0% found this document useful (0 votes)
62 views

Peng Et Al.: Deep Learning and Practice 1

This document introduces machine learning and deep learning. It explains that machine learning involves using patterns in data to build models for tasks like prediction. Deep learning is a type of machine learning where the model learns data representations in a hierarchical way by building up concepts from simpler to more complex. The document provides examples of deep learning models like feedforward neural networks and discusses the history and major developments in deep learning.

Uploaded by

huang
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
62 views

Peng Et Al.: Deep Learning and Practice 1

This document introduces machine learning and deep learning. It explains that machine learning involves using patterns in data to build models for tasks like prediction. Deep learning is a type of machine learning where the model learns data representations in a hierarchical way by building up concepts from simpler to more complex. The document provides examples of deep learning models like feedforward neural networks and discusses the history and major developments in deep learning.

Uploaded by

huang
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 8

'

Peng et al.: Deep Learning and Practice 1


$

Chapter 1

Introduction

& %
Spring 2018
ES

A
1896
CS/NCTU
'
Peng et al.: Deep Learning and Practice 2
$

Machine Learning
• Acquiring knowledge by extracting patterns from raw data
• Example: To predict a person’s wellness t from their MRI scan x by
learning patterns from the medical records {x, t} of some population
t

x Rep. φ(x) Model y - Cost

– x: MRI scan
– φ(x): data representation of MRI scan
– y ∈ (0, 1): model prediction with parameter w
1
y = fw (φ(x)) , σ(wT φ(x)), where σ(s) =
1 + e−s
– t ∈ {0, 1}: ground-truth result associated with input x
& %
Spring 2018
ES

A
1896
CS/NCTU
'
Peng et al.: Deep Learning and Practice 3
$
– Cost: some distance between y and t (e.g. ky − tk22 ), which is to
be minimized w.r.t. w over the {x, t} pairs
• Essentially, we want to find a function fw (φ(x)) to approximate t(x)
• In the present example, fw (φ(x)) bears a probabilistic interpretation of
p(t = 1|x; w)
• The setting here is termed supervised learning as the ground-truth
result t is given for each x

& %
Spring 2018
ES

A
1896
CS/NCTU
'
Peng et al.: Deep Learning and Practice 4
$

Data Representation, φ(x)


• Data representation can critically determine the prediction performance
2
t(x) = 1
t(x) = 1
300 t(x) = 0
t(x) = 0

φ2 (x) = tan−1 x2 /x1


1

200
0
x2

100
−1

0
−2 0 0.5 1 1.5 2
−2 −1 0 1 2 p
x1 φ1 (x) = x21 + x22

raw data domain feature domain


• In classic machine learning, hand-designed features are usually used; for
many tasks, it is however difficult to know what features should be used

& %
Spring 2018
ES

A
1896
CS/NCTU
'
Peng et al.: Deep Learning and Practice 5
$

Deep Learning
• A machine learning approach whose data representation is based on
building up a hierarchy of concepts, with each concept defined through
its relation to simpler concepts
• Using the previous example, this amounts to learning a function of the
following form

fw,θn ,θn−1 ,∙∙∙ ,θ1 (x) = σ(wT φθn (φθn−1 (φθn−2 (∙ ∙ ∙ φθ1 (x)))))
| {z }
Hierarchy of concepts/features

where w, θn , θn−1 , ∙ ∙ ∙ , θ1 are model parameters


• φθ (∙)’s are generally vector-valued functions, e.g. φθ (x) = σ(θx)
• Such a deep model allows to construct a complicated function f (x)
from nested composition of simpler functions φ(∙)’s

& %
Spring 2018
ES

A
1896
CS/NCTU
'
Peng et al.: Deep Learning and Practice 6
$

Example: Feedforward Deep Networks

σ(φθ3 (φθ2 (φθ1 (x))))

φθ3 (φθ2 (φθ1 (x)))

φθ2 (φθ1 (x))

φθ1 (x)

& %
Spring 2018
ES

A
1896
CS/NCTU
'
Peng et al.: Deep Learning and Practice 7
$

Classic Machine Learning vs. Deep Learning

& %
Spring 2018
ES

A
1896
CS/NCTU
'
Peng et al.: Deep Learning and Practice 8
$

History of Deep Learning


• Cybernetics (1940s-1960s): Systems inspired by biological brains
– Perceptron (Rosenblatt, 1958, 1960), Adaptive Linear Element,
ADALINE (Widrow and Hoff, 1960)
• Connectionism (1980s-1990s): Connected simple computational units
– Neocognition (Fukushima, 1980); Recurrent Neural Networks
(Rumelhart et al., 1986); Convolutional Neural Networks (LeCun et
al., 1998); Long Short-Term Memory (Hochreiter and Schmidhuber,
1997)
• Deep Learning (2006s-): Deeper networks and deep generative models
– Deep Belief Networks (Hinton et al., 2006); Deep Boltzmann
Machine (Salakhutdinov et al., 2009); Variational Autoencoder
(Kingma et al., 2014); Generative Adversarial Networks
(Goodfellow et al., 2014)
& %
Spring 2018
ES

A
1896
CS/NCTU

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy