This paper is a review of recent work using deep learning to study infant brain function with functional MRI (fMRI). The authors summarize how new computational methods can learn patterns directly from images and time series, rather than relying only on preselected measurements. They argue this shift could let researchers map each baby’s brain more reliably, follow developmental changes over time, and link early brain variation to later outcomes.

Infant fMRI is a difficult kind of data. Scans are usually short, babies move a lot, the brain changes rapidly with age, and scan conditions vary across studies and sites. Compared with methods like EEG (electroencephalography) or fNIRS (functional near‑infrared spectroscopy), fMRI gives whole‑brain coverage and access to deep structures, but it also brings high‑dimensional, noisy datasets that are hard to summarize with simple averages or fixed features.

The review describes how different deep learning tools address these problems. Autoencoders and variational models compress complex activity and connectivity patterns into compact “latent” representations that capture individual and developmental differences. Graph neural networks keep the brain’s network structure intact instead of flattening it into lists of numbers. Transformers and attention models can learn which brain regions and time windows matter together. Self‑supervised, contrastive, and generative methods can use unlabeled data to improve learning, and they may help harmonize scans from different protocols.

Why this matters: conventional analyses often focus on group averages or prespecified brain regions, which can hide individual variation. Deep methods aim to recover individualized maps, forecast developmental trajectories, and improve prediction of clinical or cognitive risk. The review highlights progress across input formatting, population and individual mapping, longitudinal forecasting, model evaluation, and attempts to relate model features back to biology.