SEXNET: A Neural Network Identifies Sex From Human Faces

Beatrice A. Golomb, David T. Lawrence, Terrence J. SejnowskiNeurIPS 1990

Introduction

Identifying a person's sex from their face is a fundamental biological pattern recognition task. While humans perform this almost instinctively, it has historically vexed traditional, rule-based Artificial Intelligence. The challenge lies in the fact that sex recognition does not follow a simple algorithm. While some features are pathognomonic (distinctive), such as facial hair or specific makeup, humans can successfully identify sex even when these explicit cues are missing or misleading.

To address this, researchers developed SexNet, a pioneering neural network designed to identify sex from low-resolution facial images. By leveraging a connectionist approach—specifically a back-propagation neural network—the authors demonstrated that AI can master this "fuzzy" visual task by learning subtle statistical regularities that are nearly impossible to codify manually.

Methodology: A Two-Stage Architecture

The researchers employed a two-stage neural network architecture to process and classify facial data, foreshadowing modern deep learning techniques:

  1. Image Compression (Feature Extraction): Facial images were sampled at a low resolution of 30x30 pixels (900 total pixels). To reduce dimensionality, these inputs were fed into a 900x40x900 fully-connected back-propagation network. Acting as an early version of an autoencoder, the network was trained to reconstruct the input at the output layer. This bottleneck forced the 40 hidden units to capture only the most essential structural "eigen-features" of the face, effectively filtering out noise.
  2. Classification (SexNet): The activations of the 40 hidden units from the compression stage served as the input for a second back-propagation network. This classifier was trained to produce a binary output: 1 for male and 0 for female.

Dataset and Benchmarking

The study utilized a highly controlled dataset of 90 facial image exemplars. The network was trained and evaluated on these images. To establish a baseline for comparison, human subjects were also asked to identify the sex of the individuals in the exact same 90 images. While small by modern standards, this dataset enabled a direct, head-to-head comparison between machine and human visual processing.

Key Results and Insights

The results of the study yielded several striking insights into both artificial and biological pattern recognition:

  • Surpassing Human Intuition: SexNet achieved an average error rate of 8.1%, notably outperforming human subjects who averaged an 11.6% error rate on the same dataset.
  • Mimicking Human Error: Interestingly, the authors noted that SexNet’s misclassifications often mirrored those made by humans. This suggests the network was not merely finding mathematical shortcuts, but was likely prioritizing the same visual cues and ambiguities as the human visual system.
  • Efficiency of Low-Resolution Data: The system proved that high-level biological categorization could be achieved using very low-resolution imagery. This demonstrated that global facial structure is often more critical than high-resolution detail for sex identification.
  • The Power of Representation Learning: Squeezing the data through 40 hidden units highlighted the importance of dimensionality reduction, reinforcing that connectionist models excel at adapting to modifiable, fuzzy patterns compared to symbolic AI.

Limitations and Future Work

Despite its pioneering success, the study acknowledged several limitations:

  • Dataset Scale: Relying on only 90 images raises questions about how well the model would generalize to a more diverse population with varying ethnicities, ages, and lighting conditions.
  • Multi-modal Cues: Humans naturally rely on supplementary cues like voice, body build, and mannerisms when facial features are ambiguous. The authors suggested that a more robust future system would need to integrate these additional data streams.
  • Algorithmic Transparency: As with many back-propagation models, the network functioned effectively, but how it recognized sex remained difficult to describe in simple rules—an early encounter with the "black box" problem that continues to challenge modern AI.

Summary updated

BibTeX
@inproceedings{golomb1990sexnet,
  author = {Beatrice A. Golomb and David T. Lawrence and Terrence J. Sejnowski},
  title = {{SEXNET: A Neural Network Identifies Sex From Human Faces}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1990},
}
Figure