Source-linked AI summary
Understanding and Comparing Deep Neural Networks for Age and Gender Classification
Sebastian Lapuschkin, Alexander Binder, Klaus-Robert Müller, Wojciech Samek
TL;DR
Deep neural networks can classify age and gender from faces, but their predictions offer little explanation of which facial features they use or how preprocessing and model choices affect them. This study compares architectures, preprocessing, and initialization on Adience and uses LRP to visualize prediction strategies. Suitable initialization can compensate for artefactual representations, and combined simple preprocessing achieves state-of-the-art gender recognition on the benchmark.
Problem
Deep neural networks have been used as black boxes, providing little information about which input aspects cause their predictions or how facial features are used.
Source: §1 · p. 2Method
The study compares four architectures, preprocessing and initialization choices on Adience, and applies LRP to visualize how models interact with face-image inputs.
Source: §1 · p. 2, §3.1 · p. 3, §4.1 · p. 4Results
Suitable pretraining can provide robust starting weights that compensate for artefactual representations, and combined simple preprocessing reaches state-of-the-art gender recognition on Adience.
Source: §1 · p. 2, §5.3 · p. 7Takeaways & Limitations
LRP visualizations connect architecture-dependent heatmap structure with the facial regions used for age and gender predictions.
Source: §5.1 · p. 6, §6 · p. 8
Abstract
from arXiv · showhide
Recently, deep neural networks have demonstrated excellent performances in recognizing the age and gender on human face images. However, these models were applied in a black-box manner with no information provided about which facial features are actually used for prediction and how these features depend on image preprocessing, model initialization and architecture choice. We present a study investigating these different effects. In detail, our work compares four popular neural network architectures, studies the effect of pretraining, evaluates the robustness of the considered alignment preprocessings via cross-method test set swapping and intuitively visualizes the model's prediction strategies in given preprocessing conditions using the recent Layer-wise Relevance Propagation (LRP) algorithm. Our evaluations on the challenging Adience benchmark show that suitable parameter initialization leads to a holistic perception of the input, compensating artefactual data representations. With a combination of simple preprocessing steps, we reach state of the art performance in gender recognition.
1. Introduction
The paper addresses the black-box nature of deep neural networks by studying what facial information they use and how preprocessing, initialization, and architecture affect age and gender classification. It compares these choices on Adience and reports that suitable pretraining can compensate for artefactual representations, while simple preprocessing steps achieve state-of-the-art gender recognition.
1. Introduction
- Age and gender recognition on Adience improved substantially in the prior three years, with reported age accuracy rising from 45.1% to 64% and gender accuracy from 77.8% to 91%.Source: §1 · p. 1
- DNNs' popularity is also linked to accessible, generic building blocks and software for training, testing, and deployment.Source: §1 · p. 1
- DNNs had often been used as black boxes, offering little information about which input features drive a prediction.Source: §1 · p. 2
- The study compares four neural-network architectures, initialization choices, and image preprocessing on Adience, and uses LRP to visualize how models use facial features.It also reports a correction to a previously cited gender result and a state-of-the-art gender-recognition result from combined simple preprocessing.Source: §1 · p. 2
- Suitable pretraining can provide robust starting weights that compensate for artefactual data representations, while combined simple preprocessing reaches state-of-the-art gender recognition on Adience.Source: §1 · p. 2
2. Related Work
Related work traces Adience results from conventional feature-based pipelines to neural networks and later pretrained models. The paper identifies architecture, pretraining, and alignment preprocessing as key factors in reported performance and notes concerns about the comparability of some gender results.
2. Related Work
- The initial Adience pipeline reported 45.1% age and 77.8% gender accuracy; a later DNN approach reported 50.7% and 86.6%, respectively.Source: §2 · p. 2
- A 3D landmark-based frontalization step raised gender accuracy to 79.3% while retaining the earlier feature-based classification pipeline.Source: §2 · p. 2
- The reported Adience state of the art was 64% accuracy for age and 91% for gender, with the cited VGG-16 age model pretrained on IMDB-WIKI and ImageNet weights.Source: §2 · p. 2
- The 91% gender result came from a commercial system with reportedly four million private training images, and its account of competing results was selective.Source: §2 · p. 2
- Architecture, pretraining, and optional alignment preprocessing are identified as three major factors behind performance improvements on Adience.Source: §2 · p. 2
- Table 1 surveys Adience age and gender recognition developments, with accuracy values reported in percent.Source: Table 1 · p. 3
3. Architectures, Preprocessing and Model Initialization
The study evaluates four architectures, two provided Adience alignments and their mixture, and available random or pretrained initializations. These choices structure comparisons of model performance and input representations.
3. Architectures, Preprocessing and Model Initialization
- The models were trained in Caffe using code based on configurations for reproducing the earlier AdienceNet results.Source: §3 · p. 3
3.1. Evaluated Models
- The evaluation includes AdienceNet, CaffeNet, GoogleNet, and VGG-16, whose architectures differ in convolutional depth, kernel arrangements, and inception modules.GoogleNet uses parallel convolution/pooling sequences within inception layers and concatenates their outputs along the channel axis.Source: §3.1 · p. 3
3.2. Data Preprocessing
- Training sets use the rotation-aligned images, affine 2D in-plane aligned images, or a mixture containing both alignment versions.The mixed set doubles the number of training samples relative to either individual version.Source: §3.2 · p. 3
3.3. Weight Initialization
- Models are compared from random initialization with models pretrained on ImageNet or IMDB-WIKI when weights are available.AdienceNet is trained from scratch; CaffeNet and GoogleNet are also tested with ImageNet initialization, while VGG-16 uses available ImageNet and IMDB-WIKI weights.Source: §3.3 · pp. 3–4
4. Visualizing Model Perception
The paper uses LRP to visualize model perception and complement its quantitative analysis. LRP decomposes predictions into input-level relevance scores that indicate support for or against a decision.
4. Visualizing Model Perception
- LRP provides qualitative insight into model perception and complements the paper’s quantitative results by visualizing how training conditions interact with the input.The authors select it as a computationally inexpensive alternative to sampling- or occlusion-based explanations that had performed well in a comparison of explanation methods.Source: §4 · p. 4
4.1. Layer-wise Relevance Propagation for DNNs
- LRP decomposes a model’s output for an input into relevance values for input components, iteratively redistributing relevance from output through hidden units toward the inputs.The method assumes the model’s decision function can be represented as a feed-forward graph, and relevance values sum to the output at each layer.Source: §4.1 · p. 4
- Relevance redistribution and pooling satisfy local conservation, so the summed relevance remains conserved across each network layer.The paper states this property for the decomposition and pooling steps.Source: §4.1 · p. 4
- For inner-product classifier layers, the authors use epsilon-decomposition with ϵ = 0.01; for ReLU convolutional layers, they use αβ-decomposition with β = −1.The epsilon term supports numeric stability, while the αβ rule handles activating and inhibiting contributions separately and maintains conservation when α + β = 1.Source: §4.1 · pp. 4–5
- The relevance values are displayed with a zero-centered color map: positive values speak for the prediction, negative values against it, and values near zero indicate little contribution.The authors distribute relevance uniformly over inputs in some lower layers so visualizations remain comparable across models with different filter sizes.Source: §4.1 · p. 5
5. Evaluation and Results
Across the Adience evaluation, results varied with architecture, alignment preprocessing, and initialization. Mixed alignment training and suitable pretraining were associated with stronger performance and robustness across settings.
5. Evaluation and Results
- +1.2% age accuracy, +2.7% age 1-off accuracy, and +1.5% gender accuracy exceeded the corresponding results in the replication.Measurements were taken after a fixed number of training steps.Source: §5 · p. 5
- −1.2% accuracy and −0.8% 1-off accuracy were the remaining differences from for mixed-set GoogleNet.The comparison averaged results across all five data splits.Source: §5 · p. 5
5.1. Remarks on Model Architecture
- Across settings, CaffeNet was outperformed by GoogleNet and VGG-16, while the best VGG-16 models surpassed the best GoogleNet models for gender classification.Age results report both accuracy and 1-off accuracy, which counts an adjacent age group as correct; gender results report accuracy, using ten-crop oversampling.Source: §5.1 · p. 5, §5 · p. 5
5.2. Observations on Preprocessing
- In-plane alignment generally performed worse than rotation alignment; the exception was randomly initialized GoogleNet for age, which lost 1% accuracy but improved 1-off prediction under rotation alignment.The authors attribute the pattern to child faces being distorted or misaligned when adult-suitable reference frames are used.Source: §5.2 · p. 6, Figure 3 · p. 8
- Training on mixed rotation- and in-plane-aligned images benefited all models and yielded good performance on test sets from both preprocessing methods.The combined training set effectively doubled the training data and supported a more robust feature set.Source: §5.2 · pp. 6–7
- Cross-method test-set swapping exposed poorer generalization for models trained on in-plane-aligned images than for models trained on rotation-aligned images.Some models tested with the other alignment method performed worse than the early SVM-based models from.Source: §5.2 · p. 7
5.3. Observations on Initialization
- ImageNet fine-tuning improved GoogleNet in both tasks and all dataset configurations, whereas CaffeNet slightly declined on age prediction but improved on gender prediction.GoogleNet's ImageNet top-5 error was 6.6%, compared with 19.6% for CaffeNet.Source: §5.3 · p. 7
- Suitable initialization produced better converged models and faster early learning, while ImageNet-pretrained GoogleNet used an ensemble of facial features rather than isolated class-associated cues.The random-initialized model's isolated cues included eyebrows and lips for female faces and nose, chin, and uncovered ears for male faces.Source: §5.3 · p. 7, Figure 4 · p. 8
6. Conclusion
The paper uses Layer-wise Relevance Propagation to examine facial cues in age and gender prediction, and finds that pretraining and simple preprocessing choices affect the models’ behavior and performance.
6. Conclusion
- On Adience, the study compares preprocessing, initialization, and architecture choices and reports state-of-the-art gender-classification performance using simple preprocessing.The conclusion also reports that appropriate pretraining counteracts overfitting and supports holistic input perception.Source: §6 · p. 8
- Layer-wise Relevance Propagation visualizations are used to investigate which facial features models rely on for age and gender prediction.Source: §6 · p. 8
- Appropriate model initialization via pretraining counteracts overfitting and leads to a more holistic perception of the input.Source: §6 · p. 8