Introduction to Machine Learning

Gianluca Costa
Gianluca Costa
Introduction to
Introduction to
Machine Learning
Machine Learning
including Deep Learning
including Deep Learning
http://gianlucacosta.info/

View Slide

Preface
Preface
●
Machine Learning – and its recent developments in
Deep Learning – is a wide branch of Artificial
Intelligence, with countless applications.
●
This brief introduction tries to concisely describe its basic
concepts – with no claim of completeness.
●
This work was inspired by:
●
the beautiful book Java Deep Learning Essentials by
Yusuke Sugomori, that includes further topics and
details, in addition to mathematical descriptions and
source code
●
Wikipedia

View Slide

AI history in brief – 3 ages
AI history in brief – 3 ages
Late 1950s - Classification based on fixed rules
Breakthroughs: tree algorithms (Depth-First, Breadth-First, ...)
Focus on: efficiency, cutting branches to simplify massive
computations, board games.
Critical problem: frame problem
1980s – Knowledge Representation
Breakthroughs: expert systems, semantic web.
Focus on: effective representations, granularity, finding a way to
conveniently input an immense knowledge base.
Critical problem: symbol grounding problem
Machine Learning (ML) – A machine can learn without being explicitly
programmed – by tuning the parameters of a model, using training data.
Breakthroughs: several methods and algorithms
Focus on: pattern recognition, based on important aspects of the dataset
(features), multi-class and multi-dimensional processing
Critical problem: at first, lack of sufficient amounts of good-quality data
(as the machine cannot evaluate such quality); then, feature engineering.

View Slide

The 3 main problems for AI
The 3 main problems for AI
●
Frame problem: a machine cannot take into account
the countless details of a realistic problem – as they
would require an infinite description and infinite
computation. Machines excel at board games because
the domain is limited and simplified.
●
Symbol grounding problem: even if we were able to
codify all the knowledge, a machine would still
manipulate symbols, not concepts.
●
Feature engineering: a machine can’t individuate
features on its own – they must be provided by the
human operator, keeping the domain in mind and often
after several iterations.

View Slide

Concepts and pandas
Concepts and pandas
PANDA
SIGNIFIED SIGNIFIER
SIGN
●
You could easily understand that the above pictures depict the same
animal even if you didn’t know the term panda – because the human brain
can naturally detect features – therefore, the signified itself.
●
On the other hand, traditional machine learning algorithms just can’t.

View Slide

Feature engineering as the root
Feature engineering as the root
problem
problem
●
If a machine could automatically detect features, the feature
engineering problem would be solved - by definition
●
Furthermore, the symbol grounding problem would be
overcome as well – because identifying features leads to
understanding concepts
●
Finally, a machine able to understand concepts would have no
frame problem
Deep Learning is a set of machine learning methods enabling
machines to automatically detect features.
In other words, via Deep Learning, a machine can discover new
concepts by itself. As an example, Google announced that a deep
neural network was able to learn what a cat is after scanning
millions of cat pictures.

View Slide

Machine Learning at a glance
Machine Learning at a glance
AI
Machine Learning
Neural Networks
Supervised Unsupervised
Support
Vector
Machine
Hidden Markov
Model
Reinforcement
Learning
Logistic
Regression Clustering
Perceptron
Deep Learning
DBN SDA
MLP
CNN

View Slide

Training and Test
Training and Test
LEARNING = TRAINING + TEST
Iterative process
Optimizes the
model parameters
by using the items
in the Training
dataset
Ensures that the model
obtained during the Training
phase is effective on another
dataset (called Test dataset)

View Slide

Supervised and Unsupervised
Supervised and Unsupervised
●
Supervised: the items in the training
dataset are labeled – that is, their
expected output is provided. ML methods
focused on categorization are usually
supervised
●
Unsupervised: items in the training set
are provided with no related output.
Unsupervised ML usually focus on the
structure of the data: one of the most
important examples is clustering

View Slide

Support Vector Machine (SVM)
Support Vector Machine (SVM)
●
Given data items as points in a dimensional space,
let’s define the support vector as the items in a
class closest to items of another class
●
We can then define the decision boundary as the
hyperplane (or set of hyperplanes) that separates
classes so that the sum of distances from each
item in the support vector to the boundary is
maximum
●
Kernel trick: if items cannot be linearly separated,
it is possible to map them to a higher dimensional
space, where linear separation is feasible – but at a
higher computational cost.

View Slide

Logistic Regression
Logistic Regression
●
It is a regression model (in statistical
analysis) where the dependent variable is
categorical – that is, discrete
●
The model parameters are:
– Weight vector
– Bias value
●
Although structurally very different from neural
networks, it actually presents several
similarities
●
Can be easily extended to multi-class
classification

View Slide

Neural Networks (NN)
Neural Networks (NN)
●
Wide variety of methods inspired by the human brain
●
Composed by artificial neurons (named units), grouped in layers and
connected by weighted links
●
Neural networks are usually trained by adjusting link weights via an
iterative process which employs information on output error to tune the
overall network.
●
Except deep learning, the input of traditional neural network method
consists of engineered, manually selected features.
●
Another common trait is the learning rate (η), a parameter describing
how much weights are affected. The learning rate:
– Should not be too small, in order to avoid stagnation in local minimum points
– Should not be too big, to allow the specific algorithm to converge
●
There are different techniques (momentum, ADAGRAD, ADADELTA ...)
to optimize the learning rate, usually involving a gradual reduction of its
value.

View Slide

Perceptron
Perceptron
●
The very first and simplest feed-forward
neural network, consisting of just 1 layer
of links (between input layer and
output layer)
●
Can only perform linear classification
for 2 classes.
●
If the items are linearly separable, the
value of the learning rate is irrelevant, as
in such case the perceptron algorithm
always converges.

View Slide

Multi-Layer Perceptron (MLP)
Multi-Layer Perceptron (MLP)
●
Requirement: extending the initial perceptron so as to
perform non-linear classification – and multi-class as well
●
Key idea: introducing a new layer between input and output
●
Input layer and output layer are called visible layers and
are not directly connected: between them there is a hidden
layer. Units too can be visible and hidden
●
The weights in the hidden layer are often randomly
initialized, to avoid constantly hitting a local maximum
●
Backpropagation: generalizes the perceptron’s tuning
algorithm by calculating the gradient of the error function –
reflecting the output error both on the hidden layer and the
input layer.
●
The output layer is often just logistic regression
●
MLP can approximate any function (both linear and non-
linear) – provided there is a suitable number of hidden units

View Slide

Stochastic Gradient Descent
Stochastic Gradient Descent
(SGD)
(SGD)
●
Whenever the parameters of the model
have to be tuned using the gradient
descent method, computations can
quickly become very intensive
●
When applying SGD, the gradient is
computed on a subset (called minibatch)
of the data
●
Can be applied to different ML methods

View Slide

Hidden Markov Model
Hidden Markov Model
●
Relies on a Markov Stochastic Process,
that is a process satisfying the Markov
Property → the future state only
depends on the present state, with no
direct dependency on previous states
●
Hidden Markov Model is widely employed
to detect sequence patterns - such as in
gene analysis or Natural Language
Processing (NLP)

View Slide

Reinforcement Learning
Reinforcement Learning
●
Unsupervised method where an agent
acts on the given environment, altering it
and receiving a positive/negative
feedback from it
●
Such feedback alters the agent’s
behavior, as it tries to maximize its
performance.

View Slide

Machine Learning – Basic steps
Machine Learning – Basic steps
1.Choose a Machine Learning algorithm
2.Perform feature engineering by selecting the features
(= aspects of raw data that will be used as input) and
possibly converting them to the format required by the
algorithm
3.Split data into a training set and a test set
4.Adjust the model and its parameters in the training
phase, until you get acceptable values of the
effectiveness indicators. You might consider
returning to point 1.
5.Apply the model to the test set, check the effectiveness
indicators and return to point 4 until satisfied.
You might also consider returning to point 1.

View Slide

Overfitting
Overfitting
●
The training dataset and the test dataset must be non-
overlapping – as the purpose of the test is to ensure that
the model is not too tailored on the training set →
overfitting problem
●
Overfitting can have 3 main causes:
– Noise in the training set; it can be due to unclean data, as well
as to exceptional situations
– Employing a training set that does not uniformly represent the
domain, but only a subset
– Ineffective feature engineering
●
To avoid overfitting:
– Increase the number of tests
– Increase the size and variety of the dataset
●
Both conditions can be satisfied by applying techniques such
as k-fold cross-validation

View Slide

Effectiveness Indicators
Effectiveness Indicators
(for binary output)
(for binary output)
Predicted: YES Predicted: NO
Actual: YES True Positive (TP) False Negative (FN)
Actual: NO False Positive (FP) True Negative (TN)
ACCURACY =
TP+TN
TP+TN +FP+FN
=
TRUE PREDICTIONS
TOTAL PREDICTIONS
PRECISION=
TP
TP+FP
=
TRUE POSITIVES
TOTAL POSITIVES
RECALL=
TP
TP+FN
=
TRUE POSITIVES
ACTUAL POSITIVES
●
Effectiveness indicators
evaluate the current model
during both training and test
●
The general formulas of such
metrics can be applied to
models with multi-class
output
2-class
Confusion
Matrix
F
1
=
2∗PRECISION∗RECALL
PRECISION +RECALL

View Slide

Limits of neural networks
Limits of neural networks
●
Perceptron cannot perform non-linear
classifications
●
Multi-layer perceptron can approximate
any function – by adding hidden units, we
can express more patterns; of course, this
requires time and computational
resources, in addition to the risk of
overfitting
●
What about adding further hidden
layers?

View Slide

Vanishing gradient problem
Vanishing gradient problem
●
Most unfortunately, adding hidden layers to a
neural network without a proper strategy
usually degrades effectiveness.
●
Vanishing gradient problem:
backpropagation becomes less and less
effective when moving from the output layer
through the hidden layers up to the input layer
– where it becomes actually ineffective, leaving
the weights almost unaltered. It gets worse:
●
As the number of hidden layers increases
●
As the number of links increases

View Slide

Deep Learning – First Idea
Deep Learning – First Idea
The very heart of the first Deep Learning algorithms is a neural
network with multiple hidden layers and a final output layer
(usually logistic regression). In lieu of standard training, there
are 2 steps:
PRE-TRAINING FINE-TUNING
Individually trains every single hidden layer, starting
from the one receving the network input and so that the
output of a trained layer becomes the input of the following
layer.
Pre-training is usually unsupervised, as it focuses on
patterns: subsequent layers get more fine-grained details.
The actual training method depends on the DL algorithm.
Supervised training of the
network as a whole, or just of
the output layer.
The vanishing gradient
problem is no more an issue,
as weights are almost correct.
Should not use the dataset
used for pre-training.

View Slide

Deep Learning – First algorithms
Deep Learning – First algorithms
●
Deep Belief Network (DBN): a feed-
forward network of Restricted
Boltzmann Machines (RBMs), whose
theory takes into account the energy of
the neural network.
●
Stacked Denoising Autoencoders
(SDA): a feed-forward network of
Denoising Autoencoders (DA), which
learn by
– Encoding: adding noise to the input
– Decoding: restoring the original value

View Slide

Evolution of Deep Learning
Evolution of Deep Learning
●
Pre-training can be skipped without
degrading effectiveness and even
enhancing performances – provided
that we introduce other strategies
●
Valuable ideas:
– Dropout →Making the network randomly
sparse
– Convolutional neural networks (CNN) →
Transform pipeline, very effective when
dealing with multi-dimensional data - such as
2D or 3D images

View Slide

Dropout
Dropout
●
Core idea: if backpropagation is affected by network
density, can we make the network sparse?
●
At every iteration, dropout randomly turns off hidden
units in the network by zeroing the weight of all their
connected links → Easy and computationally fast
●
The process is ruled by a dropout probability –
usually an engineering parameter manually chosen
●
It is similar to the encoding corruption in DAs, but:
– Dropout occurs only in hidden layers, not in visible layers
– In DAs, the very same corrupt input data is used at every
epoch, whereas dropout randomly changes its activation
masks

View Slide

Dropout effectiveness
Dropout effectiveness
●
Dropout works, but usually requires:
– More iterations than pre-training
– A more effective and efficient activation function
in layers: a typical sigmoid is not suggested, firstly
because it saturates at large values
●
Rectified Linear Unit (ReLU): f(x) = max(x, 0)
Activation function much simpler (and faster to
compute) than the sigmoid, preventing saturation
and having a simple derivative.

View Slide

Convolutional Neural Networks
Convolutional Neural Networks
N-Dimensional
Data
1-Dimensional
Features
Traditional
ML
algorithm
Convolutional
layers
Pooling
layers
Layer pipeline
Each layer applies one or more filters
(kernels) to its N-dimensional input.
Each kernel obtains M-dimensional
output (usually, M <= N) called feature
map.
Feature maps become the input for the
following layers.
Convolutional layers actually learn, as
kernel values are automatically tuned by
the algorithm.
Pooling layers behave as passive filters,
performing conversions between
convolutional layers.
Layers can be in any order – although
usually the first is convolutional and the
last is pooling (to flatten data to 1-D)
Backpropagation flows back
into the pipeline, leaving
pooling layers unaltered and
training convolutional layers.

View Slide

More about CNNs
More about CNNs
●
In lieu of fully connected neurons, CNNs are based on
convolution and pooling – which are both able to provide
translation invariance
●
Both their structure and the property of translation invariance
make convolutional neural networks suitable for image
recognition, where they have reached excellent results
●
They require several engineering choices:
– The structure of the pipeline – that is, the type and order of
layers
– Number of kernels per convolutional layer
– Size of each kernel
– Pooling functions (e.g., max-pooling or average-pooling) and
sizes
●
Machine Learning in general – and convolutional networks in
particular - usually benefit from GPU cards

View Slide

Deep Learning today
Deep Learning today
●
Deep Learning is widely adopted in 2
sectors:
– Image recognition/tagging: perhaps one
of the most successful applications, where
Deep Learning has reached better results
than humans
– Natural language processing (NLP): very
active research field → in particular, Deep
Learning in this domain requires dedicated
structures in order to deal with time and
context

View Slide

DL in lieu of traditional ML?
DL in lieu of traditional ML?
●
Silver bullets do not exist – and Deep Learning is no exception
●
Deep Learning can reach brilliant effectiveness; however, it has
drawbacks:
– Despite the considerable research activity, it’s still a very recent field
– It is not suitable for simple problems, as it requires huge amounts of data
– It is also very demanding in terms of computational time and resources
– It requires a considerable amount of trial-and-error iterations in order to
setup its many structural parameters
●
Therefore, Deep Learning usually shines for problems where
features are unknown and massive amounts of data are available.
●
For simple problems, traditional Machine Learning is often a more
suitable choice.

View Slide

A few software libraries
A few software libraries
●
TensorFlow: created by Google, written in C++ with APIs
also for Java and Python. https://www.tensorflow.org/
●
Deep Learning for Java: Open-Source, Distributed, Deep
Learning Library for the JVM. https://deeplearning4j.org/
It employs a fluent, almost declarative API starting from the
method NeuralNetConfiguration.Builder().
It also supports convolutional and shallow networks
●
Caffe: Deep learning framework made with expression,
speed, and modularity in mind.
http://caffe.berkeleyvision.org/

View Slide

Thanks for your attention! ^__^
Thanks for your attention! ^__^

View Slide

Introduction to Machine Learning

Introduction to Machine Learning

Gianluca Costa

More Decks by Gianluca Costa

Other Decks in Technology

Featured

Transcript