Increasing Completion of Neural

appliancepartAI and Robotics

Oct 19, 2013 (3 years and 7 months ago)

74 views

Increasing Completion of Neural Networks Coursework
-

1

Presented at
CIS 2011 ©
Dr Richard Mitchell
2011

Increasing Completion of Neural
Networks
Coursework


Dr
Richard
Mitchell

Senior Lecturer and University Teaching Fellow


Cybernetics
Research
Group


School
of Systems Engineering

University of Reading, UK

R.J.Mitchell@reading.ac.uk

Increasing Completion of Neural Networks Coursework
-

2

Presented at
CIS 2011 ©
Dr Richard Mitchell
2011

Overview

Part 2 module on Neural Networks assessment :


Implement an Object Oriented Neural Network



Complete suitable specified hierarchy


Apply it to real world problem of students own choice

Problem


Significant numbers not completing work (up to 33%)



Some not apply network; some not even implement NN

Paper


Describes strategies employed to overcome


This year 95% completed work



Influences for Approach

Feedback effective if students act on it to improve on future work
and learning

Glover and Brown

Most likely if

Feedback frequent, timely, sufficiently detailed

Feedback linked to purpose of assessment

Feedback understandable by students

Focus on learning by relating to future work

Gibbs and Simpson

Increasing Completion of Neural Networks Coursework
-

3

Presented at CIS 2011 © Dr Richard Mitchell 2011

Increasing Completion of Neural Networks Coursework
-

4

Presented at
CIS 2011 ©
Dr Richard Mitchell
2011

Nomenclature in Multi Layer Net

x
2
(2)

S

f

x
1

1

1

w
2

x
2
(1)

d
2
(1)

d
2
(2)

d
2
(3)

x
2
(3)

w
2,3

S

f

S

f

w
2,2

w
2,1

Data In

x
3
(2)

x
2

w
3

x
3
(1)

d
3
(1)

d
3
(2)

S

f

S

f

w
3
,2

w
3
,1

x
3

Data Out

x
m

outputs layer m

w
m

weights layer m

d
m

deltas layer m

Program with objects for each layer, not each neuron:

Relevant Equations


for each layer

Increasing Completion of Neural Networks Coursework
-

5

Presented at CIS 2011 © Dr Richard Mitchell 2011

Some commonality in functions / data : so o
-
o approach sensible

Increasing Completion of Neural Networks Coursework
-

6

Presented at
CIS 2011 ©
Dr Richard Mitchell
2011

Hierarchy

SingleLinearLayer


complete network able to compute
/ learn data

+ calc its weighted deltas

SingleSigmoidalLayer


inherits, own Calc Outputs / Delta

MultiSigmoidalLayer


hidden layer plus pointer to next

So is multilayer network

Most of its functions 2 lines
-

process own layer and next

MultiSigmoidal

Class

Base

SingleLinear

SingleSigmoidal

Classes

Increasing Completion of Neural Networks Coursework
-

7

Presented at CIS 2011 © Dr Richard Mitchell 2011

SingleLinearLayer

numInputs
,

numNeurons
,
numWeights
;

outputs, deltas, weights,
wtchanges
;

CalcOutputs
(inputs);

FindDeltas

(targets);

ChangeAllWeights
(inputs,
lparas
)

PrevLayersErrors
(errors)

Constructor (
numIn
,
numOut
);

Destructor;

Compute
Network
(data);

AdaptNetwork
(data,
lparas
);

SetTheWeights
(
initWts
);

ReturnTheWeights
(
theWts
);

SingleSigmoidaLayer

CalcOutputs
(inputs);

FindDeltas

(targets);

ErrorsToDeltas
();

Constructor (
numIn
,
numOut
);

Destructor;

MultiSigmoidalLayer

nextLayer
;

CalcOutputs
(inputs);

FindDeltas

(targets);

ChangeAllWeights
(inputs,
lparas
)

Constructor (
numIn
,
numOut
,
nxt
);

Destructor;

SetTheWeights
(
initWts
);

ReturnTheWeights
(
theWts
);

Shows name, protected parts (data
+ functions) and public interface

Strategies for Better Completion

Increasing Completion of Neural Networks Coursework
-

8

Presented at CIS 2011 © Dr Richard Mitchell 2011

For developing the neural network

Divide into series of tasks


back up help via VLE

Better, have 3 lab sessions two weeks apart

Demonstrator help in lab

Students copy code/results into template document

Easily marked, direct relevant feedback

Students make corrections before next lab

For application

Students have working network by Spring term

Five weeks to apply problem, investigate data processing,
different structures, etc

Write up as conference paper not report

Tasks in Lab Sessions

Increasing Completion of Neural Networks Coursework
-

9

Presented at CIS 2011 © Dr Richard Mitchell 2011

1)
Lab 1 : Complete
SingleLinear
; write
SingleSigmoidal


Create Project from provided files : familiarisation


Simple change, write function to return weights in net


Write code so network can learn


Write functions for
SingleSigmoidal

2)
Lab 2 : Write
MultiSigmoidal


Develop the
MultiSigmoidal


Investigate changing network size / learning
paras

3)
Lab 3 : Complete MLP


edit main program


So learn using train, validation and unseen data sets


Investigate changing network size / learning
paras

Completion Rates

Increasing Completion of Neural Networks Coursework
-

10

Presented at CIS 2011 © Dr Richard Mitchell 2011

Year

Class size

Students
reporting their
application

Students
completing MLP

2010
/
11

43

41

42

2009
/
10

56

46

50

2008
/
9

65

49

62

2007
/
8
*

71

56

50

* 2007/8 No lab sessions, but students allowed to use another’s
MLP for application

Reflections

Increasing Completion of Neural Networks Coursework
-

11

Presented at CIS 2011 © Dr Richard Mitchell 2011

Discussions with students and demonstrators show lab sessions
beneficial : help provided when needed; students self help

Students requested to reflect on comments also good

Consistent with Hughes :

“improved coursework submission attributed to learner
motivation from peer/tutor support”

Rapid relevant feedback also appreciated

Short conference paper, less work than report;

‘sold to students’ as learning opportunity as they will write paper
as part of assessment for Part 3 and 4 project

Increasing Completion of Neural Networks Coursework
-

12

Presented at
CIS 2011 ©
Dr Richard Mitchell
2011

Conclusions and Further Work

Coursework provides good practical example of object orientation,
applied to neural networks

Good use of encapsulation, small interface, inheritance

Staged tasks and rapid feedback help more to complete the neural
network

Conference paper seems to have worked well re completion of
complete assignment + good skill to have

Easier to mark


more time for giving feedback

Author to see if similar approaches can be used to better assess
whether students ‘engaging’ in Part 1


as used in ‘Engagement
system for Retention’.