shallow-training/report/report.lyx

2681 lines
46 KiB
Plaintext

#LyX 2.3 created this file. For more info see http://www.lyx.org/
\lyxformat 544
\begin_document
\begin_header
\save_transient_properties true
\origin unavailable
\textclass article
\begin_preamble
\def\changemargin#1#2{\list{}{\rightmargin#2\leftmargin#1}\item[]}
\let\endchangemargin=\endlist
\pagenumbering{gobble}
\usepackage{pxfonts}
\usepackage{color}
\definecolor{commentgreen}{RGB}{0,94,11}
\definecolor{darkblue}{rgb}{0,0,0.75}
\definecolor{darkred}{rgb}{0.6,0,0}
\end_preamble
\use_default_options true
\begin_modules
customHeadersFooters
minimalistic
todonotes
\end_modules
\maintain_unincluded_children false
\language british
\language_package default
\inputencoding utf8
\fontencoding global
\font_roman "default" "default"
\font_sans "default" "default"
\font_typewriter "default" "default"
\font_math "auto" "auto"
\font_default_family default
\use_non_tex_fonts false
\font_sc false
\font_osf false
\font_sf_scale 100 100
\font_tt_scale 100 100
\use_microtype true
\use_dash_ligatures true
\graphics default
\default_output_format default
\output_sync 0
\bibtex_command biber
\index_command default
\paperfontsize default
\spacing onehalf
\use_hyperref true
\pdf_title "Training Neural Networks With Backpropagation"
\pdf_author "Andy Pack"
\pdf_subject "EEEM005"
\pdf_keywords "EEEM005"
\pdf_bookmarks true
\pdf_bookmarksnumbered false
\pdf_bookmarksopen false
\pdf_bookmarksopenlevel 1
\pdf_breaklinks false
\pdf_pdfborder true
\pdf_colorlinks false
\pdf_backref false
\pdf_pdfusetitle true
\papersize default
\use_geometry true
\use_package amsmath 1
\use_package amssymb 1
\use_package cancel 1
\use_package esint 1
\use_package mathdots 1
\use_package mathtools 1
\use_package mhchem 1
\use_package stackrel 1
\use_package stmaryrd 1
\use_package undertilde 1
\cite_engine biblatex
\cite_engine_type authoryear
\biblio_style plain
\biblio_options urldate=long
\biblatex_bibstyle ieee
\biblatex_citestyle ieee
\use_bibtopic false
\use_indices false
\paperorientation portrait
\suppress_date true
\justification true
\use_refstyle 1
\use_minted 0
\index Index
\shortcut idx
\color #008000
\end_index
\leftmargin 1.8cm
\topmargin 2cm
\rightmargin 1.8cm
\bottommargin 2cm
\secnumdepth 3
\tocdepth 3
\paragraph_separation skip
\defskip medskip
\is_math_indent 0
\math_numbering_side default
\quotes_style british
\dynamic_quotes 0
\papercolumns 1
\papersides 1
\paperpagestyle fancy
\listings_params "language=Python,breaklines=true,frame=tb,otherkeywords={self},emph={State},emphstyle={\ttb\color{darkred}},basicstyle={\ttfamily},commentstyle={\bfseries\color{commentgreen}\itshape},keywordstyle={\color{darkblue}},emphstyle={\color{red}},stringstyle={\color{red}}"
\bullet 1 0 9 -1
\bullet 2 0 24 -1
\tracking_changes false
\output_changes false
\html_math_output 0
\html_css_as_file 0
\html_be_strict false
\end_header
\begin_body
\begin_layout Title
\size giant
Training Neural Networks with Backpropagation
\end_layout
\begin_layout Author
Andy Pack
\end_layout
\begin_layout Standard
\begin_inset VSpace 15pheight%
\end_inset
\end_layout
\begin_layout Standard
\align center
\begin_inset Graphics
filename surrey.png
lyxscale 15
width 40col%
\end_inset
\end_layout
\begin_layout Standard
\begin_inset VSpace vfill
\end_inset
\end_layout
\begin_layout Standard
\noindent
\align center
EEEM005
\begin_inset Newline newline
\end_inset
May 2021
\size large
\begin_inset Newline newline
\end_inset
Department of Electrical and Electronic Engineering
\begin_inset Newline newline
\end_inset
Faculty of Engineering and Physical Sciences
\begin_inset Newline newline
\end_inset
University of Surrey
\end_layout
\begin_layout Standard
\begin_inset Newpage newpage
\end_inset
\end_layout
\begin_layout Section*
Executive Summary
\end_layout
\begin_layout Standard
Summary here
\end_layout
\begin_layout Standard
\begin_inset Newpage newpage
\end_inset
\end_layout
\begin_layout Standard
\begin_inset ERT
status open
\begin_layout Plain Layout
\backslash
pagenumbering{roman}
\end_layout
\end_inset
\end_layout
\begin_layout Abstract
abstract
\end_layout
\begin_layout Standard
\begin_inset CommandInset toc
LatexCommand tableofcontents
\end_inset
\end_layout
\begin_layout Standard
\begin_inset Newpage pagebreak
\end_inset
\end_layout
\begin_layout Standard
\begin_inset FloatList figure
\end_inset
\end_layout
\begin_layout Standard
\begin_inset FloatList table
\end_inset
\end_layout
\begin_layout Standard
\begin_inset Newpage pagebreak
\end_inset
\end_layout
\begin_layout Right Footer
Andy Pack / 6420013
\end_layout
\begin_layout Left Footer
May 2021
\end_layout
\begin_layout Standard
\begin_inset ERT
status open
\begin_layout Plain Layout
\backslash
pagenumbering{arabic}
\end_layout
\begin_layout Plain Layout
\backslash
setcounter{page}{1}
\end_layout
\end_inset
\end_layout
\begin_layout Section
Introduction
\end_layout
\begin_layout Standard
Artificial neural networks have been the object of research and investigation
since the 1940s with
\noun on
McCulloch
\noun default
and
\noun on
Pitts
\noun default
' model of the artificial neuron
\begin_inset CommandInset citation
LatexCommand cite
key "McCulloch1943"
literal "false"
\end_inset
or
\emph on
Threshold Logic Unit
\emph default
.
Throughout the century, the development of the single and multi-layer perceptro
ns (SLP/MLP) alongside the backpropagation algorithm
\begin_inset CommandInset citation
LatexCommand cite
key "Rumelhart1986"
literal "false"
\end_inset
advanced the study of artificial intelligence.
Throughout the 2010s, convolutional neural networks have proved critical
in the field of computer vision and image recognition
\begin_inset CommandInset citation
LatexCommand cite
key "alexnet"
literal "false"
\end_inset
.
\end_layout
\begin_layout Standard
This work investigates the ability of a shallow multi-layer perceptron to
classify breast tumours as either benign or malignant.
The architecture and parameters were varied before exploring how the combinatio
n of classifiers can affect performance.
\end_layout
\begin_layout Standard
Investigations were carried out in
\noun on
Python
\noun default
using the
\noun on
TensorFlow
\noun default
package to construct, train and evaluate neural networks.
A
\noun on
Jupyter
\noun default
notebook containing the experiments and the evaluated parameters can be
seen formatted as a single script in appendix
\begin_inset CommandInset ref
LatexCommand ref
reference "sec:Source-Code"
plural "false"
caps "false"
noprefix "false"
\end_inset
.
The networks were trained using a supervised learning curriculum of labelled
data taken from a standard
\noun on
MatLab
\noun default
dataset
\begin_inset CommandInset citation
LatexCommand cite
key "matlab-dataset"
literal "false"
\end_inset
from the
\noun on
Deep Learning Toolbox
\noun default
.
For this binary-classification problem there are two formats for the network,
a single output node (threshold of 0.5 to differentiate classes) or two
output nodes to create a one-hot vector.
As the labels were formatted as one-hot vectors, two output nodes with
a softmax activation function were used.
The number of parameters associated with the employed architectures of
varying hdiden nodes can be seen in appendix
\begin_inset CommandInset ref
LatexCommand ref
reference "app:Network-Parameter-Counts"
plural "false"
caps "false"
noprefix "false"
\end_inset
while a graph of the constructed network can be seen in appendix
\begin_inset CommandInset ref
LatexCommand ref
reference "sec:Network-Graph"
plural "false"
caps "false"
noprefix "false"
\end_inset
.
\end_layout
\begin_layout Standard
Section
\begin_inset CommandInset ref
LatexCommand ref
reference "sec:exp1"
plural "false"
caps "false"
noprefix "false"
\end_inset
investigates the effect of varying the number of hidden nodes on test accuracy
along with the number of epochs that the MLPs are trained for.
Section
\begin_inset CommandInset ref
LatexCommand ref
reference "sec:exp2"
plural "false"
caps "false"
noprefix "false"
\end_inset
builds on the previous experiment by using reasonable parameter values
to investigate performance when using an ensemble of models to classify
in conjunction.
The effect of varying the number of nodes and epochs throughout the ensemble
was considered in order to determine whether combining multiple models
could produce a better accuracy than any individual model.
Section
\begin_inset CommandInset ref
LatexCommand ref
reference "sec:exp3"
plural "false"
caps "false"
noprefix "false"
\end_inset
investigates the effect of altering how the networks learn by changing
the optimisation algorithm.
Two additional algorithms to the previously used are considered and compared
using the same test apparatus of section
\begin_inset CommandInset ref
LatexCommand ref
reference "sec:exp2"
plural "false"
caps "false"
noprefix "false"
\end_inset
.
\end_layout
\begin_layout Section
Hidden Nodes & Epochs
\begin_inset CommandInset label
LatexCommand label
name "sec:exp1"
\end_inset
\end_layout
\begin_layout Standard
This section investigates the effect of varying the number of hidden nodes,
\begin_inset Formula $n_{h}$
\end_inset
, in the single hidden layer of a shallow multi-layer perceptron.
This is compared to the effect of training the model with different numbers
of epochs.
Throughout the experiment, stochastic gradient descent with momentum is
used as the optimiser, variations in both momentum and learning rate are
presented.
The learning rate and momentum coefficient used during training are denoted
\begin_inset Formula $\eta$
\end_inset
and
\begin_inset Formula $\beta$
\end_inset
respectively.
\end_layout
\begin_layout Subsection
Results
\end_layout
\begin_layout Standard
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-14-error-rate-curves.png
lyxscale 50
width 33col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.05$
\end_inset
,
\begin_inset Formula $\beta=0$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-14"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-12-error-rate-curves.png
lyxscale 50
width 33col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.1$
\end_inset
,
\begin_inset Formula $\beta=0$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-12"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-13-error-rate-curves.png
lyxscale 50
width 33col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.5$
\end_inset
,
\begin_inset Formula $\beta=0$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-13"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
Varied hidden node performance results over varied training lengths for
\begin_inset Formula $\eta=0.05,0.1,0.5$
\end_inset
,
\begin_inset Formula $\beta=0$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-12,14"
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Plain Layout
\end_layout
\end_inset
\end_layout
\begin_layout Standard
Figure
\begin_inset CommandInset ref
LatexCommand ref
reference "fig:exp1-test2-12,14"
plural "false"
caps "false"
noprefix "false"
\end_inset
visualises the test performance of hidden nodes up to
\begin_inset Formula $n_{h}=128$
\end_inset
over training periods up to 100 epochs in length.
In general, the error rate can be seen to decrease when the models are
trained for longer.
Increasing
\begin_inset Formula $n_{h}$
\end_inset
decreases the error rate and increases the gradient with which it falls
to a minimum limit.
As the learning rate increases, the speed with which the network converges
increases.
For
\begin_inset Formula $\eta=0.05$
\end_inset
, networks with large
\begin_inset Formula $n_{h}$
\end_inset
begin converging after 30 epochs.
This is after only 15 epochs for
\begin_inset Formula $\eta=0.1$
\end_inset
and almost immediately for
\begin_inset Formula $\eta=0.5$
\end_inset
.
\end_layout
\begin_layout Standard
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-14-error-rate-std.png
lyxscale 50
width 33col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.05$
\end_inset
,
\begin_inset Formula $\beta=0$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-14-std"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-12-error-rate-std.png
lyxscale 50
width 33col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.1$
\end_inset
,
\begin_inset Formula $\beta=0$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-12-std"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-13-error-rate-std.png
lyxscale 50
width 33col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.5$
\end_inset
,
\begin_inset Formula $\beta=0$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-13-std"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
Varied hidden node performance standard deviation results over varied training
lengths for
\begin_inset Formula $\eta=0.05,0.1,0.5$
\end_inset
,
\begin_inset Formula $\beta=0$
\end_inset
, note the larger
\begin_inset Formula $y$
\end_inset
scale for
\begin_inset Formula $\eta=0.5$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-12,14-std"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Standard
The standard deviations for the above discussed results of figure
\begin_inset CommandInset ref
LatexCommand ref
reference "fig:exp1-test2-12,14"
plural "false"
caps "false"
noprefix "false"
\end_inset
can be seen in figure
\begin_inset CommandInset ref
LatexCommand ref
reference "fig:exp1-test2-12,14-std"
plural "false"
caps "false"
noprefix "false"
\end_inset
.
In general, prior to the networks beginning to converge the standard deviation
is close to 0.
As previously described, this takes place at lower epochs for higher learning
rates.
Once the networks start converging, the standard deviation of the test
error rate increases.
Increasing the learning rate also increases the variance in test error
rates, the max value for
\begin_inset Formula $\eta=0.5$
\end_inset
is double that of the lower
\begin_inset Formula $\eta$
\end_inset
experiments within the first 20 epochs.
\end_layout
\begin_layout Standard
\begin_inset Flex TODO Note (inline)
status open
\begin_layout Plain Layout
more std stuff and test/train splits
\end_layout
\end_inset
\end_layout
\begin_layout Standard
The effect of varying momentum can be seen in figure
\begin_inset CommandInset ref
LatexCommand ref
reference "fig:exp1-momentums"
plural "false"
caps "false"
noprefix "false"
\end_inset
, a fixed learning rate of
\begin_inset Formula $\eta=0.01$
\end_inset
was maintained throughout.
The meaning of momentum and its effect on training is discussed in section
\begin_inset CommandInset ref
LatexCommand ref
reference "subsec:Stochastic-Gradient-Descent"
plural "false"
caps "false"
noprefix "false"
\end_inset
.
Without momentum (
\begin_inset CommandInset ref
LatexCommand ref
reference "fig:exp1-test2-11"
plural "false"
caps "false"
noprefix "false"
\end_inset
), it can be seen that the network does not begin to converge within 100
epochs.
This is also the case for
\begin_inset Formula $\beta=0.3$
\end_inset
, it is only by
\begin_inset Formula $\beta=0.5$
\end_inset
that some of the evaluated architectures begin to converge.
The test error rates for the 32 and 64 node series' begin to decrease after
64 epochs with
\begin_inset Formula $n_{h}=64$
\end_inset
nodes descending faster.
With
\begin_inset Formula $\beta=0.7$
\end_inset
, the 32 and 64-node networks begin to converge earlier, after 32 epochs
while the remaining architectures down to 2 nodes begin to converge after
64 epochs.
Finally, with
\begin_inset Formula $\beta=0.7$
\end_inset
, all of the evaluated architectures have convered by 64 epochs.
\end_layout
\begin_layout Standard
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-7-error-rate-curves.png
lyxscale 50
width 45col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.01$
\end_inset
,
\begin_inset Formula $\beta=0.9$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-7"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-10-error-rate-curves.png
lyxscale 50
width 45col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.01$
\end_inset
,
\begin_inset Formula $\beta=0.7$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-10"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-8-error-rate-curves.png
lyxscale 50
width 45col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.01$
\end_inset
,
\begin_inset Formula $\beta=0.5$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-8-1"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-9-error-rate-curves.png
lyxscale 50
width 45col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.01$
\end_inset
,
\begin_inset Formula $\beta=0.3$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-9"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp1-test2-11-error-rate-curves.png
lyxscale 50
width 45col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
\begin_inset Formula $\eta=0.01$
\end_inset
,
\begin_inset Formula $\beta=0.0$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-test2-11"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
Varied hidden node performance results over varied training length with
different momentum coefficients
\begin_inset CommandInset label
LatexCommand label
name "fig:exp1-momentums"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Subsection
Discussion
\end_layout
\begin_layout Standard
From the presented results, it can be seen that, generally, increasing either
learning rate or momentum increases the speed of convergence.
\end_layout
\begin_layout Standard
Increasing the number of hidden nodes also increases the speed of convergence.
However, it is worth noting that a large number of nodes is not required
to achieve a highly performant accuracy.
A single hidden node for a total of 14 parameters, with enough training,
was able to achieve similar results to a 64-node network of 770 or 5 times
as many parameters.
\end_layout
\begin_layout Section
Ensemble Classification
\begin_inset CommandInset label
LatexCommand label
name "sec:exp2"
\end_inset
\end_layout
\begin_layout Standard
A horizontal ensemble of
\begin_inset Formula $m$
\end_inset
models was constructed with majority vote in order to investigate whether
this could improve performance over that of any single model.
In order to introduce variation between models of the ensemble, a range
of hidden nodes and/or epochs could be defined.
When selecting parameters throughout the ensemble,
\begin_inset Formula $m$
\end_inset
equally spaced values within the range are selected
\begin_inset Foot
status open
\begin_layout Plain Layout
For
\begin_inset Formula $m=1$
\end_inset
, the average of the range is taken
\end_layout
\end_inset
.
\end_layout
\begin_layout Standard
The statistic
\emph on
agreement
\emph default
,
\begin_inset Formula $a$
\end_inset
, is defined as the proportion of models under the meta-classifier that
correctly predict a sample's class when the ensemble correctly classifies.
It could also be considered the confidence of the meta-classifier, for
one horizontal model
\begin_inset Formula $a_{m=1}\equiv1$
\end_inset
.
As error rates are presented as opposed to accuracy, this is inverted by
\begin_inset Formula $d=1-a$
\end_inset
to
\emph on
disagreement
\emph default
, the proportion of incorrect models when correctly group classifying.
Alongside the disagreement and ensemble test accuracy, the average individual
accuracy for both test and training data are also presented.
\end_layout
\begin_layout Subsection
Results
\end_layout
\begin_layout Standard
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp2-test8-error-rate-curves.png
lyxscale 50
width 50col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
Ensemble classifier performance results for
\begin_inset Formula $\eta=0.03$
\end_inset
,
\begin_inset Formula $\beta=0.01$
\end_inset
, nodes =
\begin_inset Formula $1-400$
\end_inset
, epochs =
\begin_inset Formula $5-100$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp2-test8"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Standard
An experiment with a fixed epoch value throughout the ensemble is presented
in figure
\begin_inset CommandInset ref
LatexCommand ref
reference "fig:exp2-test10"
plural "false"
caps "false"
noprefix "false"
\end_inset
.
Nodes between 1 and 400 were selected for the classifiers with a learning
rate,
\begin_inset Formula $\eta=0.15$
\end_inset
and momentum,
\begin_inset Formula $p=0.01$
\end_inset
.
The ensemble accuracy can be seen to be fairly constant throughout the
number of horizontal models with 3 models being the least accurate with
a higher standard deviation.
3 horizontal models also shows a significant spike in disagreement and
individual error rates which gradually decreases as the number of models
increases.
\end_layout
\begin_layout Standard
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/exp2-test10-error-rate-curves.png
lyxscale 50
width 50col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
Ensemble classifier performance results for
\begin_inset Formula $\eta=0.15$
\end_inset
,
\begin_inset Formula $\beta=0.01$
\end_inset
, nodes =
\begin_inset Formula $1-400$
\end_inset
, epochs = 20
\begin_inset CommandInset label
LatexCommand label
name "fig:exp2-test10"
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Plain Layout
\end_layout
\end_inset
\end_layout
\begin_layout Subsection
Discussion
\end_layout
\begin_layout Standard
From the data of figure
\begin_inset CommandInset ref
LatexCommand ref
reference "fig:exp2-test10"
plural "false"
caps "false"
noprefix "false"
\end_inset
, 3 horizontal models was shown to be the worst performing configuration
with lower ensemble accuracy and higher disagreement.
This is likely due to larger proportion that a single model constitutes.
When correct, three models may only have a disagreement of 1/3 or 0 and
thus the final value will lie somewhere between these two.
As the number of horiztonal models increases, the number of acceptable
disagreement values increases.
\end_layout
\begin_layout Section
Optimiser Comparisons
\begin_inset CommandInset label
LatexCommand label
name "sec:exp3"
\end_inset
\end_layout
\begin_layout Standard
Throughout the previous experiments the stochastic gradient descent optimiser
was used to change the networks weights but there are many different optimisati
on algorithms.
This section will present investigations into two other optimisation algorithms
and discuss the differences between them using the horizontal ensemble
classification of the previous section.
\end_layout
\begin_layout Standard
Prior to these investigations, however, stochastic gradient descent and
the two other subject algorithms will be described.
\end_layout
\begin_layout Subsection
Optimisers
\end_layout
\begin_layout Subsubsection
Stochastic Gradient Descent
\begin_inset CommandInset label
LatexCommand label
name "subsec:Stochastic-Gradient-Descent"
\end_inset
\end_layout
\begin_layout Standard
Gradient descent and the closely related stochastic and mini-batch gradient
descent are popular optimisation algorithms in the machine learning space.
\end_layout
\begin_layout Standard
The aim of the neural networks in question are to make correct classifications
on sample data being fed-forward, ideally the networks classification would
be equal to the provided label.
A loss function,
\begin_inset Formula $J$
\end_inset
, is defined as the difference between the predicted ouput and the target
labelled output
\begin_inset Foot
status open
\begin_layout Plain Layout
There are many different options for the loss function including mean squared
error and categorical cross-entropy.
Although they have significant differences, this coverage of optimisation
algorithms does not rely on a specific loss function.
\end_layout
\end_inset
, it follows that we are aiming to minimise this as much as possible.
In order to improve the network, the values of the parameters,
\begin_inset Formula $\theta$
\end_inset
, must be changed with the intention of reducing the loss value.
From a set of starting weights,
\begin_inset Formula $\theta_{0}$
\end_inset
, this could be completed by finding the gradient of
\family roman
\series medium
\shape up
\size normal
\emph off
\bar no
\strikeout off
\xout off
\uuline off
\uwave off
\noun off
\color none
\begin_inset Formula $J$
\end_inset
w.r.t
\family default
\series default
\shape default
\size default
\emph default
\bar default
\strikeout default
\xout default
\uuline default
\uwave default
\noun default
\color inherit
\begin_inset Formula $\theta_{0}$
\end_inset
\family roman
\series medium
\shape up
\size normal
\emph off
\bar no
\strikeout off
\xout off
\uuline off
\uwave off
\noun off
\color none
.
Formally this would be
\begin_inset Formula $\nabla_{\theta_{0}}J\left(\theta_{0}\right)$
\end_inset
, the first derivative of the loss function with respect to the current
weights.
In order to reduce the loss, the gradient should be subtracted from the
current weight, a scale factor,
\begin_inset Formula $\eta$
\end_inset
, or the learning rate is defined to apply a tuneable proportion of the
gradient to the starting values.
\end_layout
\begin_layout Standard
In order to iteratively apply this algorithm, the form below is used for
time steps,
\begin_inset Formula $t$
\end_inset
.
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
\theta_{t+1}=\theta_{t}-\eta\cdot\nabla_{\theta_{t}}J\left(\theta_{t}\right)
\end{equation}
\end_inset
\end_layout
\begin_layout Standard
The differences between standard or batch gradient descent and the previously
mentioned variants is how many samples are fed-forward as part of the optimisat
ion algorithm.
Standard gradient descent propagates and calculates weight changes for
the entire training dataset in a single iteration of the algorithm.
Stochastic gradient descent, instead, processes only one sample during
an iteration.
Mini-batch strikes a balance between the two, the speed of stochastic gradient
descent is retained as more weight updates are made, however the path through
the error surface can be noisier than vanilla gradient descent.
Therefore, although the algorithm is colloquially referred to as gradient
descent or SGD, more strictly as a batch size of 35 was used for this work,
mini-batch gradient descent is being used.
\end_layout
\begin_layout Standard
\noun on
Tensorflow's
\noun default
implementation of SGD also includses a momentum parameter.
Momentum aims to help a network increase the speed of convergence and reduce
oscillations by reinforcing dimensions (weights) that are changing in a
consistent direction while slowing dimensions that are changing direction
rapidly
\begin_inset CommandInset citation
LatexCommand cite
key "paperspace-mom-rmsprop-adam"
literal "false"
\end_inset
.
Momentum introduces a memory element to the descent by including a portion,
\begin_inset Formula $\beta$
\end_inset
, of the previous step's weight delta or
\emph on
velocity
\emph default
in subsequent iterations.
\end_layout
\begin_layout Standard
The introduction of momentum can be described as below
\begin_inset CommandInset citation
LatexCommand cite
key "tf.keras.optimizers.SGD"
literal "false"
\end_inset
,
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
v_{t}=\beta\cdot v_{t-1}-\eta\cdot\nabla_{\theta_{t}}J\left(\theta_{t}\right)\label{eq:sgd-momentum}
\end{equation}
\end_inset
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
\theta_{t+1}=\theta_{t}+v_{t}
\end{equation}
\end_inset
\end_layout
\begin_layout Standard
As previously presented (figure
\begin_inset CommandInset ref
LatexCommand ref
reference "fig:exp1-momentums"
plural "false"
caps "false"
noprefix "false"
\end_inset
), momentum can significantly increase convergence speed.
\end_layout
\begin_layout Subsubsection
RMSprop
\end_layout
\begin_layout Standard
Although gradient descent is a powerful optimisation algorithm, there are
drawbacks.
One limitation is that the learning rate,
\begin_inset Formula $\eta$
\end_inset
, is a scalar applied to all gradients.
As a result, smaller gradients as would be found at saddle points move
slowly.
An alternative would be to expand the single scalar to a learning rate
per parameter that could move dynamically throughout the training process,
known as adaptive learning rate optimisation.
\end_layout
\begin_layout Standard
One such algorithm is RMSprop
\begin_inset CommandInset citation
LatexCommand cite
key "rmsprop-hinton"
literal "false"
\end_inset
or
\emph on
root mean square propagation
\emph default
, an unpublished algorithm that builds on previous adaptive algorithms such
as Rprop and Adagrad.
These aimed to overcome the shortcomings of SGD by using just the sign
of the calculated gradients and allowing the learning rate alone to define
the size of the step.
Instead of a constant or defined learning rate schedule, each learning
rate
\emph on
floats
\emph default
and is scaled up or down based on whether it is consistently changing in
the same direction each iteration.
\end_layout
\begin_layout Standard
Equations for RMSprop can be seen below
\begin_inset CommandInset citation
LatexCommand cite
key "understanding-rmsprop"
literal "false"
\end_inset
.
For conciseness, the previously defined derivative of the loss function
w.r.t to the current parameters is shortened,
\begin_inset Formula $g_{t}=\nabla_{\theta_{t}}J\left(\theta_{t}\right)$
\end_inset
.
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
E\left[g^{2}\right]_{t}=\alpha\cdot E\left[g^{2}\right]_{t-1}+\left(1-\alpha\right)\cdot g_{t}^{2}\label{eq:rmsprop-expected-value}
\end{equation}
\end_inset
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
\theta_{t+1}=\theta_{t}-\frac{\eta}{\sqrt{E\left[g^{2}\right]_{t}+\epsilon}}g_{t}\label{eq:rmsprop-update}
\end{equation}
\end_inset
\end_layout
\begin_layout Standard
As previously mentioned, only the sign of the gradient is used, this can
be achieved by dividing
\begin_inset Formula $g$
\end_inset
by the magnitude
\begin_inset Formula $|g|$
\end_inset
.
RMSprop extends this by instead dividing by the exponential average of
squared gradients, equation
\begin_inset CommandInset ref
LatexCommand ref
reference "eq:rmsprop-expected-value"
plural "false"
caps "false"
noprefix "false"
\end_inset
.
In this equation,
\begin_inset Formula $\alpha$
\end_inset
constitutes the gradient decay rate, a value of 0.9 is suggested
\begin_inset CommandInset citation
LatexCommand cite
key "understanding-rmsprop"
literal "false"
\end_inset
.
\begin_inset Formula $\epsilon$
\end_inset
is a small constant on the order of
\begin_inset Formula $1\times10^{-7}$
\end_inset
that stops the algorithm from dividing by 0.
\end_layout
\begin_layout Subsubsection
Adam
\end_layout
\begin_layout Standard
Adam or
\emph on
adaptive moment estimation
\emph default
is an optimisation algorithm that combines the adaptive learning rates
of RMSprop with the previously described momentum
\begin_inset CommandInset citation
LatexCommand cite
key "adam-paper"
literal "false"
\end_inset
.
Like RMSprop, the exponential average of squared gradients is maintained,
compare equations
\begin_inset CommandInset ref
LatexCommand ref
reference "eq:rmsprop-expected-value"
plural "false"
caps "false"
noprefix "false"
\end_inset
and
\begin_inset CommandInset ref
LatexCommand ref
reference "eq:adam-squared-grad-accum"
plural "false"
caps "false"
noprefix "false"
\end_inset
.
In addition to this, however, the exponential average of gradients is maintaine
d with a similar function to momentum, compare equations
\begin_inset CommandInset ref
LatexCommand ref
reference "eq:sgd-momentum"
plural "false"
caps "false"
noprefix "false"
\end_inset
and
\begin_inset CommandInset ref
LatexCommand ref
reference "eq:adam-momentum"
plural "false"
caps "false"
noprefix "false"
\end_inset
.
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
m_{t}=\beta_{1}\cdot m_{t-1}+\left(1-\beta_{1}\right)g_{t}\label{eq:adam-momentum}
\end{equation}
\end_inset
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
v_{t}=\beta_{2}\cdot v_{t-1}+\left(1-\beta_{2}\right)g_{t}^{2}\label{eq:adam-squared-grad-accum}
\end{equation}
\end_inset
\end_layout
\begin_layout Standard
These two equations constitute the eponymous moments,
\begin_inset Formula $m_{t}$
\end_inset
is the first moment or mean while
\begin_inset Formula $v_{t}$
\end_inset
is the second moment or the uncentered variance of the gradients.
As these moments are initialised at zero, these estimations tend to bias
towards 0.
The original authors correct the bias using the below
\begin_inset CommandInset citation
LatexCommand cite
key "adam-paper"
literal "false"
\end_inset
.
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
\hat{m}_{t}=\frac{m_{t}}{1-\beta_{1}^{t}}
\end{equation}
\end_inset
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
\hat{v}_{t}=\frac{v_{t}}{1-\beta_{2}^{t}}
\end{equation}
\end_inset
\end_layout
\begin_layout Standard
This leaves the update step itself, described below.
Similarities can be seen between the previous RMSprop update step (equation
\begin_inset CommandInset ref
LatexCommand ref
reference "eq:rmsprop-update"
plural "false"
caps "false"
noprefix "false"
\end_inset
) and that of Adam.
The RMSprop momentum term
\family roman
\series medium
\shape up
\size normal
\emph off
\bar no
\strikeout off
\xout off
\uuline off
\uwave off
\noun off
\color none
\begin_inset Formula $E\left[g^{2}\right]_{t}$
\end_inset
has been replaced by the equivalent
\begin_inset Formula $v_{t}$
\end_inset
while the calculated gradient,
\begin_inset Formula $g_{t}$
\end_inset
has been replaced by the exponentially decaying average gradient,
\begin_inset Formula $m_{t}$
\end_inset
.
\end_layout
\begin_layout Standard
\begin_inset Formula
\begin{equation}
\theta_{t+1}=\theta_{t}-\frac{\eta}{\sqrt{\hat{v}_{t}+\epsilon}}\hat{m}_{t}\label{eq:adam-update}
\end{equation}
\end_inset
\end_layout
\begin_layout Subsection
Results
\end_layout
\begin_layout Standard
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename /home/andy/dev/py/shallow-training/graphs/exp3-test1-error-rate-curves.png
lyxscale 30
width 100col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
Ensemble classifier performance results for SGD, RMSprop and Adam optimisation
with
\begin_inset Formula $\eta=0.1$
\end_inset
,
\begin_inset Formula $\beta=0.0$
\end_inset
, nodes = 16, epochs =
\begin_inset Formula $1-100$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp3-test1"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Standard
\begin_inset Float figure
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename /home/andy/dev/py/shallow-training/graphs/exp3-test7-error-rate-curves.png
lyxscale 30
width 100col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
Ensemble classifier performance results for SGD, RMSprop and Adam optimisation
with
\begin_inset Formula $\eta=0.1$
\end_inset
,
\begin_inset Formula $\beta=0.9$
\end_inset
, nodes =
\begin_inset Formula $1-400$
\end_inset
, epochs =
\begin_inset Formula $50-100$
\end_inset
\begin_inset CommandInset label
LatexCommand label
name "fig:exp3-test7"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Subsection
Discussion
\end_layout
\begin_layout Standard
In suggesting a optimal algorithm it is worth considering the intended domains
for RMSprop and Adam.
As newer algorithms, there tends to a focus on deep convolutional networks
which implies a somewhat different set of requirements.
This is not to say that the algorithms are inappropriate for the presented
applications, as demonstrated, these more complex algorithms were able
to outperform the employed gradient descent with optional momentum.
\end_layout
\begin_layout Section
Conclusions
\end_layout
\begin_layout Standard
\begin_inset Newpage newpage
\end_inset
\end_layout
\begin_layout Standard
\begin_inset CommandInset label
LatexCommand label
name "sec:bibliography"
\end_inset
\begin_inset CommandInset bibtex
LatexCommand bibtex
btprint "btPrintCited"
bibfiles "references"
options "bibtotoc"
\end_inset
\end_layout
\begin_layout Section
\start_of_appendix
Network Parameter Counts
\begin_inset CommandInset label
LatexCommand label
name "app:Network-Parameter-Counts"
\end_inset
\end_layout
\begin_layout Standard
\begin_inset Float table
placement H
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Tabular
<lyxtabular version="3" rows="9" columns="2">
<features tabularvalignment="middle">
<column alignment="center" valignment="top">
<column alignment="center" valignment="top">
<row>
<cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
Hidden Nodes
\end_layout
\end_inset
</cell>
<cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" rightline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
Trainable Parameters
\end_layout
\end_inset
</cell>
</row>
<row>
<cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
1
\end_layout
\end_inset
</cell>
<cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
14
\end_layout
\end_inset
</cell>
</row>
<row>
<cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
2
\end_layout
\end_inset
</cell>
<cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
26
\end_layout
\end_inset
</cell>
</row>
<row>
<cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
4
\end_layout
\end_inset
</cell>
<cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
50
\end_layout
\end_inset
</cell>
</row>
<row>
<cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
8
\end_layout
\end_inset
</cell>
<cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
98
\end_layout
\end_inset
</cell>
</row>
<row>
<cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
16
\end_layout
\end_inset
</cell>
<cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
194
\end_layout
\end_inset
</cell>
</row>
<row>
<cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
32
\end_layout
\end_inset
</cell>
<cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
386
\end_layout
\end_inset
</cell>
</row>
<row>
<cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
64
\end_layout
\end_inset
</cell>
<cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
770
\end_layout
\end_inset
</cell>
</row>
<row>
<cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
128
\end_layout
\end_inset
</cell>
<cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" rightline="true" usebox="none">
\begin_inset Text
\begin_layout Plain Layout
1,538
\end_layout
\end_inset
</cell>
</row>
</lyxtabular>
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
Number of trainable parameters for architectures of varying numbers of hidden
nodes
\begin_inset CommandInset label
LatexCommand label
name "tab:trainable-params"
\end_inset
\end_layout
\end_inset
\end_layout
\begin_layout Plain Layout
\end_layout
\end_inset
\end_layout
\begin_layout Section
Source Code
\begin_inset CommandInset label
LatexCommand label
name "sec:Source-Code"
\end_inset
\end_layout
\begin_layout Standard
\begin_inset CommandInset include
LatexCommand lstinputlisting
filename "../nncw.py"
lstparams "caption={Formatted Jupyter notebook containing experiment code},label={notebook-code}"
\end_inset
\end_layout
\begin_layout Section
Network Graph
\begin_inset CommandInset label
LatexCommand label
name "sec:Network-Graph"
\end_inset
\end_layout
\begin_layout Standard
\begin_inset Float figure
placement H
wide false
sideways false
status open
\begin_layout Plain Layout
\noindent
\align center
\begin_inset Graphics
filename ../graphs/tensorboard-graph.png
lyxscale 50
width 100col%
\end_inset
\end_layout
\begin_layout Plain Layout
\begin_inset Caption Standard
\begin_layout Plain Layout
Single hidden layer neural network as graphed by
\noun on
Tensorboard
\noun default
\begin_inset CommandInset label
LatexCommand label
name "fig:tensorboard"
\end_inset
\end_layout
\end_inset
\end_layout
\end_inset
\end_layout
\end_body
\end_document