ducnh1022

What is MDD
make any kind of change to a model as well as to the code generated from that model
= round-trip engineering = fordward engineering + reserver engineering

1. Increased likelihood of scope creep due to ease of change

easier to extend function
an existing meta-model may contain more functionality than specified in the requirements

=> impact supplier & client => confused to identify original requirements & change request

ducnh1022

1 what is antenna

way of converting guided waves in waveguide, feeder cable or transmission line into radiating waves

art of antenna design is to ensure this process efficiently as possible

2 conditions for radiation

charges + not uniform motion = reversing direction = direction changing or oscillating in periodic motion

ducnh1022

Architecture
fundamental 
define guideline
communicate with stakeholder
cross-cutting concern
manage uncertainty
conceptual intergrity

what kind of storage, how modules interact, where recovery system

ducnh1022

Purpose
Data -> classifier -> intelligence

Input x -> sentence sentiment classifier -> y (positive or negative)

App:
Spam filtering
image classification

Impact of classification

ducnh1022

find best classifier => maximize likelihood over all possible w0 w1 w2

Data likelihood
quality metric -> probability of data

learn logistic regression model with maximum likelihood estimation MLE

finding best linear classifier with gradient ascent

convergence criteria, optimize when dl/dw = 0 but in practice we stop when dl/dw < epsilon

ducnh1022

step 1 start with empty tree
step 2 select a feature to split data
FOr each split of tree
  step3 if nothing more to, make predictions
  step4 ohter wise go to step 2 and continue recurse on this split
  
Feature split learning = decision stump learning

what better: split on credit or term

ducnh1022

Learning simpler decision trees

early stopping -> limit depth, use classification error on limit depth of tree, if very few data points

Pruning: simplify tree after learning algorithm terminates
cost function

step 1: consider a split
step 2: computer total cost C(t) of split

ducnh1022

combine multiple simple classifier -> ensemble classifiers

y hat = sign(f(x))

Adaboost

start same weight for all points alpha = 1/N

For t = 1..T
learn f(t) with data weight alpha

ducnh1022

Logític clasìication -> stochastic ooptimization 
-> data and parameter tuning -> deep networks  
-> regularization -> convolutional networks 
-> embeddings -> recurrent models

deep learning apply in all field, reasearcher, engineer, data scientist

ducnh1022

Precision: fraction of positive predictions that are actually positive
recall: fraction of positive data predicted to be positive

optimistic = low precision high recall
pesstimistic = high precision low recall

trade off