Do Infants Smile Like Adults?

Mapping emotional development with deep learning

Project poster — “Becoming: seven signals, five ages, two ways of seeing.” An infant facing a mirror beside red, green, blue and black silhouettes ageing into an adult, on acid yellow.

06
Year — 2025

Role

MSc project — University of Edinburgh
Age-estimation module & cross-age evaluation (team of 3)

Code — GitHub

A smile at nine months and a smile at thirty look similar — but do they mean the same thing to a machine? One ResNet was trained purely on infant expressions, another purely on adult faces, and both were tested across five developmental stages. Their disagreement is the finding: infant expressiveness fades linearly with age, while adult-like expression emerges nonlinearly — crystallising not in adulthood, but in late childhood.

Problem

Facial expressions are an infant’s first language — the channel through which they signal needs, and an early window into developmental health. Yet nearly all emotion-recognition research is built on adult faces, and developmental psychology lacks measurable tools for pinpointing when a child’s expressions become adult-like. Atypical expression patterns can be early indicators of developmental disorders; the question is practical, not just philosophical.

Controlled lab captures of infant expressions from the Tromsø Infant Faces database — Tromsø Infant Faces — expert-validated lab captures

In-the-wild adult faces from the AffectNet dataset — AffectNet — a million faces in the wild

Approach

The networks were treated as instruments of measurement rather than products. Where the infant-trained model succeeds, faces still carry infant features; where the adult-trained model succeeds, adult-like expression has emerged — classification accuracy repurposed as a developmental ruler.

Because AffectNet carries no age labels, a pretrained ResNet-50 age estimator segmented 26,270 images into five developmental brackets. Both classifiers combined cross-entropy with a center-based distance loss, pulling same-emotion features together and pushing different emotions apart. When textbook oversampling collapsed — synthetic interpolated faces are not faces — the fix was a simpler, principled weighted sampling strategy.

The ResNet-50 age estimation pipeline that automatically segmented the dataset into developmental stages — The age-labelling pipeline that made the study possible

The adult-trained model reads teenagers better than adults themselves — adolescent faces may be the most legible of all.

Bar chart showing the infant model’s accuracy declining linearly from infants to adults — The infant model — a clean linear fade, 55.7% to 28.5%

Bar chart showing the adult model peaking on older children and teenagers rather than adults — The adult model — a nonlinear peak in late childhood

Study

Images across five life stages: 26,270
Emotion categories: 7
Peak cross-age accuracy: 77.7%
Infant-model decline: 55.7% → 28.5%
Chance baseline: 14.3%

Findings

Even at its worst, the infant model stayed well above chance on adult faces — a core of expressive features survives the journey from infancy to adulthood. Confusion-matrix analysis surfaced developmental asymmetries: the infant model read surprise near-perfectly in its own domain, but conflated adult happiness with neutrality and sadness. The conclusion for the field: emotion-recognition systems should be age-aware — and deep learning can serve as a measurable complement to observational developmental psychology.

Grid of confusion matrices for both models across every age group — Not just where the models failed — how

In collaboration with

University of Edinburgh

Next case study

07 / 11