Week 5: Training Neural Networks

Question 1

What is required for training a dataset?

Answer 1

Input (x) and ground truth (y)

Question 2

Explain how input and ground truth is used to train data.

Answer 2

1. Ground truth is provided by humans
2. Provide corrupted input
3. Compares prediction from the neural network (y hat) to the ground truth
4. Based on the prediction and the ground truth, you are given an error
5. Based on the error, the weight and bias ia adjuted to bring it closer to the gorund truth

Question 3

Ground truth

Answer 3

verified, true data used fro training

Question 4

Why do we need a model if you already have the ground truth?

Answer 4

The dataset you input is only a fraction of the amount of data in the real word where you don’t know the ground truth. We want to train the neural network with the data we do have so we can apply it to similar scenarios where we don’t know the answer.

you can’t use neural network training fot one task and apply it to a different task
E.g. you can’t train a network to predict the cost of a house in Los Angeles and apply it to New York

Question 5

Is the ground truth supervised or unsupervised?

Answer 5

Supervised - you label each object in an image

Question 6

Supervised data

Answer 6

annotated data (done by humans)

Question 7

One hot encoding

Answer 7

Turns every label into a unique binary vector

The length is equal to the number of unique categories in the data

Each vector contains a single 1 at the position corresponding to the category

0s everywhere else

Question 8

Explain how one-hot is used for training.

Answer 8

You compare the one-hot ground truth to the predicted activation outputs

If one hot ground truth doesn’t match activation output, there is an error

Question 9

What is a softmax function?

Answer 9

You turn all of the output numbers into a probability

If you have a large activation in the output, this will become the highest probability

Question 10

Why is the dataset squared when making a prediction

Answer 10

You don’t want the error rate to be a negative number

Question 11

What does the loss number give you a prediction of?

Answer 11

an estimation of how your neural network is currently performing

Question 12

How do you calcluate the loss?

Answer 12

1. (Predicted output - ground truth output)^2
2. Add up all of the indivudual training values
3. Multiply by (1/n)

*n = number in training dataset

refer to notes on page 19

Question 13

Explain training loss when you first begin training a neural network

Answer 13

At the beginning, the training loss will be very high

Begins by plugging in random weights, but the loss gradually decreases over time

X represent how many times you change the weight (i.e. steps), Y is training loss (i.e. the error)

Training loss will never reach 0

Question 14

According to Historian Lisa Gitelman…

Answer 14

every field has its own defintion of what qualfies as data

Question 15

Explain the CU Colorado Springs AI contoversy

Answer 15

-Professor installed camera to collect images of students and faculty to train dataset
-Did so without their permission
-Front facing images are too easy - wanted people to be looking away
-State senate banned facial recognition technology in Colorado

Question 16

Explain how the Duke research used facial recognition.

Answer 16

-Duke research used multiple cameras to track the same person across campus
-A number is associated with each person, making it possible to track an individual
-Don’t need to add a neuron for every pearson. Instead, there are two neurons that identify if it is the same person or not – it becomes a yes or no question

Question 17

OpenAI Lawsuit with NYT

Answer 17

Open AI uses articles from NYT to train it, despite IP laws
With AI, what qualifies as fair use - sometimes if it contributes to society it is allowed

Question 18

How is DeepSeek trained?

Answer 18

from data in other AI models

Question 19

Explain CAPTCHA

Answer 19

A turing test used to prove that you are human by clicking on the object it tells you to (Usually cars and traffic lights because they want to collect data to train self-driving cars)

Labeling is expensive but this provides a cheap and efficent way to label

Question 20

What are some ethical issues for collecting data to train AI

Answer 20

1. Consent (e.g. using artists’ work for training)
2. Cheap labor (need cheap labor for supervised learning)
   e.g. OpenAI used Kenyan workers on less than $2 per hour to make chatgpt less toxic

Question 21

Components of ImageNet

Answer 21

One component: not possible without iphone data
Second component: the labelling aspect
(paying students to label was too expensive so posted the job on Amazon platform that enabled people all around the world to sign up- outsourcing labor made it cheaper)

Question 22

Limitation of loss minimization

Answer 22

This model seems good in theory but can cause a lot of problems
Google racist mistake identifying two black people as “gorillas”

Question 23

How does the example of the gorilla miscategorization highlight an issue with training?

Answer 23

The issue with this model is that the errors have the same weight (either 0 or 1)

They treat all miscategorization the same
(E.g treats miscategorization a dog bread and labeling a human as a gorilla in the same way)

Instead, we should change what we place most importance on when evaluating an error (e.g. prioritizing accurately categorizing humans) – will prioritize dropping the error that is most important (e.g. labelling a human as a gorilla)

But model will be less motivated to correct errors of things that are less important

Question 24

Loss meaning

Answer 24

the error - to what degree does the prediction differ from the ground truth

Question 25

How did Google solve the gorilla problem

Answer 25

Google banned the label “gorilla” from the image classifier Another approach might be to train it with more images of gorilla and humans (technical solution) But even with 99.9% accuracy, there are never 100% accurate and there is still a risk of a mistake Google wanted to avoid this mistake again, even with high accuracy

Question 26

What is the challenge with training data and bias

Answer 26

we don't know our own value system and biases Whatever bias we have is amplified by AI The images we see on generative AI are a result of bias and then these images reinforce our bias (it is a cycle) E.g. gender distribution in occupational portraits created by AI generators

Question 27

Two camps for approaching bias in AI

Answer 27

As long as no one is tweaking the data, it is what it is Others want to combat AI reflecting the bias of humans

Question 27

issue with addressing AI bias

Answer 27

Researchers are actively fighting AI bias but sometimes they go too far E.g. getting images of founding father wrong by changing their race Model developed by a few developers who make key decisions, which influence how correcting bias is approached