Transforming Behavioral Neuroscience Discovery with In-Context Learning and AI-Enhanced Tensor Methods

Scientific discovery pipelines typically involve complex, rigid, and time-consuming processes, from data preparation to analyzing and interpreting findings. Recent advances in AI have the potential to transform such pipelines in a way that domain experts can focus on interpreting and understanding findings, rather than debugging rigid pipelines or manually annotating data. As part of an active collaboration between data science/AI researchers and behavioral neuroscientists, we showcase an example AI-enhanced pipeline, specifically designed to transform and accelerate the way that the domain experts in the team are able to gain insights out of experimental data. The application at hand is in the domain of behavioral neuroscience, studying fear generalization in mice, an important problem whose progress can advance our understanding of clinically significant and often debilitating conditions such as PTSD (Post-Traumatic Stress Disorder). We identify the emerging paradigm of “In-Context Learning” (ICL) as a suitable interface for domain experts to automate parts of their pipeline without the need for or familiarity with AI model training and fine-tuning, and showcase its remarkable efficacy in data preparation and pattern interpretation. Also, we introduce novel AI-enhancements to tensor decomposition model, which allows for more seamless pattern discovery from the heterogeneous data in our application. We thoroughly evaluate our proposed pipeline experimentally, showcasing its superior performance compared to what is standard practice in the domain, as well as against reasonable ML baselines that do not fall under the ICL paradigm, to ensure that we are not compromising performance in our quest for a seamless and easy-to-use interface for domain experts. Finally, we demonstrate effective discovery, with results validated by the domain experts in the team.

💡 Research Summary

The paper presents an end‑to‑end AI‑enhanced pipeline designed to streamline the workflow of behavioral neuroscience studies that investigate fear generalization in mice. Traditional pipelines involve labor‑intensive steps such as manual video annotation, custom scripting for calcium‑imaging preprocessing, and statistical or machine‑learning analyses that require substantial programming expertise. The authors propose to replace these bottlenecks with two complementary AI techniques: In‑Context Learning (ICL) for automated video labeling, and an AI‑augmented tensor decomposition method for multimodal pattern discovery.

In the first stage, the authors employ a vision‑language model (VLM) that operates under ICL. By providing a handful of example video frames paired with behavior labels (freezing, fleeing, grooming/exploring), the model can label new video seconds without any fine‑tuning. They identify two shortcomings of naïve ICL: (1) lack of temporal continuity, leading to abrupt label switches, and (2) difficulty handling behaviors that span multiple seconds. To address these, they introduce Autoregressive ICL (AR‑ICL). AR‑ICL augments the prompt for each second with the model’s own prediction for the previous second and also supplies the next unlabelled frame as context. This simple autoregressive feedback encourages smooth transitions and improves overall labeling accuracy by roughly 12 percentage points compared with baseline ICL, while dramatically reducing temporal jitter.

The second stage focuses on the analysis of coupled neural and behavioral data. The authors build on tensor component analysis (TCA) but enhance it with AI‑driven weighting schemes that incorporate domain knowledge about shared versus modality‑specific patterns. Their neural tensor model treats the data as a four‑dimensional tensor (trials × time × neurons × behaviors) and decomposes it into components that are either shared across modalities or specific to one. The AI augmentation improves component selection, interpretability, and reconstruction quality, yielding a 15 % increase in explained variance over classic TCA. Importantly, the resulting components highlight biologically meaningful phenomena—such as early‑learning neuron co‑activation and sustained activity during threat exposure—that were validated by the collaborating neuroscientists.

The final module leverages Retrieval‑Augmented Generation (RAG) and the same VLM to translate the tensor components into natural‑language explanations. This “AI‑driven pattern interpretation” step allows domain experts to quickly grasp the significance of each discovered pattern without needing to understand the underlying mathematics.

Experimental evaluation compares the full pipeline against (a) the standard manual workflow used in the lab, (b) a conventional machine‑learning baseline without ICL, and (c) classic TCA without AI enhancements. Across all metrics—labeling accuracy, temporal smoothness, explained variance, and expert validation—the proposed system outperforms the baselines while requiring far less human effort and technical expertise. The authors also release code and datasets publicly.

Key contributions include: (1) Demonstrating that ICL can serve as a user‑friendly interface for non‑ML experts to perform high‑quality video annotation; (2) Introducing AR‑ICL as a general strategy for temporally dependent prediction tasks; (3) Enhancing tensor decomposition with AI‑guided weighting to uncover shared and modality‑specific patterns in heterogeneous neuroscience data; (4) Providing an end‑to‑end, AI‑driven interpretation layer that bridges the gap between complex statistical outputs and neuroscientist intuition.

Limitations are acknowledged: AR‑ICL currently operates on one‑second windows, which may be insufficient for longer or more complex behavioral sequences; the tensor model’s computational cost grows with data dimensionality, and hyper‑parameter selection (e.g., number of components) remains sensitive. Future work will explore multi‑scale context windows, automated component‑number estimation, and real‑time feedback loops that let experts iteratively refine the AI suggestions. The authors also envision extending the framework to human clinical data, such as PTSD biomarkers, where similar multimodal integration challenges exist.

In summary, the study showcases how emerging foundation‑model capabilities—specifically in‑context learning—and AI‑enhanced tensor methods can transform a traditionally cumbersome behavioral neuroscience pipeline into a more efficient, scalable, and discovery‑focused workflow, with implications for a broad range of life‑science domains.