Week 3: How AI Works Flashcards
Input layer
How the network receives information from the outside world
Output layer
How you get answers from your network and how the solution can be represented
Explain how you’d build a neural network from an image.
- Begin with pixels
- Use a number to represent the color
- Pixel values are flattened
- Design a neuron corresponding to one of the pixels.
- One neuron acts as the input
What is the number of neurons in an image?
the total number of pixels in an image
(length x width)
How is the output layer classified?
binary classification (i.e. yes/no response)
End-to-end AI model
a model that handles the entire process, from raw data to final output, without any additional components
*this is the type of model for image processing
What is AI based on?
probabilities
The probabilities of all possible outcomes must sum to 1
What would the output be for an image?
The output should represent the probability that the input image contains a target image
You can decide the threshold of yes (e.g. 0.5 probability)
ŷ
probability of yes
same for no
What are the inputs and outputs for a Zillow “Zestimate”?
-inputs: location, number of levels, number of bedrooms, urban proximity, etc. (each input acts as a neuron)
-output: prediction of the price
For inputs that don’t have a numerical value (e.g. style of construction), you can assign an arbitrary number. As long as this stays consistent, the model can still make predictions
What do we want with hidden layers?
For each hidden layer, you want it to gradually move into one output
*Class verification system
-Each neuron corresponding to one student
-Output number is the number of students in class, with each corresponding to one neuron
-Which neuron is activated the most (i.e. strong activation, i.e. high probability) for a particular student
Emotion recognition
-Image as input
-Output for every possible expression
-Build a neuron with every possible expression
-Highest activated neurons corresponds to emotion
How do video games demonstrate that not every AI problem is a classification problem?
-this is a control problem
What are the inputs and outputs for a video game?
-input: the pixels contained in an image of the current game board
-ouput: the action you take
-Each moment, you have one neuron corresponding to one of the actions that you can take at that moment. If one neuron output has really high activation, it takes that action
As the game advances, the model is taking a new input image/frame rapidly
