Week 3: AlphaGo and Computations Flashcards
Why is the AlphaGo game so difficult?
-board game is 19x19 so 361 neurons
-state space is 10^172
-game tree size is 10^ 360
How does AI use a neural network for AlphaGo?
-Build a neuron layer that corresponds to each number within the 361 spaces
-Neural network learns the strength of each positon
-Model learn the pattern of the game board, known as the policy network
policy network meaning
You take the input of the image and you get a distribution of all the outputs you should take
same as end-to-end system
AlphaGo model
-a hybrid model
-includes both self-learning and 30 millions games played by professional players
-looks at the game board and the action that a real player actually took to train AI
Explain AlphaZero
-Self Play with zero human knowledge
-AI competes with AI
-Uses the outcome of the game play to train their model (in the beginning they act randomly)
-Based on outcome they update their neural network
Can use alpha zero to play different games
Difference between human and neural network learning
- Deep network has high performance with a lot of experience (takes a lot of hours spent playing the game, but can outperform humans)
- Humans learn new games much more quickly– we require very little “data” while AI requires a lot
How would you design AI to be able to identify the joints?
-input: image
-output: identifying each joint
To design the output layer: assign a number to each joint – the number tells you the x and y position
E.g. Left elbow
X position: 280
Y position: 300
Components of one neuron
- Summation process
- Bias
Summation Process
Sum all the numbers from your input
-This is a weighted summation. the sigma value indicates the weighted sum.
-x represents the pixel value
Weighted Summation
Some input has higher weight and some has lower weight
*Bias
-a constant that gets added at the end of the weighted sum
Why do you need the bias?
-Some neurons get excited too easily and need to calm down a bit
-Others are too passive and need a boost
Who sets the weight and the bias?
the neural network, not humans
What does a fully connected neural network mean?
Every neuron is connected to every neuron
Since H1 and H2 are connected to the same input neurons and therefore always receive the same inputs, does this mean that H1 and H2 will always have the same input summation?
No because you have different weights and biases per neuron
Activation function
Takes the weighted sum and bias and transforms it to produce the neuron’s output, which is then passed to the next neuron
Activation function must be nonlinear
An activation function must be…
nonlinear
-Input (x axis ) will be weighted sum
-Y axis will be the outcome
How does simple classification work
In order to determine which category a new data point belongs to, it makes a classification by drawing a line (on this side of line you belong to this category, on the other side of the line you belong to that category)
However, data is not always this simple - not every problem can be solved for a line (for example, some data is represented as a circle)
How is the problem of messy data addressed?
Instead of changing the model, the hidden layer of neurons can transform the data into something that can be segmented with a line
Cartesian System vs Polar System
-Cartesian uses two perpendicular lines (x and y axis) to plot points while polar system uses a distance from the origin and an angle
-Wouldn’t be able to segment them with the cartesian system
-Neural networks wants to learn the transformation function so that points on the cartesian system can be put into the polar system
Two ways a point can be represented
- Cartesian System
- Polar System
*How can a cartesian point be transformed into a polar point?
Polar X Coordinate
1. identify the two cartesian coordinates
2. square both the x and y coordinates, then add
3. take the square root of this value
Polar Y Coordinate
atan^2(y,x) x (180/pi)
Linear transformation
-Transforming cartesian data into polar is not a linear function
-A neural network learns how to combine these small pieces of action functions to achieve the transformations that works best for a particular task
What are the hidden layers really and what do they do?
They are a giant function that transform the raw data in one coordination into another coordination
What does the ouput layer represent in a yes/no question?
the probability that the input image contains a target object (e.g. face or cup)
