Assigment 1b - ImageNet and MobileNet

Week2: ImageNet and MobileNet

 Jianhao Ma

ImageNet

My initial impression with ImageNet is about how slow it is. But I thought that makes sense regarding it is a non-profitable research project with a huge dataset. I tried to find how big the dataset is and figured out that the statistics hasn't been updated for 10 years.

Later I realized that not only the search engine lacks consistency, the tree view is also taking forever. 
As I browsed through ImageNet, I started to have a sense that it's one of those early Image Classification project that's been forbidden by time. Can't neglect that there are many privacy concerns, especially under "person" synsets. I think that they would have be more careful when selecting synsets of illegal occupations and minority groups if it's a recent project. But I wouldn't blame that on ImageNet because that's the common theme of fabricating any datasets with all resources on the Internet. It's too much works for handpicking and you cannot rely on algorithms being ethnically appropriate. That would be another essential challenge for Machine Learning: teach them to be ethical. 

MobileNet

I picked 10 photos I shot recently and applied them to the p5 MobileNet image classifier, some of those images are quite abstract. The result is surprisingly good. Six images are classified correctly. Two images are given close guesses that my two year old nephew would make. However, two images are being classified as something that's not correlated at all. Those are:

In summary, objects that are bigger than 1/4 in proportion to the image are better classified; images with better lighting are better classified; and objects that have color contrasts with the backgrounds are better classified. 


Comments

Post a Comment

Popular posts from this blog

Assigment 4

Image
 Image Transfer on P5 Response to Dr. Fiebrink's Speech: There's one part in this speech that intrigues me the most, which is using Machine Learning in creative works does not need to be smart and delicate. It is OK to be dumb and funny, because that's what an unmatured technology genuinely looks like, and it can still be creative. The Edges to Cat example is so fun to play with that I'm kind of addict to it.  Attempt to do Image Transfer on P5 with Pose Regression I watched the videos posted and was most interested in the pose regression example. But the RGB value shift looks not "machine learningly" enough, because it's possible to make some similar project without neural network. Therefore I tried to make one with image transfer. Firstly, I need to load image in data format in P5. I decided to try pixels first, because they retrieve and store image data directly. Then I console logged the "pixels" object in draw, to test out how much the data ...
Read more

Assignment 2A

Image
2A: Excavating AI, Teachable Machine Project Jianhao Ma Moral Machines The reading "Excavating AI" brings up an idea that the tasks of training AI should not just be making them smarter, but to apply political consideration and moral standard as well, especially when related to people. I would not care about such idea years ago, because I thought that technical breakthroughs is the one and foremost priority in emerging technology industries, why bother? However, when the application of machine learning started entering our lives, I changed my mind. Everyday, there are many information processed/filtered by machine learning tools being fed to us, textual or non textual. They could be the advertisement you watch on social media, the pop-up news on your phone. In other words, many of the lower products of machine learning are mediums. Therefore, they have the inherent nature of being political biased and ideologically influential. The fantasy that "technologies should be un...
Read more

A9 - Draw my Drum Machine

Image
My Box Sequencer and Circle Visualization A Dumb GIF illustration: I just followed the steps we learned in class and stylized the box sequencer a little bit by shrinking the size, add strokes, and changing colors. Those were the easy part. Then I recalled that Luisa encouraged me to try making a circular sequencer with arc(x,y,w,h,start,stop). I felt confident about it because I've used arc several times in P5. However, when I actually coded it, I found it's more complicated than I thought.  The first difficulty assigning orders. At first the bigger arc overlapped the smaller ones and hided them underneath. I fixed that through changing the w and h from adding up to decreasing down.  The second challenge was the mousePressed function. It's doing not simple math with mouseX, mouseY, and w anymore. In order to make it work, I have to create a class of arcs and assign the arcs that represent each drum note to objects of that class. It's too much labor to do ...
Read more