Reflect on the relationship between labels and images in a machine learning image classification dataset. Who has the power to label images and how do those labels and machine learning models trained on them impact society?

The reading reveals the fact that the interpretation and classification of images are actually influenced by factors like politics, ideology, prejudices, history etc. In the digital age, Magritte’s example shows that there is always a gap between labels and images. However, machine learning are trying to assume the unstable and non-absolute relation between images and labels can somehow be fixed as “objective”.

As for who has the power to label images, I think there are several stakeholders involved.

• Designers of algorithm: These people decide the underlying logic that machines follow to make interpret and classify images. Tech companies are the direct creators of machine learning systems and they shape the data, inputs, and design of processes that directly affect the model.
• Government：Behind tech companies, political factors may also play a role. Government ideologies can subtly influence how data is collected, labeled, and regulated.
• Cultural and social context: Broader cultural and social values and norms shape both the way images are labeled and how those labels are interpreted.

As for how do those labels and machine learning models impact society:

• These may cause direct inconveniences, such as misidentification by facial recognition systems.
• Moreover, existing biases can be reinforced through machine learning outcomes. Minority groups or those with limited voice may become even more disadvantaged, creating a new layer of injustice.

Coding Exercise

For this assignment, I tried to make a Paper-Scissors-Stone game.

For the image classification part, the three classes correspond to the three hand gestures. At first, I collected about 40 images for each gesture, but during testing I found the accuracy was not good enough. So I added more images with different angles of the hand gestures, and eventually built about 150 images for each class to improve the accuracy of Teachable Machine.

https://drive.google.com/file/d/181CZyJ5HeEThmgGPY4gnJdpEbSx6IBVA/view?usp=sharing

For the coding part, I basically made some adjustment and changes based on the Image Classification example code.

After importing the machine learning model derived from teachable machine, I tried to randomized the computer’s choice and made a comparison with the player’s choice.

At first, the result was always "You're lose," and I thought it was because the judgment was made before the model was completely loaded. So I added a condition to check if the model was completely loaded before making the judgment.

I also noticed that there were some errors. For instance, though it worked well in the teachable machine web page, the model sometimes failed to recognize gestures after being imported into p5js. The model wasn't very stable, so I wanted to add a three-second countdown before capturing the player's image and selected the label. I wondered if this would make it more stable so I created this countdown effect.

Code for countdown: