Making Machine Learning Easy with H2O and WebFlux

1
Making Machine Learning Easy
Deep Learning with H2O and Spring WebFlux
Licensed under CC BY 3.0 license.

2
Disclaimer
All information presented in this document and all supplementary materials
and programming code represent only my personal opinion and current
understanding and has not received any endorsement or approval by IPT -
Intellectual Products and Technologies or any third party. It should not be
taken as any kind of advice, and should not be used for making any kind of
decisions with potential commercial impact.
The information and code presented may be incorrect or incomplete. It is
provided "as is", without warranty of any kind, express or implied, including but
not limited to the warranties of merchantability, fitness for a particular purpose
and non-infringement. In no event shall the author or copyright holders be
liable for any claim, damages or other liability, whether in an action of
contract, tort or otherwise, arising from, out of or in connection with the
information, materials or code presented or the use or other dealings with this
information or programming code.

3
3
Trayan Iliev
– CEO of IPT – Intellectual Products & Technologies
– Oracle® certified programmer 15+ Y
– End-to-end reactive fullstack apps with Go, Python, Java, ES6/7,
TypeScript, Angular, React and Vue.js
– 12+ years IT trainer
– Voxxed Days, jPrime, Java2Days, jProfessionals, BGOUG, BGJUG, DEV.BG
speaker
– Lecturer @ Sofia University – courses: Fullstack React, Multimedia with
Angular & TypeScript, Spring 5 Reactive, Internet of Things (with SAP),
Multiagent Systems and Social Robotics, Practical Robotics & Smart Things
– Robotics / smart-things/ IoT enthusiast, RoboLearn hackathons organizer
About Me

4
Where to Find The Code and Materials?
https://github.com/iproduct/course-ml

5
Machine Learning + Big Data in Real Time +
Cloud Technologies
=> The Future of Intelligent Systems

6
6
Asynchronous Event Streams
in Real-time
6

7
Need for Speed :)
The Go gopher was designed by Renee French. (http://reneefrench.blogspot.com/) The design is licensed under the Creative Commons 3.0 Attributions license.

8
Batch Processing
Extract Transform Load Extract
Transform
Load Transform
Transform

9
DB
Synchronous
A
A
B
B
DB
Asynchronous
A
B
C
D
A
B
C
D
Synchronous vs. Asynchronous IO

10
10
http://www.reactivemanifesto.org
Reactive Manifesto

11
• Message Driven – asynchronous message-passing allows to
establish a boundary between components that ensures loose
coupling, isolation, location transparency, and provides the means
to delegate errors as messages [Reactive Manifesto].
• The main idea is to separate concurrent producer and consumer
workers by using message queues.
• Message queues can be unbounded or bounded (limited max
number of messages)
• Unbounded message queues can present memory allocation
problem in case the producers outrun the consumers for a long
period → OutOfMemoryError
Scalable, Massively Concurrent

12
“Conceptually, a stream is a (potentially never-ending) flow of data
records, and a transformation is an operation that takes one or more
streams as input, and produces one or more output streams as a result.”
Apache Flink: Dataflow Programming Model
Data / Event / Message Streams

13
The idea of abstracting logic from execution is hardly new -- it was the
dream of SOA. And the recent emergence of microservices and
containers shows that the dream still lives on.
For developers, the question is whether they want to learn yet one more
layer of abstraction to their coding. On one hand, there's the elusive
promise of a common API to streaming engines that in theory should let
you mix and match, or swap in and swap out.
Tony Baer (Ovum) @ ZDNet - Apache Beam and Spark:
New coopetition for squashing the Lambda Architecture?
Data Stream Programming

14
Direct Acyclic Graphs - DAG

15
Event Sourcing – Events vs. Sate (Snapshots)

16
https://commons.wikimedia.org/w/index.php?curid=34963987, By Textractor - Own work, CC BY-SA 4
Query = λ (Complete data) = λ (live streaming data) * λ (Stored data)
Lambda Architecture - I

17
https://commons.wikimedia.org/w/index.php?curid=34963986, By Textractor - Own work, CC BY-SA 4
Query = λ (Complete data) = λ (live streaming data) * λ (Stored data)
Lambda Architecture - II

18
Apache
ZooKeeper
MySQL /
PostgreSQL
HDFS /
Amazon S3
Lambda Architecture - Druid Distributed Data Store
https://commons.wikimedia.org/w/index.php?curid=33899448 By Fangjin Yang, GFDL

19
Kappa Architecture
Query = K (New Data) = K (Live streaming data)
• Proposed by Jay Kreps in 2014
• Real-time processing of distinct events
• Drawbacks of Lambda architecture:
−It can result in coding overhead due to comprehensive processing
−Re-processes every batch cycle which may not be always beneficial
−Lambda architecture modeled data can be difficult to migrate
• Canonical data store in a Kappa Architecture system is an append-only
immutable log (like Kafka, Pulsar)

20
Kappa Architecture II
Query = K (New Data) = K (Live streaming data)
• Multiple data events or queries are logged in a
queue to be catered against a distributed file
system storage or history.
• The order of the events and queries is not
predetermined. Stream processing platforms
can interact with database at any time.
• It is resilient and highly available as handling
terabytes of storage is required for each node
of the system to support replication.
• Machine learning is done on the real time basis
Data Lake
Input
Data
Streaming
Layer
Serving
Layer

21
Zeta Architecture
• Main characteristics of Zeta architecture:
− file system (HDFS, S3, GoogleFS),
− realtime data storage (HBase, Spanner, BigTable),
− modular processing model and platform (MapReduce, Spark, Drill, BigQuery),
− containerization and deployment (cgroups, Docker, Kubernetes),
− Software solution architecture (serverless computing – e.g. Amazon Lambda)
• Recommender systems and machine learning
• Business applications and dynamic global resource management
(Mesos + Myriad, YARN, Diego, Borg).

22
• Apache Spark is an open-source
cluster-computing framework. Spark
Streaming, Spark Mllib
• Apache Storm is a distributed
stream processing – streams DAG
• Apache Samza is a distributed real-
time stream processing framework.
Distributed Stream Processing – Apache Projects:

23
• Apache Flink - open source stream
processing framework – Java, Scala
• Apache Kafka - open-source stream
processing (Kafka Streams), real-
time, low-latency, high-throughput,
massively scalable pub/sub
• Apache Beam – unified batch and
streaming, portable, extensible
Distributed Stream Processing – Apache Projects II

24
24
Example: Internet of Things (IoT)
CC BY 2.0, Source: https://www.flickr.com/photos/wilgengebroed/8249565455/ ,
Radar, GPS, lidar for navigation and obstacle avoidance ( 2007 DARPA Urban Challenge )

26
26
• Raspberry Pi 2 (quad-core ARMv7 @
900MHz) + Arduino Leonardo clone A-Star
32U4 Micro
• Optical encoders (custom), IR optical array,
3D accelerometers, gyros, and compass
MinIMU-9 v2
• IPTPI is programmed in Python, Java and
Go using: Wiring Pi, Numpy, Pandas,
Scikit-learn, Pi4J, Reactor, RxJava,
GPIO(Go)
IPTPI: Raspberry Pi + Ardunio Robot

27
IPTPI: Raspberry Pi + Ardunio Robot
27
3D accelerometers, gyros,
and compass MinIMU-9 v2
Pololu DRV8835
Dual Motor Driver
for Raspberry Pi
Arduino Leonardo clone
A-Star 32U4 Micro
USB Stereo
Speakers - 5V
LiPo Powebank
15000 mAh

29
Encoder
Readings
ArduinoData
Fluxion
Arduino
SerialData
Position
Fluxion
Robot
Positions
Command
Movement
Subscriber
RobotWSService
(using Reactor)
Angular 2 /
TypeScript
MovementCommands
IPTPI Reactive Streams

31
31
Functional Reactive Programming
31

32
32
We live in a Connected Universe
... there is hypothesis that all the things in
the Universe are intimately connected, and
you can not change a bit without changing
all.
Action – Reaction principle is the
essence of how Universe behaves.
Imperative and Reactive

33
Imperative vs. Reactive
• Reactive Programming: using static or dynamic data flows and
propagation of change
• Example: a := b + c
• Functional Programming: evaluation of mathematical functions, avoids
changing-state and mutable data, declarative programming
• Side effects free => much easier to understand and predict the program
behavior.
Ex. (RxGo): observable := rxgo.Just("Hello", "Reactive", "World", "from", "RxGo")().
Map(ToUpper). // map to upper case
Filter(LengthGreaterThan4) // greaterThan4 func filters values > 4
for item := range observable.Observe() {
fmt.Println(item.V)
}

34
34
• According to Connal Elliot's (ground-breaking paper at
Conference on Functional Programming, 1997), FRP is:
(a) Denotative
(b) Temporally continuous
• FRP is asynchronous data-flow programming using the building
blocks of functional programming (map, reduce, filter, etc.)
and explicitly modeling time
Functional Reactive Programming (FRP)

35
35
Reactive Programming
• ReactiveX (Reactive Extensions) - open source polyglot (http://reactivex.io):
Rx = Observables + Flow transformations + Schedulers
• Go: RxGo, Kotlin: RxKotlin, Java: RxJava, JavaScript: RxJS, Python: RxPY,
C#: Rx.NET, Scala: RxScala, Clojure: RxClojure, C++: RxCpp, Ruby: Rx.rb,
Python: RxPY, Groovy: RxGroovy, JRuby: RxJRuby, ...
• Reactive Streams Specification (http://www.reactive-streams.org/):
• Publisher – provider of potentially unbounded number of sequenced
elements, according to Subscriber(s) demand (backpressure):
onNext* (onError | onComplete)
• Subscriber, Subscription
• Processor = Subscriber + Publisher

36
ReactiveX: Observable vs. Iterable
By ReactiveX - Creative Commons Attribution 3.0 License, Source: http://reactivex.io/intro.html
You can think of the Observable class as a “push” equivalent to Iterable, which is a “pull.” With an Iterable, the
consumer pulls values from the producer and the thread blocks until those values arrive. By contrast, with an
Observable the producer pushes values to the consumer whenever values are available. This approach is more
flexible, because values can arrive synchronously or asynchronously.

37
ReactiveX Observable – Marble Diagrams
By ReactiveX - Creative Commons Attribution 3.0 License, Source: http://reactivex.io/documentation/observable.html

38
Example: CombineLatest
https://projectreactor.io/core/docs/api/, Apache Software License 2.0

39
Redux == Rx Scan Opearator
Source: RxJava 2 API documentation, http://reactivex.io/RxJava/2.x/javadoc/

40
40
• PULL-based (Cold Event Streams) – Cold streams (e.g. RxJava
Observable / Flowable or Reactor Flow / Mono) are streams
that run their sequence when and if they are subscribed to.
They present the sequence from the start to each subscriber.
• PUSH-based (Hot Event Streams) – emit values independent of
individual subscriptions. They have their own timeline and events
occur whether someone is listening or not. When subscription is
made observer receives current events as they happen.
• Example: mouse events
Hot and Cold Event Streams

41
41
Converting Cold to Hot Stream
Source: RxJava 2 API documentation, http://reactivex.io/RxJava/2.x/javadoc/

42
Introduction to Machine Learning
42

43
What Is Machine Learning?
• Machine Learning (ML) –the study of computer algorithms that improve automatically
through experience. A subset of Artificial Intelligence (AI).
• ML algorithms build a model based on sample data, known as "training data", in order
to make predictions or decisions without being explicitly programmed to do so.
• ML algorithms are used in a wide variety of applications, where it is difficult or
unfeasible to develop conventional algorithms for the task – e.g. computer vision,
email filtering, digital assistants, natural language processing (NLP), recommendations,
personalized online advertising, chatbots, fraud detection, cybersecurity, medical
image analysis, self-driving cars, self-driving databases, and much, much more ...
• ML is closely connected with computational statistics, mathematical optimization, data
mining, predictive analytics.

44
Machine Learning and Artificial Intelligence
• ML learns and predicts based on passive observations, whereas AI implies an agent
interacting with the environment to learn and take actions that maximize its chance of
successfully achieving its goals.
Judea Pearl, The Book of Why, 2018
Pictures by: Yakoove - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=95589683
By Yakoove - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=95589710

45
Types of Learning Algorithms: Supervised Learning
• Supervised learning – build a mathematical model of a set of data that contains both
the inputs and the desired outputs.
• Training data – a set (matrix) of training examples (feature vectors), having one or
more inputs and the desired output. Through iterative optimization of an objective
function, supervised learning algorithms learn a function that can be used to predict
the output associated with new inputs.
• Examples: active learning, classification and regression.
• Active learning – learning algorithm can interactively query a user (teacher or oracle) to
label new data points with the desired outputs (optimal experimental design).
• Classification algorithms – when the outputs are restricted to a limited set of values
• Regression algorithms – when the outputs may have any numerical value within a range.
• Similarity learning – using a similarity function measuring how similar two objects are –
ranking, recommender systems, visual identity tracking, face and speaker verification.

46
Types of Learning Algorithms: Unsupervised Learning
• Unsupervised learning – takes a data set that contains only inputs, and find structure in
the data, like grouping or clustering of data points by identifying data commonalities
• Algorithms learn from test data that has not been labeled, classified or categorized.
• Applications: density function estimation, summarizing and explaining data features.
• Cluster analysis is the assignment of a set of observations into subsets (called clusters)
so that observations within the same cluster are similar according to one or more
predesignated criteria, while observations drawn from different clusters are dissimilar.
• Different clustering techniques make different assumptions on the structure of the data,
often defined by some similarity metric and evaluated, for example, by internal
compactness, or the similarity between members of the same cluster, and separation,
the difference between clusters. Other methods are based on estimated density and
graph connectivity.

47
Types of Learning Algorithms: Semi-supervised Learning
• Semi-supervised learning – falls between unsupervised learning (without any labeled
training data) and supervised learning (with completely labeled training data). During
training, it uses a smaller labeled data set to guide classification and feature extraction
from a larger, unlabeled data set.
• In weakly supervised learning, the training labels are noisy, limited, or imprecise; however,
these labels are often cheaper to obtain, resulting in larger effective training sets.

48
Types of Learning Algorithms: Reinforcement Learning
• Reinforcement learning – concerned with how software agents ought to take actions in
an environment so as to maximize some notion of cumulative reward
• Due to its generality, the field is studied in many other disciplines, such as game theory,
control theory, operations research, information theory, simulation-based optimization,
multi-agent systems, swarm intelligence, statistics and genetic algorithms.
• In ML, the environment is typically represented as a Markov
decision process (MDP). Many reinforcement learning algorithms
use dynamic programming techniques.
• Reinforcement learning algorithms do not assume knowledge of
an exact mathematical model of the MDP, and are used when
exact models are infeasible.
• Examples: reinforcement learning algorithms are used in
autonomous vehicles or in learning to play a game against human

49
Types of Learning Algorithms: Feature learning – I
• Feature learning (representation learning algorithms) – discover better representations of
the inputs provided during training. Examples: principal components analysis and cluster
analysis.
• Attempt to preserve the information in their input but also transform it in a way that makes
it useful, often as a pre-processing step before performing classification or predictions.
• Allows a machine to both learn the features and use them to perform a specific task.
• Feature learning can be either supervised or unsupervised. In supervised feature learning,
features are learned using labeled input data. Examples: artificial neural networks,
multilayer perceptrons, and supervised dictionary learning. In unsupervised feature
learning, features are learned with unlabeled input data. Examples: dictionary learning,
independent component analysis, autoencoders, matrix factorization, and clustering.

50
Types of Learning Algorithms: Feature learning – II
• Manifold learning algorithms attempt to do so under the constraint that the learned
representation is low-dimensional.
• Sparse coding algorithms attempt to do so under the constraint that the learned
representation is sparse, meaning that the mathematical model has many zeros.
• Multilinear subspace learning algorithms aim to learn low-dimensional representations
directly from tensor representations for multidimensional data, without reshaping them into
higher-dimensional vectors.
• Deep learning algorithms discover multiple levels of representation, or a hierarchy of
features, with higher-level, more abstract features defined in terms of (or generating)
lower-level features. It has been argued that an intelligent machine is one that learns a
representation that disentangles the underlying factors of variation that explain the
observed data.

51
51
Machine Learning Libraries
51

52
Top Machine Learning Libraries - I
• Apache Spark Mllib – fast, in-memory, big-data processing, which has also made it a
go-to framework for ML, highly scalable, can be coupled with any Hadoop data
source, supports Java, Scala, R and Python, new ML algorithms are constantly being
added, and existing enhanced.
• H20 + AutoML - open source, in-memory, distributed, fast, and scalable machine
learning and predictive analytics platform allowing to learn from big data, written in
Java, uses Distributed Key/Value store for accessing data, models, objects, etc. across
all nodes and machines, uses distributed Map/Reduce, utilizes Java Fork/Join multi-
threading, data read in parallel, distributed across the cluster, and stored in memory in
a columnar format in a compressed way, data parser has built-in intelligence to guess
the schema of the incoming dataset and supports data ingest from multiple sources in
various formats, APIs: REST, Flow UI, H2O-R, H2O-Python, supports Deep Learning, Tree
Ensembles, and Generalized Low Rank Models (GLRM).

53
Top Machine Learning Libraries - II
• Microsoft Azure ML Studio - Microsoft’s ML framework runs in their Azure cloud, offering
high-capacity data processing on a pay-as-you-go model. Its interactive, visual
workspace can be used to create, test and iterate ML “experiments” using the built-in
ML packages; they can then be published and shared as web services. Python or R
custom coding is also supported. A free trial is available for new users.
• Amazon Machine Learning - like Azure, the Amazon Machine Learning framework is
cloud-based, supporting Amazon S3, Redshift and RDS. It combines ML algorithms with
interactive tools and data visualizations which allow users to easily create, evaluate
and deploy ML models. These models can be binary, categoric or numeric – making
them useful in areas such as information filtering and predictive analytics.

54
Top Machine Learning Libraries - III
• Microsoft Distributed Machine Learning Toolkit (DMTK) - uses local data caching to
make it more scalable and efficient on computing clusters that have limited resources
on each node. Its distributed ML algorithms allow for fast training of gradient boosting,
topic and word-embedding learning models:
• LightLDA: Scalable, fast and lightweight system for large-scale topic modeling.
• LightGBM: LightGBM is a fast, distributed, high performance gradient boosting (GBDT, GBRT,
GBM or MART) framework based on decision tree algorithms, used for ranking, classification
and many other machine learning tasks.
• Distributed word embedding: a parallelization of the Word2Vec algorithm
• Distributed Multi-sense Word Embedding (DMWE) - a parallelization of the Skip-Gram
Mixture algorithm used for polysemous words.
• Google TensorFlow - open-source ML toolkit that has been applied in areas ranging
from machine translation to biomedical science. Data flows are processed by a series
of algorithms described by a graph. This allows for a flexible, multi-node architecture
that supports multiple CPUs and GPUs. The recently released Tensorflow Lite adds
support for neural processing on mobile phones.

55
Top Machine Learning Libraries - IV
• Caffe - developed by Berkeley AI Research, Caffe claims to offer one of the fastest
implementations of convolutional neural networks. It was originally developed for
machine vision projects but has since expanded to other applications, with its
extensible C++ code designed to foster active development. Both CPU and GPU
processing are supported.
• Veles - developed and released as open source by Samsung, Veles is a distributed ML
platform designed for rapid deep-learning application development. Users can train
various artificial neural net types – including fully connected, convolutional and
recurrent. It can be integrated with any Java application, and its “snapshot” feature
improves disaster recovery.

56
Top Machine Learning Libraries - V
• Massive Online Analysis (MOA) - developed at the University of Waikato in New
Zealand, this open-source Java framework specializes in real-time mining of streaming
data and large-scale ML. It offers a wide range of ML algorithms, including
classification, regression, clustering, outlier detection and concept drift detection.
Evaluation tools are also provided.
• Torch - Torch ML library makes ML easy and efficient, thanks to its use of the Lua
scripting language and a GPU-first implementation. It comes with a large collection of
community-driven ML packages that cover computer vision, signal processing, audio
processing, networking and more. Its neural network and optimization libraries are
designed to be both accessible and flexible.

57
57
Deep Learning
57
Convolutional Neural Networks

58
Deep Learning
• Deep learning (deep structured learning) - artificial neural networks with representation
learning – supervised, semi-supervised or unsupervised.
• Deep-learning architectures such as deep neural networks, deep belief networks,
recurrent neural networks (RNN) and convolutional neural networks (CNN) have been
applied to fields including computer vision, machine vision, speech recognition,
natural language processing, audio recognition, social network filtering, machine
translation, bioinformatics, drug design, medical image analysis, material inspection
and game programing, where they have produced results comparable to and in
some cases surpassing human expert performance.
• Artificial neural networks (ANNs) were inspired by information processing and
distributed communication nodes in biological systems. ANNs have various differences
from biological brains. Specifically, neural networks tend to be static and symbolic,
while the biological brain of most living organisms is dynamic (plastic) and analog.

59
Artificial Neural Networks (ANNs)
• ANN is based on a collection of connected units or
nodes called artificial neurons, which model
the neurons in a biological brain.
• Synapses can transmit a signal to other neurons,
“signal" is a real number, output is computed by some
non-linear function of the sum of its inputs -> threshold.
• The connections are called edges - typically have
weight that adjusts as learning proceeds.
• Typically, neurons are aggregated into layers. Different
layers may perform different transformations on their
inputs. Signals travel from the first layer (the input
layer), to the last layer (the output layer), possibly after
traversing the layers multiple times.
By Egm4313.s12 (Prof. Loc Vu-Quoc) - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=72816083
CC BY-SA 2.0, https://commons.wikimedia.org/w/index.php?curid=125222

60
Feed Forward Networks (FFN)
• FFN - artificial neural network wherein
connections between the nodes
do not form a cycle. As such, it is different
from its descendant: recurrent neural
networks.
By Mcstrother - Own work, CC BY 3.0, https://commons.wikimedia.org/w/index.php?curid=9516009
By Mcstrother - Own work, CC BY 3.0, https://commons.wikimedia.org/w/index.php?curid=9515940

61
ANNs, Multilayer Perceptron (MLP)
• MLP - class of feedforward artificial neural network (ANN),
three layers: input layer, hidden layer, and output layer,
nonlinear activation function, supervised learning
technique called backpropagation for training.
CC BY-SA 3.0, https://en.wikipedia.org/w/index.php?curid=8201514

62
Error Backprpagation
Source: https://en.wikipedia.org/wiki/Multilayer_perceptron

63
Deep Learning
• The adjective "deep" in deep learning comes from the use of multiple layers in the
network. Early work showed that a linear perceptron cannot be a universal classifier,
and that a network with a nonpolynomial activation function with one hidden layer of
unbounded width can be so.
• Deep learning – unbounded number of layers of bounded size, which permits practical
application and optimized implementation, while retaining theoretical universality
under mild conditions.
• Layers can be heterogeneous and to deviate widely from biologically informed
connectionist models, for the sake of efficiency, trainability and understandability,
whence the "structured" part.

64
Convolutional Neural Networks (CNN)
• Convolutional Neural Network (CNN, or ConvNet) - a class of deep neural networks, most
commonly applied to analyzing visual imagery, also known as shift invariant or space
invariant artificial neural networks (SIANN), based on their shared-weights architecture and
translation invariance characteristics, have applications in image and video recognition,
recommender systems, image classification, medical image analysis, natural language
processing, brain-computer interfaces, and financial time series.
• CNNs are regularized versions of multilayer perceptrons – MPLs usually mean fully
connected networks, that is, each neuron in one layer is connected to all neurons in the
next layer -> overfitting, typical ways of regularization include adding measurement of
weights to the loss function.
• CNNs take a different approach towards regularization – they take advantage of the
hierarchical pattern in data and assemble more complex patterns using smaller and
simpler patterns -> lower connectedness and complexity.

65
Picture by: By Sven Behnke - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=82466022
By Original file: Avimanyu786SVG version: Tukijaaliwa - File:AI-ML-DL.png, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=90131352

66
• Convolutional networks were inspired by biological processes in that the connectivity
pattern between neurons resembles the organization of the animal visual cortex.
Individual cortical neurons respond to stimuli only in a restricted region of the visual
field known as the receptive field. The receptive fields of different neurons partially
overlap such that they cover the entire visual field.
• CNNs use relatively little pre-processing compared to other image classification
algorithms. This means that the network learns the filters that in traditional algorithms
were hand-engineered. This independence from prior knowledge and human effort in
feature design is a major advantage.

67
Picture by: https://wiki.seg.org/wiki/Machine_learning

68
Recurrent Neural Networks (RNN)
Picture by: https://wiki.tum.de/display/lfdv/Video+Analysis

69
RNN for Video Analysis
Picture by: https://wiki.tum.de/display/lfdv/Video+Analysis

71
Deep Leaning with H2O - I
• purely supervised training protocol for regression and classification tasks
• fast and memory-efficient Java implementations based on columnar compression and
fine- grain Map/Reduce
• multi-threaded and distributed parallel computation to be run on either a single node
or a multi-node cluster
• fully automatic per-weight adaptive learning rate for fast convergence
• optional specification of learning rate, annealing and momentum options
• regularization options include L1, L2, dropout, Hogwild! and model averaging to
prevent model overfitting
• grid search for hyperparameter optimization and model selection

72
Deep Leaning with H2O - II
• model checkpointing for reduced run times and model tuning
• automatic data pre- and post-processing (one-hot encoding and standardization) for
categorical and numerical data
• automatic imputation of missing values
• automatic tuning of communication vs computation for best performance
• model export in plain java code for deployment in production environments
• export of weights and biases as H2O frames
• additional expert parameters for model tuning
• deep autoencoders for unsupervised feature learning and anomaly detection
capabilities

73
capabilities

74
capabilities

75
capabilities

76
Receiver Operating Characteristic (ROC)
Receiver operating
characteristic (ROC) curve -
graphical plot that illustrates
the diagnostic ability of a
binary classifier system as its
discrimination threshold is
varied.

77
Receiver Operating Characteristic (ROC)
• condition positive (P) / negative (N) - the number of real positive / negative cases in the data
• true positive (TP) - eqv. with hit, true negative (TN) - eqv. with correct rejection
• false positive (FP) - eqv. with false alarm, Type I error, false negative (FN) - eqv. with miss, Type II error

78
78
Distributed Learning: Ensemble
Learning
78

79
Ensemble Learning
• Ensemble methods – using multiple learning algorithms to obtain better predictive
performance than could be obtained from any of the constituent learning algorithms
alone. The term ensemble usually means using the same type of base learners.
• Even if the hypothesis space contains hypotheses that are very well-suited for a
particular problem, it may be very difficult to find a good one. Ensembles combine
multiple hypotheses to form a better hypothesis. Resultant hypothesis, is not necessarily
contained within the hypothesis space of the building models – can be more complex.
• Ensemble over-fits the training data easier than a single model -> Bagging.
• Ensembles yield better results when there is a diversity among the models. Using variety
of strong learning algorithms, has been shown to be more effective than using
techniques that attempt to dumb-down the models in order to promote diversity.
• number of independent classifiers == number of classes => highest accuracy

80
Bagging – e.g. Random Forest
• Bagging (Bootstrap AGGregatING) – involves having each model in the ensemble vote
with equal weight. In order to promote model variance, bagging trains each model in
the ensemble using a randomly drawn subset of the training set.
• Example: Random Forest algorithm combines random decision trees with bagging to
achieve very high classification accuracy. Given a training set X = x1, ..., xn with
responses Y = y1, ..., yn, bagging repeatedly (B times) selects a random sample with
replacement of the training set and fits trees to these samples + feature bagging:
1. For b = 1, ..., B:
2. Sample, with replacement, n training examples from X, Y; call these Xb, Yb.
3. Train a classification or regression tree fb on Xb, Yb, by selecting a limited number of features
(ayttributes) to be used during the training (feature bagging).
4. After training, predictions for unseen samples x' can be made by averaging the predictions /
taking the majority vote from all the individual regression trees on x‘.

81
Decision Trees. Random Forests

82
Boosting
• Boosting - ensemble meta-algorithm for primarily reducing bias, and also variance,
based on the question posed by Kearns and Valiant (1988, 1989"Can a set of weak
learners create a single strong learner?"
• In supervised learning, and a family of machine learning algorithms that convert weak
learners to strong ones
• Weak learner – a classifier that is only slightly correlated with the true classification (it
can label examples better than random guessing). In contrast, a strong learner is a
classifier that is arbitrarily well-correlated with the true classification.
• Informally, [the hypothesis boosting] problem asks whether an efficient learning
algorithm […] that outputs a hypothesis whose performance is only slightly better than
random guessing [i.e. a weak learner] implies the existence of an efficient algorithm
that outputs a hypothesis of arbitrary accuracy [i.e. a strong learner]."

83
Boosting - Basic Algorithm II
By Sirakorn - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=85888769

84
Boosting – Basic Algorithm
1. Equal weight (uniform probability
distribution) is given to the sample
training data (say D1) initially. Let i = 1
2. This data (Di) is then given to a base
learner (say Li).
3. Mis-classified instances by ensemble
L1, L2, …Li are assigned a weight higher
than the correctly classified instances
(keeping the total probability
distribution = 1) => Di+1 – boosted data
4. Let i = i + 1. Go to step 2 until error
becomes acceptable (or other finish
criteria is met).
D1 D2 Di
L1 L2 Li

85
AdaBoost
• AdaBoost - the output of the other learning algorithms ('weak learners') is combined into a
weighted sum that represents the final output of the boosted classifier. AdaBoost is
adaptive in the sense that subsequent weak learners are tweaked in favor of those
instances misclassified by previous classifiers. AdaBoost is sensitive to noisy data and
outliers. The individual learners can be weak, but as long as the performance of each one
is slightly better than random guessing, the final model can be proven to be strong learner.
• Every learning algorithm tends to suit some problem types better than others, and typically
has many different parameters and configurations to adjust before it achieves optimal
performance on a dataset. AdaBoost (with decision trees as the weak learners) is often
referred to as the best out-of-the-box classifier. When used with decision tree learning,
information gathered at each stage of the AdaBoost algorithm about the relative
'hardness' of each training sample is fed into the tree growing algorithm such that later
trees tend to focus on harder-to-classify examples.

86
Discrete AdaBoost
Source: https://en.wikipedia.org/wiki/AdaBoost

87
Gradient Boosting
Hastie, T.; Tibshirani, R.; Friedman, J. H. (2009). "10. Boosting and Additive Trees". The Elements of Statistical Learning (2nd ed.). New York: Springer. pp. 337–384. ISBN
978-0-387-84857-0. Archived from the original on 2009-11-10.

88
Thank’s for Your Attention!
Trayan Iliev
IPT – Intellectual Products & Technologies
http://iproduct.org/
https://github.com/iproduct
https://twitter.com/trayaniliev
https://www.facebook.com/IPT.EACAD

Making Machine Learning Easy with H2O and WebFlux

More Related Content

What's hot (20)

Similar to Making Machine Learning Easy with H2O and WebFlux (20)

More from Trayan Iliev (12)

Recently uploaded (20)

Making Machine Learning Easy with H2O and WebFlux