Fast and Flexible Algorithms for trend filtering
Mathematics on Intelligent Systems- 4
Anirudh Edpuganti - CB.EN.U4AIE20005
Jyothis Viruthi Santosh - CB.EN.U4AIE20025
Onteddu Chaitanya Reddy - CB.EN.U4AIE20045
Pillalamarri Akshaya - CB.EN.U4AIE20049
Pingali Sathvika - CB.EN.U4AIE20050
Team-2
Fast and Flexible Algorithms for trend filtering
Contents
Contents
Contents
Contents
Contents
Contents
Trend Filtering
Trend Filtering
Faces a trade-off between 2 objectives
Trend Filtering
Faces a trade-off between 2 objectives
Minimizing the residual noise
Trend Filtering
Faces a trade-off between 2 objectives
Minimizing the residual noise
Maximizing the smootheness
Trend Filtering
Faces a trade-off between 2 objectives
Minimizing the residual noise
Maximizing the smootheness
Trend Filtering
Faces a trade-off between 2 objectives
Minimizing the residual noise
Maximizing the smootheness
Output points
Trend Filtering
Faces a trade-off between 2 objectives
Minimizing the residual noise
Maximizing the smootheness
Input points
Evenly spaced
Trend Filtering
Faces a trade-off between 2 objectives
Minimizing the residual noise
Maximizing the smootheness
Trend Filter estimate
Trend Filtering
Faces a trade-off between 2 objectives
Minimizing the residual noise
Maximizing the smootheness
Trend Filtering
Faces a trade-off between 2 objectives
Minimizing the residual noise
Maximizing the smootheness
Discrete Difference Operator
Trend Filtering
Faces a trade-off between 2 objectives
Minimizing the residual noise
Maximizing the smootheness
Captures the smoothess between every set of k+2 points
Trend Filtering
Minimizing the residual noise
Maximizing the smootheness
Hold Parallely
Trend Filtering
Minimizing the residual noise
Maximizing the smootheness
Trend Filtering
Minimizing the residual noise
Maximizing the smootheness
Captures the trade-off between the 2 objectives
Trend Filtering
Minimizing the residual noise
Maximizing the smootheness
Captures the trade-off between the 2 objectives
Regularisation parameter
Trend Filtering
So, our trend filtering estimate becomes
Discrete Difference Operator (D)
Discrete Difference Operator (D)
Constant-Order Trend Filtering
Discrete Difference Operator (D)
Constant-Order Trend Filtering
Discrete Difference Operator (D)
Constant-Order Trend Filtering
Observations
Discrete Difference Operator (D)
Constant-Order Trend Filtering
Observations
1D fused lasso problem
Discrete Difference Operator (D)
Linear Trend Filtering
Discrete Difference Operator (D)
Linear Trend Filtering
Discrete Difference Operator (D)
Linear Trend Filtering
linear trend filtering problem
Discrete Difference Operator (D)
Simpler and Better
Discrete Difference Operator (D)
Discrete Difference Operator (D)
where
Discrete Difference Operator (D)
Discrete Difference Operator (D)
where
Discrete Difference Operator (D)
Discrete Difference Operator (D)
Understanding the dimensions
Discrete Difference Operator (D)
Understanding the dimensions
Discrete Difference Operator (D)
Understanding the dimensions
Let k=1
Discrete Difference Operator (D)
Understanding the dimensions
Let k=1
Discrete Difference Operator (D)
Understanding the dimensions
Let k=1
Discrete Difference Operator (D)
Understanding the dimensions
Let k=1
Discrete Difference Operator (D)
Understanding the dimensions
Let k=1
Discrete Difference Operator (D)
Understanding the dimensions
Let k=1
Discrete Difference Operator (D)
Understanding the dimensions
For k
Discrete Difference Operator (D)
Understanding the dimensions
ADMM Algorithm
ADMM Algorithm
Our problem
ADMM Algorithm
Our problem
Augumented Lagrangian
ADMM Algorithm
Augumented Lagrangian
Using the scaled dual variable
Also, let
ADMM Algorithm
Augumented Lagrangian
ADMM Algorithm
Augumented Lagrangian
ADMM Algorithm
Augumented Lagrangian
Manipulation
ADMM Algorithm
Augumented Lagrangian
ADMM Algorithm
Augumented Lagrangian
ADMM Algorithm
Augumented Lagrangian
ADMM Algorithm
Augumented Lagrangian
Backsubstituting everything
ADMM Algorithm
Augumented Lagrangian with
scaled dual variable
ADMM Algorithm
Augumented Lagrangian with
scaled dual variable
ADMM Algorithm
Augumented Lagrangian with
scaled dual variable
ADMM Algorithm
Augumented Lagrangian with
scaled dual variable
ADMM Algorithm
Augumented Lagrangian with
scaled dual variable
ADMM Algorithm
ADMM Algorithm
ADMM Algorithm
Augumented Lagrangian with
scaled dual variable
ADMM Algorithm
This can be rewritten as
ADMM Algorithm
This can be rewritten as
ADMM Algorithm
Augumented Lagrangian with
scaled dual variable
Specialized ADMM
Specialized ADMM
Text
Problem
Specialized ADMM
Text
Problem
Augumented Lagrangian
Specialized ADMM
Text
Augumented Lagrangian
Specialized ADMM
Text
Augumented Lagrangian
Specialized ADMM
Text
WHYYY ???
Why Lasso over Ridge ?
Why Scalable form ?
Why Specialized over Standard ?
Why not (n-1) ?
Discrete Difference Operator (D)
Understanding the dimensions
Let k=1
Discrete Difference Operator (D)
Understanding the dimensions
Let k=1
Discrete Difference Operator (D)
Understanding the dimensions
Let k=1
Why k+2 points ?
Why k+2 points ?