily been given the platform that it deserves within the mainstream machine learning research world: it is often dismissed as falling outside of machine learning and only interesting and relevant within a narrow application. 

The algorithmic innovations introduced in application-driven research are often confused with simple engineering tasks, and their utility to machine learning research is underappreciated. This is unfortunate, as computer scientists and AI experts have a lot to learn from those who are using their tools on the ground.

In order for AI to be used in ways that actually help people, we need to focus on impact from the start. Thinking consciously about how what we build is being used throughout the innovation process is essential to ensure the right kinds of downstream impacts we want to see in the world.

We believe that the machine learning community is starting to see the value in application-driven innovation. We call for a better recognition of the duality between these two approaches, and for empowering them both as complementary angles on machine learning research.

Innovation should be driven by what we want to achieve, not just by what we can build.