Let’s start reading this assignment with a song that gives you a slightly drunken feeling:
1. NYU campus ID card & smart phone
link: https://teachablemachine.withgoogle.com/models/76B3LB5kP/
I tried Hubert Lu’s classifier and it turned out to be pretty functional. The confidence value was greater than 90% most of the time. ID card was easier to be put on the right label and hardly ever went wrong.
For the smart phone, I found that with the case it would be more likely to be classified as an ID card, even just showing the front side. Another interesting thing was that when I put my phone on the left side of my body, it was more likely to be recognized as an ID card, possibly because the phones in samples were on the right side.
An extra funny video of my model:
2. perrier bottle & wine bottle
link: https://teachablemachine.withgoogle.com/models/OLl8myGS0/
link: https://editor.p5js.org/AislynnLi/sketches/4ODYKiuii
It was just a joke when I chose these two objects. But I was indeed curious about whether they could be successfully distinguished by model.
I used two kinds of perrier bottles and two kinds of wine bottles to be my samples. At first, I only used a white background, but I found that whenever people and perrier were in the same frame together, it was categorized as wine. I guessed it was because the wine bottle has a richer color on it. So I added the “none” label and a lot of samples with human backgrounds, but that added a lot of variables.
Due to the time limit and all the variables, the result was not that successful(compared to campus ID card & smart phone), but it is still beyond my expectation.
As for the sketch, I wanted to let it visually shows the view of people after drinking wine or mineral water.
- The “perrier” label:
At first, it was just a normal camera view. Then I thought it would be better if there were some bubbles floating up. But if the bubbles were drawn with p5, it would be too inconsistent with the original picture. That’s when I found copy()
which could extract part of the image. I thought that would be the bubbles.
So this is the first version:
With “speed” to be the y position of the bubble, it gets smaller to let the bubble float up. And I made the vision of the bubble part slightly bigger than the original part to create the amplified sensation. The floor(random(0,5))
makes the bubble shake a little bit.
Next, I thought it would be a little weird if there was only one bubble. So I used class so that there could be multiple bubbles.
Here’s the code:
Now there are two more problems. The first one is that it is too slow now, but it is not lagging when there is no model. Therefore, probably the overlapping use of the two makes it too slow. The second is that bubbles are square. I tried to use a mask to deal with it, but I found that it becomes very, very slow. So I give up for now.
- The “wine” label:
Firstly I used three images overlaid and staggered each image a bit. While each image only contains one color, I tried to create a sense of double vision.
Here’s the code:
I used gridSize = (frameCount % 30) + 5
to change the size of the rectangles to create a feeling of eyes out of focus.
Then, I wanted the whole view to shake a little, like a drunken man who can’t stand still.
Here’s the code:
xMove = map(cos(frameCount * 0.4), 0, 1, -10, 30);
yMove = map(sin(frameCount * 0.1), 0, 1, -5, 5);
- The “none” label:
I tried to make an in-between version of the two, like half “wine” and half “perrier”. But then the conversion would be a bit too abrupt. Finally, I made it into pure black, like the feeling of closing your eyes. Or…just to be consistent with “none”, when you have drank none.
3. Reading Response
- Hundreds of AI tools have been built to catch covid. None of them helped:
The first issue is the poor quality of duplicated data from unknown sources. The second issue is the processing of data. Missing the important features, failing to find the errors, and incorporation bias leads to inaccurate results. What’s more, there are unknown implementations that bring a bad reputation to the AI tools.
- Are we witnessing the dawn of post-theory science:
The author believes that there won’t be the end of theory. From the perspective of AI, there’s still prejudice. From the human perspective, humans want to know “why”, humans don’t trust things without knowing why, and we have intuition. But the new way of learning(using big data and getting results without asking why) has gradually stepped into real-life use and has gone beyond what theory-learning can accomplish. “How much predictability are we willing to give up for interpretability?”
First of all, I feel that my own learning mode is to go in the mode of machine learning. In many cases, when I did not learn the basic principles carefully step by step, I would start to do problems or practical exercises, reverse learning and summarize through examples. This feeling is like building up layer after layer of the net, some knowledge just by the net pocket, and some knowledge just fell into the gap of the net. The knowledge that fell into the gaps made me feel very unconfident. So I understand the distrust that people have for AI because they can’t explain why.
And when something I hadn’t learned before came up, I was easily overwhelmed, but those who learned from the theory perspective, those who figured out the “why” step by step, were able to solve these problems by reasoning. But models have much more data than I can have, which means the “net” will be much denser. But even a dense net can encounter new problems. It raises a question I didn’t think through for now: is the inference based on theory that humans have and the prediction of machines the same process?
The current epidemic has also brought some new changes. There are claims that the process of developing our vaccine will change because the virus is mutating so rapidly. Our original theory to countless rounds of experiments to put into use may be too slow to apply. Perhaps we will be forced to use the data-learning-prediction model of AI because it is rapid enough.
But if such a situation does arise, how do people address the issue of prejudice? The scariest thing is that people are unable to detect the existence of bias until there are serious consequences.
I do feel uneasy about dealing with the black box.