首页 > > 详细

代写3RD 调试Python作业

Call  for  Papers  for  the  
3RD  ANU  ANNUAL  BIO-­‐INSPIRED  COMPUTING  STUDENT  CONFERENCE  http://cs.anu.edu.au/~tom/conf/ABCs2020/  
also  being  used  for  
COMP4660/8420  Assignment  1:  Neural  Networks    
    1  
Submission  Due:  Sunday  3rd  May  at  11:55pm  
 
Context    
 
Neural  networks  research  prior  to  the  deep  learning  boom  has  a  lot  to  teach  us  still.  The  
neural  networks  part  of  the  course  focuses  on  the  research  in  this  area  from  1986  (back-­‐‑
propagation  algorithm  published)  into  the  1990s  and  early  2000s.  Later  papers  tend  to  be  too  
esoteric  to  be  useful  for  course  based  learning  and  I  prefer  you  to  read  papers  rather  than  just  
textbooks.    
 
The  assignment  will  provide  you  some  experience  with  taking  two  papers:  one  with  a  dataset,  
and  another  with  a  technique,  and  you  will  do  a  mix  of  data  encoding  /  analysis  on  the  dataset  
and  implementing  /  applying  a  technique  from  the  paper.  Thus,  each  of  your  assignments  will  
be  unique,  but  you  will  still  be  able  to  usefully  discuss  with  other  students,  as  there  will  be  
other  people  using  the  dataset  (but  different  techniques/papers)  and  other  people  using  the  
same  technique  (but  on  different  datasets).  
 
To  be  possible  as  an  assignment,  we  stick  to  simple  (by  modern  standards)  techniques,  which  
fits  well  with  the  chronological  focus  from  1986  to  early  2000s.  Some  of  the  datasets  are  from  
the  same  period,  and  some  are  from  later  papers.  
 
Wattle  will  shortly  have  a  way  for  you  to  select  the  technique,  then  the  dataset.  
 
I  have  uploaded  a  number  of  papers  relating  to  the  neural  network  lectures  to  Wattle,  most  of  
the  technique  papers  will  be  one  of  these.  
 
Perhaps  it  is  worth  noting  that  some  of  the  techniques  developed  in  that  early  neural  
networks  period  a  likely  be  of  use  today.  My  example  is  the  use  of  an  input  feature  selection  
technique  of  mine  from  1997  which  has  recently  been  implemented  in  the  H2O  AI  package.  In  
Google  type:  “Gedeon  method”  H2O  and  you  get  a  surprising  number  of  recent  hits.  I  should  
also  mention  that  I  have  no  connections  with  H2O  and  found  out  about  this  when  a  Korean  
researcher  emailed  me  to  ask  me  about  the  function  in  the  H2O  package  and  I  discovered  that  
it  was  based  on  my  paper  J.    Similarly,  please  also  note  that  the  CNN  architecture  dates  to  
the  1980s  (Fukushima’s  Neocognitron  in  1980  to  LeCun’s  coining  of  the  term  in  1989)  but  was  
not  competitive  given  the  computing  power  and  datasets  generally  available  at  that  time.  Also  
please  note  that  most/many  deep  learning  methods  have  fully  connected  layers  at  the  end  
which  are  simple  neural  networks  as  in  the  first  part  of  the  course.  
 
Deliverables  
1. A  report  (about  4-­‐‑6  pages  of  text  content  with  a  MAXIMUM  of  10  pages,  including  
   
  2  
references,  diagrams,  graphs  and  tables).  Remember  to  keep  your  report  clear  and  
concise.  I  say  that  the  text  content  should  be  about  4-­‐‑6  pages  to  signal  that  we  will  not  
be  counting  lines,  but  if  it  is  10  pages  long  and  it  is  all  diagrams  then  it  is  very  clear  
that  there  is  too  little  text.  Conversely,  if  it  is  10  pages  long  and  there  is  just  one  
diagram  then  there  is  too  much  text,  but  6  pages  with  one  diagram  is  fine,  and  so  on.  
2. PyTorch  or  Python  source  code  file(s).  
3. The  dataset  you  used  for  your  assignment,  in  original  and  preprocessed  form,  and  a  
copy  of  the  paper.  (I  know  we  provided  it,  but  this  makes  it  easier  for  marking  if  
everything  is  in  one  spot.)  
4. A  copy  of  the  technique  paper  (for  consistency  and  to  help  the  markers).  Both  the  
dataset  and  technique  papers  must  be  cited  within  your  report,  along  with  at  least  4  
more  papers.  
Submission  Method  
Please  submit  your  assignment  via  the  EasyChair  conference  management  system:  
https://www.easychair.org/conferences/?conf=abcs2020.    
You  will  submit  a  pdf  for  the  report,  the  name  must  start  with  your  uNumber,  and  a  zip  
folder  of  the  other  deliverables.    
Objectives  
The  purpose  for  this  assignment  is  for  you  to:  
• Develop  a  good  understanding  of  data  management  for  artificial  neural  networks  and  
enhance  your  skills  in  implementing  neural  network  techniques  in  PyTorch  /  Python  
• Enhance  your  approach  to  investigating  a  data  set  and  setting  up  a  problem  for  an  
artificial  neural  network  to  solve  a  classification/regression  problem  
• Develop  a  good  understanding  for  reporting  investigations  in  a  conference  style  paper  
• A  small  amount  of  experience  in  using  Google  scholar  to  find  citing/cited  papers  
Task  Description  
Your  task  is  to:  
1. devise  a  classification  or  regression  problem  to  investigate  using  the  data  set  provided  
–  this  can  be  reproducing,  extending  or  modifying  the  problem  which  is  described  in  
the  paper  related  to  that  dataset;  
2. implement  a  neural  network  in  PyTorch  /  Python  to  solve  the  problem  and  implement  
a  method  to  determine  the  performance  of  the  neural  network;    
3. implement  a  technique  from  the  literature  (paper  selected  as  described  above)  and  
determine  its  benefit  or  lack  of  benefit;  
4. compare  your  results  with  results  published  in  a  research  paper  reporting  results  on  
the  data  set  you  chose  (see  below);  and  
5. write  a  report  on  your  work  
Data  set  
Choose   a  data   set   from   the  datasets  provided.  Wattle  will   have   a  way   for   you   to   select   the  
combination  of  dataset  and  technique.  You  can  modify  the  prediction  task  from  that  originally  
described  in  the  paper.  You  should  find  it  difficult  to  obtain  near  perfect  results,  as  you  will  be  
   
  3  
marked  on  the  steps  taken  to  achieve  better  results  and  how  they  compare  with  your  previous  
results,  not  on  how  high  %  accurate  your  results  are.  The  paper  related  to  the  dataset  may  be  
of   a   similar   age   to   the   technique  paper,  or  will  be  using  unusual  data   so   there  will  be   little  
need  to  try  to  compete  with  the  best  modern  techniques.  
Your   report   must   describe   which   dataset   you   used,   what   modifications   you   made   to   the  
encodings  if  you  needed  to  do  so,  and  cite  the  paper  associated  with  the  data  set  to  provide  
some  context  to  your  results.  You  must  also  cite  the  technique  paper(s)  you  use.  You  should  
also   cite   some   other   papers   which   are   somehow   relevant.   Other   academic   papers   can  
generally  be   found  by  using  Google   scholar.  Google   scholar  will   have   links   to  one  or  more  
paper   repositories.   You   can   get   access   if   you   do   this   from   ANU   campus   as   the   library  
subscribes   to   most   of   the   large   document   repositories.   Some   papers   are   in   multiple  
repositories  so  if  you  cannot  find  a  free  copy  then  ask  your  tutor  for  advice.  From  off  campus,  
if  you  log  into  the  ANU  Virtual  Proxy  server  then  usually  you  can  access  the  same  electronic  
resources   that  would   be   available   on   campus.  You   should   also  mention   in   your   report   any  
different  topologies  and  analyses  used  in  your  experimentation.  
Design  of  a  Problem  
Detail  in  your  report  what  you  want  to  model  in  the  data  set  and  explain  what  the  inputs  and  
outputs  you  will  use  to  develop  the  neural  network  model.  
Implementation  
Choose  an  appropriate  measure  to  report  the  results  produced  by  your  neural  networks.  You  
can  use  the  measure  used  in  the  research  papers  that  have  the  results  of  their  predictions  on  
the  data  set  you  have  used  and  would  like  to  compare  your  results  with  –  if  the  paper  did  not  
do  predictions,  then  the  paper  provides  a  context  for  your  predictions  instead.  Remember  to  
cite  the  papers  in  your  report.    
Report  
The   report   must   be   in   the   style   of   an   academic   paper   and   must   conform   to   the   Springer  
Lecture   Notes   in   Computer   Science   conference   paper   proceedings   format,   but   with   the  
margins   changed   to   2   cm   and   header/footer   to   1   cm.   The   template   for   the   report   can   be  
downloaded   from   Wattle   by   clicking   the   link   named   “AssignmentReportTemplate-­‐‑LNCS-­‐‑
Office2007.zip”  and  needs  the  margins  to  be  modified.   If  you  prefer   to  use  Latex   ,   there   is  a  
link   on   the   ABCs   conference   website   which   takes   you   to   Springer’s   current   pages.   Any  
template  from  there  will  also  need  the  margins  to  be  modified.  Use  the  Springer  citation  style  
as  found  in  the  template  file  you  use.  
Your  report  should  have  a  meaningful  title,  which  indicates  what  you  have  done.  "ʺCOMP8420  
Assignment"ʺ  is  not  a  meaningful  title.  Your  title  and  content  should  NOT  mention  the  course:  
we  are  modelling  the  assignment  so  that  you  are  making  a  conference  paper  submission.  Your  
u   number   should   be   showing   in   your   email   address.   Your   affiliation   would   be   "ʺResearch  
School   of   Computer   Science,   Australian   National   University"ʺ   etc.   Please   do   not   forget   to  
include  your  name  as  the  single  author.  
A  suggested  structure  for  the  report:  
• Abstract  –  A  paragraph  that  summarises  the  work  you  did,  the  results  you  found  and  
whether   it  was   better,   same   or  worse   than   a   published   research   paper   for   the   same  
   
  4  
dataset.  An  abstract  is  similar  to  an  executive  summary  of  the  entire  report.    
• Introduction   –   A   description   of   the   motivation   for   the   choice   of   the   data   set,   the  
problem   that   you   modelled   using   a   neural   network   and   an   outline   of   the  
investigations  that  you  carried  out  using  the  model.  This  section  should  also  include  a  
brief   background   to   the   problem   and   the   methods   used   perform   the   analysis.  
Remember  to  use  citations!  
• Method   –   A   description   of   the   technique(s)   you   implemented   and   details   of   the  
investigations  or  tests  you  conducted  using  the  technique.  
• Results  and  Discussion  –  Presentation  of  results  from  the  investigations  and  detailed  
analysis  of  the  results  including  comparison  of  your  results  with  the  results  published  
in  a  research  paper  on  the  data  set.  Remember  to  use  citations!  
• Conclusion  and  Future  Work  –  A  statement  on  your  findings  and  how  your  work  can  
be   extended   or   how  might   it   be   improved.   Even   if   you   have   conducted   a   thorough  
investigation  there  is  always  work  left  to  do.  Outlining  future  work  is  VERY  important  
as   it   shows   that   you   have   thought   about   the   problem   and   have   a   deeper  
understanding  then  just  stating  a  conclusion.  
A  comment:  you  should  read  papers  in  relevant  areas  of  the  literature  (e.g.  on  similar  topics)  
to  get  more  of  a  feel  for  how  to  write  and  layout  academic  papers.  This  forms  part  of  the  
learning  you  should  get  out  of  advanced  courses  such  as  COMP4xxx/COMP8xxx  given  that  
the  ANU  is  a  research  intensive  University.  
 
Peer  Marking  (optional)  
Conference  papers  need  reviewers,  and  conferences  need  a  Program  Committee.  The  plan  is  
that  all  of  you  are  the  program  committee,  the  program  chair(s)  will  assign  you  some  papers  
to  review  and  you  will  either  do  them  or  not  (the  peer  marking  is  optional).  You  will  get  
marks  for  how  well  you  do  the  reviews.  How  well  means  whether  you  gave  appropriate  
criticism  and  helpful  suggestions  or  not  (a  too  critical  or  too  uncritical  review  can  be  equally  
bad).  I  hope  many  of  you  will  chose  to  do  the  peer  marking.  Any  low  marks  can  be  redeemed  
in  the  final  exam.  There  will  be  more  details  about  the  peer  marking  when  we  get  closer  to  it,  
the  reports  need  to  be  submitted  before  any  peer  marking  is  possible.  
 
 
   
   
  5  
Assessment  Guide  (  /30  i.e.  marked  out  of  30)   worth  15%  of  overall  course  marks  
Abstract  (  /2)  
Clear  and  concise  abstract  summarising  the  work  done  
Method  section  /  Data  Set  and  Model  Design  (  /8)  
Valid  reasons  for  choosing  the  prediction  using  the  data  set  
Good  level  of  problem  complexity  
Clear  and  valid  investigation  aims  
Model  design  that  clearly  serves  the  purpose  of  the  investigations  
Appropriate  measures  used  to  determine  the  performance  of  the  neural  network  and  
predictions  
Good  explanation  of  the  model  design  
Appropriate  choice  for  the  inputs  and  outputs  of  the  prediction  model  with  valid  
reasons  
Evidence  of  good  understanding  of  the  relevant  literature  
Results  and  Discussion  (  /6)  
Good  methods  used  to  evaluate  the  model  including  an  appropriate  split  of  the  train,  
test  /  validation  data  
Good  level  of  detail  used  to  analyse  results  
Conclusion  and  Future  Work  (  /2)  
Appropriate  conclusion  of  the  work  
Appropriate  work  suggested  to  extend  and/or  improve  the  work  
References  and  citations  (  /2)  
References  used  as  appropriate  
Presentation  (  /6)  
Structure  of  the  code  is  legible  and  well  organised  
Good  report  structure  
Report  is  legible,  clear  and  concise  
Clear  presentation  of  results  including  appropriate  use  of  figures,  tables  and  graphs  
Report  conforms  to  required  style  (including  the  use  of  appropriate  language)  and  
length  
Consistent  style  used  for  citing  references  
Correct  grammar  and  spelling    
Implementation  (  /4)  
Evidence  of  good  code  design  -­‐‑  appropriate  level  of  modularity,  encapsulation  and  
reusability  of  the  code  
Code  is  comprehensible  with  appropriate  names  of  coding  items  e.g.  code  files,  
functions,  variables  
Code  executes  without  errors  
Good  comments  in  the  code  

联系我们
  • QQ:99515681
  • 邮箱:99515681@qq.com
  • 工作时间:8:00-23:00
  • 微信:codinghelp2

联系我们 - QQ: 99515681 微信:codinghelp2
© 2014 www.7daixie.com
程序代写网!