Does AI see a rabbit in the Moon? Why did humans see a rabbit in the Moon?

SHOJI Daigo / JAXA Lunar and Planetary Exploration Data Analysis Group (JLPEDA)

In Asian countries, there is a culture that a rabbit lives on the Moon. On the other hand, it has been reported that in European cultures, people tend to see a human figure or a human face in the Moon. This work tested what artificial intelligence sees in the Moon. The appearance of the Moon changes with time, season, and the latitude where we observe the Moon. Here, it was assumed that ancient people usually observed the Moon at early evening (~20:00). First, using CLIP*1, an AI model capable of classifying objects based on given classes, the surface patterns of the Moon were analyzed to determine whether the lunar surface was more similar to “rabbit” or “face”. Using several images with different appearance and contrast of the Moon, the lunar surface pattern tended to be classified to “rabbit” as the latitude decreases. This result is consistent with the fact that the oldest texts about the Moon rabbit were edited in India and Southern China, and that of the face on the Moon was written in Europe by the Greek Philosopher Plutarch. As a next step, the lunar images were judged by seven AI models*2 trained for classification of 1,000 types of objects. The probability that these AI models classified the images as “rabbit” was very low. Thus, the lunar surface patterns do not resemble a rabbit. However, in several images, CLIP and ConvNeXt*3 classified the lunar surface as a rabbit with relatively high probabilities, comparable to the top 10 objects out of 1,000. In contrast to AI, humans can communicate and share interpretations with each other. Even if only a few people imagined a rabbit in the Moon at first, this perspective might propagate through their society with communication.

Research Summary

Fig. 1
Fig. 1: The reasons the rabbit and the Moon are related: the similarity between the lunar surface pattern and the rabbit, and the similarity between their behavior (cyclic appearance), have been indicated.

In Asian countries, including Japan, there is a culture that a rabbit lives on the Moon. This culture is very old, and we can find references to the rabbit on the Moon in Indian texts written approximately 2,500 years ago. On the other hand, in European countries, it has been reported that there is a culture to see a human figure or a human face in the Moon. Typically, it is believed that people see a rabbit or a face because the pattern of the lunar mare resembles the shape of a rabbit or a face. On the other hand, in the field of cultural anthropology, it has been suggested that the rabbit and the Moon are related because they share a common behavior. The Moon appears cyclically due to its waxing and waning phases, and rabbits bear descendants frequently (appear cyclically). Thus, both the Moon and the rabbit became symbols of “fertility”. I named the similarity of pattern "static similarity" and the similarity of behavior/function "dynamic similarity" (Fig. 1).

Fig. 2
Fig. 2: The appearance of the Moon in January and July at different time and latitude. These appearances were calculated by the simulation software Stellarium setting the year at 500 BCE (the year when the oldest Indian text about the Moon rabbit was edited). Image credit: NASA/JPL.

If dynamic similarity is the main factor behind this culture, are the lunar surface pattern and the rabbit’s shape not similar? For humans, because of cultural biases, it is difficult to evaluate the similarity between them. In this work, using AI, the similarity between the Moon and the rabbit was analyzed. The appearance of the Moon changes with time, season, and the latitude from which we observe it (Fig. 2). This is because the relative angle between our eyes and the Moon changes due to these factors. Here, considering the lunar patterns observed at different latitudes, it was tested whether the lunar pattern resembles either "rabbit" or "face." For the test, CLIP*1, an AI model capable of classifying objects based on given classes, was used. The appearance of the Moon also changes with time and season (Fig. 2). Here, it was assumed that ancient people frequently observed the Moon in the early evening (8:00 PM). In addition, images from January were used because the area associated with the rabbit’s ears appears to stand vertically during this season at low latitudes. For the AI test, several images with different area of the lunar mare and contrast were prepared (Fig. 3).

As a result, the lunar surface pattern was judged as resembling a rabbit rather than a face as latitude decreases (Fig. 4). This is consistent with the fact that the oldest texts mentioning the Moon rabbit originated in India and Southern China. CLIP tended to focus on the central area of the images (Fig. 4). Interestingly, the magnitude of concentration on the area related to ear of rabbit was not so strong.

Fig. 3
Fig. 3: Lunar images used for the AI tests. To remove the effect of color (to consider only the pattern of the lunar mare), black and white images with different area of the mare and contrast were prepared. These images were rotated to match the appearances from each latitude at 8:00 PM in January as shown in Fig 2.
Fig. 4
Fig. 4: Lunar posture at 500 BCE, 8:00 PM in January, and probabilities judged by CLIP. Colors indicate the attention map, where red areas were more focused when judging the class. The appearance of the Moon was determined by referencing the simulation software Stellarium. The attention maps were generated using the code by Chefer et al. (https://github.com/hila-chefer/Transformer-MM-Explainability/blob/main/CLIP_explainability.ipynb).

As a next test, using the images that CLIP had judged as "rabbit" in the previous test, the seven AI models*2 trained for the classification of 1,000 types of objects judged the lunar patterns. These AI models classified the lunar images as "rabbit" with significantly low probabilities. Thus, the AI models did not regard the surface pattern of the Moon as a rabbit. However, in several images, CLIP and ConvNeXt*3 judged the lunar pattern as a rabbit with relatively high probabilities (Fig. 5). These probabilities correspond to those of the top 10 selected objects out of 1,000 (Fig.5).

Fig. 5
Fig. 5: Lunar images judged by CLIP (left) and ConvNeXT (right) with relatively high probabilities for the “Rabbit”. The top 10 objects out of 1,000, and the probabilities of each object and “Rabbit” are shown in the tables. The 1,000 objects correspond to those of ImageNet-1K*4, which is the image data set for AI classification. The probabilities for the “Rabbit” were calculated as the sum of the three types of rabbits in ImageNet: “Angora,” “hare,” and “wood_rabbit.

Even though state-of-the-art AI models were used, the classification of obscure images, such as the lunar mare, was inconsistent across architectures. The number of ancient people who initially saw a rabbit on the Moon might have been small. However, in contrast to AI, humans can communicate with each other and change their cognition. Even if only a few people originally saw a rabbit on the Moon, this perception could evolve into a cultural belief. Of course, the effects of their behavior and other factors (e.g., the spread of Buddhism) must also be considered. In the future, will AI also be able to classify the world based on the behavior and function of each object (i.e., can AI create symbols)? In the future, when people frequently visit the Moon, what will we see on the Moon? These questions require further discussion.

Terminologies

  • *1 CLIP: An AI model developed by OpenAI in 2021. A characteristic feature of CLIP is its ability to classify objects into categories, which was not used for training.
  • *2 Seven AI models: This work used Resnet 50, ViT, BiT, SWSL, ConvNeXt, Noisy Student, and CLIP. Trained data (weight) of these models are published, and everyone can use them.
  • *3 ConvNeXt: An AI model developed by Liu et al. of Meta in 2022. Based on the conventional model, Resnet, improvements were added resulting in classification with high accuracy.
  • *4 ImageNet: An image and label dataset for object classifications by AI. There are a 1000-objects (1000-classes) version (referred to as ImageNet-1K) and an approximately 21,000-objects version (referred to as ImageNet-21K or ImageNet-22K).

Information

Journal Title AI & Society
Full title of the paper Classification of the lunar surface pattern by AI architectures: does AI see a rabbit in the Moon?
DOI https://doi.org/10.1007/s00146-024-02145-1
Publish date 12 December 2024
Author(s) Daigo Shoji
ISAS or JAXA member(s) among author(s) SHOJI Daigo (JAXA Lunar and Planetary Exploration Data Analysis Group, JLPEDA)

Links

Author

SHOJI Daigo

SHOJI Daigo
Researcher at German Aerospace Center(DLR) and Earth-Life Science Institute, Tokyo Institute of Technology from 2015 to 2020. From 2020, researcher at ISAS/JAXA.