Artificial vs Human Intelligence: Age Group Classification from Ambiguous Images of Young People

Introduction

Predicting a person’s age is complicated by human perception and cultural cognition complexities. In this era of technological advancement, we find ourselves at the crossroads of Artificial Intelligence (AI) and Human Intelligence (HI), particularly in their abilities to classify abstract and ambiguous visual content.

What makes our project remarkable compared to regular age classification tasks is the choice of ambiguous images of young people. And what makes the images ambiguous are the elements that make predicting a person’s age harder. For example, cigarettes in the mouth or sunglasses. These features make the face classification by age group more challenging, and such ambiguous images are what we call “noisy face data”. By classifying ambiguous images of young individuals into age groups, this project aims to refine the nuances in age group classification through the collaboration of artificial intelligence and human intelligence.

Background

The previous study on classifying the ambiguous images of events by humans and AI served as a foundation for our research idea [1]. The former experiment included classifying ambiguous scenes on event types, while we are analysing how AI and HI will succeed in age group classification based on ambiguous images of faces. In our work, we will explore that this “noise” in noisy face data is perhaps meaningful for humans, as it influences how the images are classified differently from how it influences AI.

The authors of “Notes on the Semiotics of Face Recognition” translate a quote from Italian semiotician Patrizia Magli: “The perception of a face is primarily based on its recognition. To recognize a face means to assign it to a class; and, in this respect, the perceptive experience seems already directed. by the cultural filter: the filter of culture lies in between the face that is in front of us and our irresistible vocation to interpret” [2].

Humans have glasses of cultural cognition through which they perceive the world. Unlike computers that calculate similarity patterns, humans attach meanings to certain cultural symbols. In this way, the faces are being analysed too. Certain features in appearance play a significant role in human age recognition from faces. The qualities of being bold, having grey hair, no teeth, eyebrows, beard, face tattoos, natural and neutral makeup or no makeup at all, etc. are more often associated with a person being older. On the other hand, face paint, bright colours in makeup and hair, fun facial expressions, unusual hairstyles, no facial hair, body modifications, etc. are usually represented as a sign of youth.

Data Collection

The project’s dataset, which we named the “Noisy face dataset” consists of ambiguous images of young people, was gathered manually from the internet [3]. Studies show that ‘’faces from a social in-group, such as people of a similar age, receive more in-depth processing and are processed holistically’’ [4]. Therefore, to draw more attention, interest, and precision in the classification task of the ambiguous images of people, we have picked the age range that included images of people from 15 to 30 years old. The age range is divided into 5 classes: 16–18, 19–21, 22–24, 25–27, and 28–30 years. Our gathered dataset is a collection of 250 diverse images of young people, each falling into one of five age classes. We’ve designed a two-fold experimental approach to bring this exploration to life.

Experiment Part 1: Exploring Natural Intelligence

The first part of our adventure puts human intelligence to the test. We’re keen to see how well people can navigate through a sea of abstract and not-so-clear visual information. The challenge is to accurately place images of individuals into the correct age categories. We’ve selected 30 participants through a simple yet effective convenience sampling method for the survey. These participants will embark on their classification quest using a Google Form, including a selection of 50 test images as shown in Figure 1.

Figure 1. There are 50 test images of dataset, 10 images from each class.

Experiment Part 2: Exploring Artificial Intelligence

The second chapter of our exploration evaluates how well AI can handle the same classification task. To conduct this cutting-edge experiment, we’ve chosen a state-of-the-art AI model, such as Keras pre-trained ResNet50, known for its expertise in image classification [5]. The challenge for our AI part is similar to that faced by our human participants. It will classify the same 50 test images, ensuring a fair and direct comparison between human and machine intelligence. However, there’s a twist in the AI’s approach. While our human classifiers relied on their natural intelligence and perception, the AI will be trained on 200 images.

Get Ali Zeynalli’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Experimental Results

Part 1: Testing Natural Intelligence

Figure 2. The confusion matrix of results of the survey in experiment Part 1.

Figure 3. The classification report of results of the survey in experiment Part 1.

According to the Figure 2 and Figure 3, the results for human classification of age categories of ambiguous images show varying precision, recall, and F1-score across different age groups:

• 16–18: This category has the highest precision and F1-score, indicating that participants were better at correctly identifying individuals in this age group than others.
• 19–21 and 22–24: These categories have relatively low precision and recall, suggesting participants found it challenging to accurately distinguish between these age ranges.
• 25–27: Similar to the previous two age groups, precision and recall are low, with a slightly higher recall, indicating participants were slightly more successful at identifying all relevant instances of this age group.
• 28–30: The precision and recall are notably better than for ages 19–27 but still not as high as for the youngest group.

The overall accuracy is 0.36, which suggests a generally moderate performance by human participants, with some difficulty in classifying ambiguous images across the age spectrum.

Part 2: Testing Artificial Intelligence

Figure 4. The confusion matrix of test images on the best model, ResNet50 in experiment Part 2.

Figure 5. The classification report of test images on the best model, ResNet50 in experiment Part 2.

According to the Figure 4 and Figure 5, the results for AI model classification of age categories show varying precision, recall, and F1-score across different age groups:

• 16–18: The AI struggled the most with the youngest age group, showing the lowest precision and F1-score.
• 19–21 and 22–24: The AI performed slightly better in these age groups than the youngest, with the 22–24 group showing the highest recall.
• 22–24 and 28–30: The AI demonstrated the highest recall in these age groups, indicating it was more successful in identifying all relevant instances in the dataset for these ages. However, precision is not as high, suggesting more false positives were in the classifications.
• 25–27: The AI had the lowest recall for this age group, indicating it missed many individuals who should have been classified into this category. This suggests the model struggled to identify distinctive features for this age range.

The overall accuracy for the AI model is 0.28, which is lower than that of the human participants, indicating a relatively uniform performance across the different age groups but at a lower level than human classifiers.

Evaluation

Human Strengths: Humans showed a more vital ability to classify the youngest age group (16–18) correctly, potentially due to more apparent distinguishing features that are easier for humans to recognize or the artefacts or cultural symbols on images that indicated belonging to the 16–18 age group more strongly.

AI Weaknesses: AI had particular difficulty with the youngest and the 25–27 age groups, possibly due to less defined age-specific features that the model could learn from or perhaps because of the quality and variety of training data. One of the limitations of this experiment is the size of the training data. Due to the lack of time, we could gather only 250 images, 200 of which were used for training, but it is not enough to train a good AI model.

Overall Performance: Humans outperformed AI with an overall accuracy of 0.36 compared to 0.28 for AI. This could be attributed to the human ability to perceive and interpret subtle nuances in facial elements and meaningful cultural symbols that an AI model like ResNet50 might not capture without extensive training data. While humans and AI struggled with classifying ambiguous images into specific age categories, humans showed a slightly better ability to interpret such images. This might suggest that, while powerful, current AI models still lack the nuanced understanding that comes naturally to humans. One of such natural abilities is to engage in semiotic analysis. While a human can interpret an image of a human with a fun posture, facial expression and a cigarette as a sign of being young and funky, the AI would treat these features as a noise that disturbs the proper face analysis. Even though the AI’s performance could improve with more training data, better feature extraction techniques, and more advanced modelling techniques, we expect that, unlike dealing with standard face images, the ambiguous images of faces will remain a challenge.

Conclusion

In our exploration, we have conducted surveys and experimented with the pre-trained AI model, ResNet50, to classify ambiguous images of young people into age classes. The results indicate an average accuracy of approximately 36% for surveyed individuals, while the best-performing AI model achieves an average accuracy of around 28%. These findings begin a nuanced understanding of the interplay between human and artificial intelligence in the age classification of ambiguous face images, opening avenues for further investigation and refinement of human and AI capabilities in this domain.

References

1. Kate Sanders et al., “Ambiguous Images with Human Judgments for Robust Visual Event…,” OpenReview, May 23, 2023, https://openreview.net/forum?id=6Hl7XoPNAVX.
2. Remo Gramigna and Cristina Voto, “Notes on the Semiotics of Face Recognition,” Sign Systems Studies 49, no. 3–4 (2021): 338–60, https://doi.org/10.12697/sss.2021.49.3-4.05, page 349.
3. “Ages: Famous Birthdays.” AllFamous.org. Accessed January 7, 2024. https://allfamous.org/age.
4. Michael E. Nicholls, Owen Churches, and Tobias Loetscher, “Perception of an Ambiguous Figure Is Affected by Own-Age Social Biases,” Scientific Reports 8, no. 1 (2018), https://doi.org/10.1038/s41598-018-31129-7, page 1.
5. “Image Classification Using ResNet-50 Network.” Accessed January 7, 2024.https://www.mindspore.cn/tutorial/training/en/r1.1/advanced_use/cv_resnet50.html

Contributors: Ali Zeynalli, Altay Garayev, Palina Y.