A Dynamical Cartography of the Epistemic Diffusion of Artificial Intelligence in Neuroscience

TL;DR

The paper investigates how artificial intelligence knowledge diffuses within neuroscience over five decades using a semantic knowledge cartography based on SPECTER embeddings, 2D UMAP, and HDBSCAN. It analyzes both the spatial embedding of AI concepts and the diffusion of AI‑related publications via a large citation network, employing the radius of gyration and coreness measures to quantify diffusion and integration. The results show that AI is present across neuroscience subfields but remains concentrated in a subset of domains and tends to diffuse within local subspaces rather than across the entire knowledge map, challenging notions of broad conceptual genericity while illustrating broad practical applicability. These findings highlight the need to consider metrology diffusion and social‑epistemic factors when assessing AI's integration into science, with implications for fostering cross‑subfield diffusion and collaborative AI research.

Abstract

Neuroscience and AI have an intertwined history, largely relayed in the literature of both fields. In recent years, due to the engineering orientations of AI research and the monopoly of industry for its large-scale applications, the mutual expansion of neuroscience and AI in fundamental research seems challenged. In this paper, we bring some empirical evidences that, on the contrary, AI and neuroscience are continuing to grow together, but with a pronounced interest in the fields of study related to neurodegenerative diseases since the 1990s. With a temporal knowledge cartography of neuroscience drawn with advanced document embedding techniques, we draw the dynamical shaping of the discipline since the 1970s and identified the conceptual articulation of AI with this particular subfield mentioned before. However, a further analysis of the underlying citation network of the studied corpus shows that the produced AI technologies remain confined in the different subfields and are not transferred from one subfield to another. This invites us to discuss the genericity capability of AI in the context of an intradisciplinary development, especially in the diffusion of its associated metrology.

Figures (17)

• Figure 1: Building step of the dataset provided by Fontaine et al. (fontaine_epistemic_2024-1) and used in this study. The internal citation network is defined by the citation links between the publications within the set $\mathcal{P}$.
• Figure 1: Summary of the robustness checks on the selection of fields of study situated at levels 2 and 3 within the concept network of our dataset $\mathcal{P}$. Top: Topological indicators describing the studied time-aggregated networks. The ratios of the number of nodes and total weights compared to those of the all-levels network $G_{012345}$ are also indicated to evaluate the loss of nodes and edge weight when selecting specific level ranges. Bottom: NetLSD distance matrix between these networks, computed with the method exposed in tsitsulin_netlsd_2018.
• Figure 1: Temporal log return $R_t$ of the citation RoG of the papers published in 1976 (A) and in 1989 (B). The time series are aggregated into clusters $K_i$ in each year with a k-means algorithm, based on an optimal number of clusters identified with a standard elbow method. $N_p$ is the number of articles within the studied cluster $K_i$ and therefore the number of different log-returned RoGs plotted for this cluster. The red curves represent the centroid of the cluster time series.
• Figure 2: Schematic representation of the pipeline used to generate the neuroscience knowledge map below. First, SPECTER converts the textual elements of the articles within our corpus (called $p_i$) into 768-dimensional vectors each ($v_i$), then UMAP transforms the latter into 2-dimensional vectors ($v'_i$) given the data structure provided by SPECTER. The final list of 2D points are used to draw the map.
• Figure 2: Temporal average coreness $\langle\tilde{c}\rangle$ of the AI-related concepts in the cumulative temporal concept networks of each cluster within each networks -- normalized by the highest coreness returned by the $k$-core decomposition of the network at a given year $t$. The higher the coreness $\langle\tilde{c}\rangle$, the closer the AI-related concepts are to the core of the network, and vice versa. We only consider AI-related concepts included in the giant component of each cluster's concept network. The error bars are the standard errors produced by the distribution of corenesses of all AI-related concepts within each cluster at each year, i.e. the ratio of the standard deviation to the square root of the number of entities present in the distribution.