Machine Learning Assignment 1

MobileNet:

It correctly classifies remote control, umbrella, curtain, broom, lipstick, light, teddy bear, backpack, ruler, telephone. It does not classify mirror, glasses, boots, passport, scissors, glue stick, hat, pen. I feel like if the object is very distinctive, modern, and doesn’t have reflection, it might classify. It also doesn’t have to be daily used or common.

The most egregious misclassification: My boots are classified as a vacuum. My hat is classified as a mop.

Being wrong but I know why: I showed the model the inner part of my hat and it classified it as a conch.

Here’s my hat:

 

When I showed the vertical version of the remote control, the model could tell it was a remote control. But when I showed the remote control in a horizontal way, the model classified it as a digital clock. I also tried the umbrella in different states. When the umbrella was closed, it could not recognize it. When the umbrella was open, it could recognize it. Partially obscured, the teddy bear can be classified as an Egyptian cat.

If it is given the picture on the phone, it will always be classified as iPod prioritize than other labels. But the model can still classify the object.

The labels I found: bonnet, band-aid, Windsor tie, wig, lipstick, matchstick, projectile, missile, lighter, punching bag, violin, fiddle, shower curtain, abaya, cloak shower curtain, harp, shopping cart, sax, saxophone, wardrobe, closet, press, refrigerator, icebox, spotlight, spot, milk can, window screen, geyser, window shade, switch, projector, digital clock, suit, bonnet, poke bonnet, sunglasses, shades ski mask, mask, shower cap…… I found a list of labels on the internet. Although I’m not sure if it’s the right one, here’s the link: https://github.com/leferrad/tensorflow-mobilenet/blob/master/imagenet/labels.txt.

I think these labels are more precise than comprehensive. They focus more on the details than to cover more categories. For instance, it recognizes specifically sunglasses instead of glasses. I guess it was trained specifically on these labels because they are easier to be distinguished. In regard to all kinds of animal labels, maybe the model is trained on them because they need to be more finely differentiated by breed, for example, different kinds of cats. 

 

A People’s Guide to Artificial Intelligence:

P11:

machine learning(sorry) / her / chess competition / siri

P14:

1. 1. I identify with existential intelligence the most. Because considering those things that have nothing to do with survival is what makes human life more meaningful.
   2. I would prioritize interpersonal intelligence. Since it’s about communication, better communication creates better understanding and understanding is the foundation of getting the right request.

P20:

1. 1. Handwriting Recognition kind of surprised me since handwriting is a personal and various thing. At first, I thought it was something that only humans could do, but then I realized that humans also recognize handwriting through the accumulation of experience, so maybe the data on handwriting can help the computer recognize it.
   2. In some photoshop apps distinguish the different “object” parts, “background” parts and help you get rid of or change the parts. Also in some albums that can help you categorize images by faces.

P23:

1. 1. mobile phone/ macbook / ipad
   2. When I want to send something to my friends on Wechat, my phone has a list of suggestions of whom to send and they are correct most of the time. It knows who I talk to most and who I’ve been talking to lately.
   3. Email inbox — spam filtering/ conversational systems(helping you to reply)
      Check depositing — facial recognition/ handwriting recognition
      Texting and mobile keyboards — conversational systems(guessing the words you want to use) / handwriting recognition
      Netflix — recommendation engines / machine translation(maybe human sometimes)
      Google — recommendation engines / machine translation / virtual assistance
      Social media platforms — recommendation engines / machine translation / spam filtering(?)
      Automated message systems — conversational systems/spam filtering