The image dataset on ImageNet was quite interesting to explore. I found it intriguing how some animal categories, such as dog breeds, had about 30 different images, while others like cats only had 3-4 breeds represented. Some animals, like the axolotl, had just a single image. This disparity might reflect an average human's knowledge—people are more likely to be familiar with various dog breeds than with an axolotl. I also noticed differences in image composition: some, like the cockroach, appeared with other objects, while others, like the hartebeest, included multiple subjects in one image. These variations could affect image classification accuracy. Perhaps the dataset would benefit from more consistent image resolution, size, and neutral backgrounds.

For the image classification p5 code examples, I noticed significant differences in the model's recognition abilities. Animals were generally recognized accurately. However, it struggled with animated characters—for instance, it identified Po the Panda as a "tin can opener." The model also had difficulty with religious and cultural imagery. It misclassified a picture of women from India as a "cobra," a Greek statue as a "footstand," and the American flag as a "window shade." I believe the image size affected the model's accuracy and confidence score. Interestingly, it couldn't accurately recognize AI-generated images either. I found it particularly surprising that an AI-generated picture of a cat in Hogwarts attire was identified as a "clock" rather than a cat, despite the model's ability to recognize most cat breeds. Some of its classifications were completely wrong, such as the model recognizing Earth as “dough”.

The emoji scavenger hunt was quite enjoyable too. It recognized some emojis much quicker than others, like the laptop. When pointing it at unrelated objects, it accurately figured out what it saw. However, for some items, I had to point repeatedly for it to recognize them. I was also able to trick it by using my hand for some prompts.