What a baby-photo classifier taught me about training data

In 2019, the original author used Google’s Teachable Machine to build a small classifier for “happy” and “unhappy” baby photos. The model was a playful experiment, not a tool for judging a child’s state.

The exercise exposed several machine-learning lessons quickly. A simple interface can make training easy while leaving data quality, bias, privacy, and evaluation to the person building the model.

This update does not reproduce the personal family photos from the original article.

What the experiment did

The author supplied two sets of images, trained a binary classifier, and tested it with new images. Teachable Machine handled the model workflow through a visual interface built around web-based machine learning.

The model appeared to distinguish the two classes within the small sample. That result did not show whether it understood facial expression. It may have learned lighting, camera angle, background, clothing, or features unique to one child.

The first limit was the label

“Happy” and “unhappy” are interpretations, not direct visual facts. A photo captures one moment and may not reveal the person’s actual state.

Before training, define what the label means, who applies it, and where disagreement is expected. If the target cannot be observed reliably, more images do not fix the definition.

The sample was too narrow for a general claim

Training on one child cannot show performance for other children, ages, skin tones, expressions, cameras, or settings. Images from the same phone and room can also make a random test split look stronger than real-world use.

Use a held-out set collected under different conditions. Review errors by relevant group and environment. A classroom demonstration should stay within that limited purpose.

Personal images need a data boundary

Photos of children are sensitive. A responsible experiment states why the images are used, who agreed, where they are processed, how long they are kept, and whether a trained model can be shared.

Use non-sensitive or purpose-made images when the learning goal does not require personal data. Delete source and exported artifacts when the exercise ends unless there is a justified retention need.

Visual tools teach the workflow, not every risk

Teachable Machine is useful for exploring collection, training, testing, and export. More complex work still needs repeatable data preparation, evaluation, versioning, monitoring, and review.

The project is associated with TensorFlow.js, and its community code is available on GitHub. These links are retained as references to the original experiment; confirm current behaviour before building on them.

A good visual demo makes model behaviour visible. Its most useful lesson is how quickly a small dataset can create confidence that the evidence does not support.