By Bob de Vos and Isabell Trinh
From HAL-9000 to R2D2 to Jarvis, machines with artificial intelligence have appeared in fiction countless times. When OpenAI launched GPT-4, fiction turned reality in one fell swoop—suddenly, machines could actually understand and communicate with us. But what if we could harness the same technical principles behind large language models (LLMs) to revolutionize how we interpret complex data?
We are currently developing foundation models to interpret spectral data—a type of data with applications in healthcare, food safety, environmental monitoring, and, in our case, diagnostics. Foundation models form the backbone of many large language models (LLMs), and ours can do for spectral data what GPT has done for text: simplify complexity and unlock transformative applications.
Our diagnostic test uses surface-enhanced Raman spectroscopy (SERS) to generate spectra of bacteria, enabling us to identify which species they belong to. You can imagine two spectra of the same bacterial species as two people describing the same person: they might not use the exact same words, and both descriptions may be incomplete, but you can tell who they’re talking about—especially when listening to both.
With our foundation models, we can understand the language of spectra the same way we interpret the meaning of words in a sentence. With minimal training to establish a basic description, our models can perform tasks such as determining which species a spectrum represents. This is called few-shot learning, which is particularly valuable when learning new tasks with limited data.
Beyond accelerating and improving identification, our foundation models also simplify the storage, retrieval, and analysis of spectral data. The best part? They work with any kind of spectral data. Whether you’re a small lab or an enterprise team, foundation models can seamlessly integrate into your workflow.
In the upcoming blog posts, we will explore how. Follow us on LinkedIn and visit our website to learn more!
