RESUME KULIAH TAMU

[Resume Kuliah Tamu]

Hari/Tanggal : Rabu/28 Februari 2017
Judul              : Form Perceptron to Deep Neural Network
Pembicara     : Adi Chris Bangun (Machine Learning Engineer) di-Traveloka
Topiknya 

 Bagaimana machine learning itu bekerja?
-Really on the past
-Masa depan itu sama dengan apa yang terjadi sebelumnya.
-Mengdiscover rules pada suatu program

 Can We Should Use Machine Learning?
1.Recognizing hotdog or not hot dog from a given picture.  (Yes and we should)
2.Classifying an integer as primer or not.  (yes, but we sholudn't)
3 .Predict whether you'll like Black Panther movie, given you've watched captain america. (yes and we should)
4. Predict number of sales by the end of this year, given some oberved data. (yes we should)
5.Computing the shortest path between Jakarta and Surabaya. (no, we cant)

The essence of Machine Learning!
1.Paling utama harus punya datanya (Data)
2. Harus ada pola(pattern)
3.Gaada permasalah MTK yang  langsung solved (no Formulaic Expression exist to solve the pattern)
Problem kita punya data dan polanya juga gabisa langsung menyelesaikan atau Solved suatu permasalahan tersebut.

So, What is Machine Learning?
Manual : Shopkeeper decides what music to recommend for every customer.
Automation : Computer decides what music to recommend for every customer, using rules created by the shopkeeper.
Machine

Components of Learning
Example :
Input dan ouputnya apa
input : × (Customer Apllication)
output : y (Good/Bad Customer)
Target function : ideal function , ideal credit approval formula
Data -> historical records
-Machine Learning Framework
>>Theres is no right answer -> tidak pernah tau masa depan gimana.  Semakin banyak masa lalu yg dikenang semakin banyak masa depan yg ingin diproses.
>> intinya harus punya data yang cukup besar dan juga algoritma yang di pakai.
<<<<contoh ML pada traveloka
-klasifikasi gambar
-Rangking flight
Kunci : yang penting ada domain yang benar-benar kita pahami.

What is Deep Learning?
Neural net with Deeper Layers
>>>> kenapa sekarang org pake deep learning?
tipikal problemnya Di ML ada dua yaitu: linear problem dan non linear problem.
Non linear problem bisa dibuat garis lurus kenapa?  in non linear problem, we still draw a straight line.  However, we draw it in the new representation of our data.
A concrete example :
XOR Probelm
>>>>HOW?
Neural Network :
>>> general idea :
Perceptron : Algortima elegan , simple dan merupakan pondasi bagi algoritma lain.
<<<< to learn the values of the weights w that are then multiplied with the input features in order to make a decision whether a neuron fires or not.
>>>>>A very important component : activation function<<<<<
perceptron tidak pakai activation function.
+++++Heaviside Step Function+++++
>>Activation Function : Sigmoid
How does it Work - Neural Network Prediction
>>>predict = Forward Propagation
>>>Training Neural Network
))Training = Minimize Loss! (untuk mengubah value)
>>>Training Neural Network = Gradient Descent
>>> Backpropagation
+ Forward propagation : we get the prediction of our model
+ we then compute the L,  loss,  between our prediction and the true label
+ to calculate gradien,
Deep learning

Why Deep Learning?
>>Deeper networks proven to be more expressive and can extact more rebust feature
>>Compared to other classification model, such SVM,  with DL,  we dont do (much)  feature extraction.
>>The availability of large amounts of labelled data as well as GPUs which can process this data in parallel at high speeds enables these models to be much faster than previous methods.
Ga cum sekedar high dan juga banyak di pakai perusahaan-perusahaan besar.
>>>Current DL Architecture

What I wish I took More seriously
+ Linear Algebra -> Seriously guys, take this course seriously
+Statistic -> Doing Machine Learning without Statistic is like using library stack in c++ without knowing why it works the it is
+Probability, probability and probability -> all that we have now in Machine Learning is based on the assumption under so many uncertainties.