Trend  #1:  Scale  driving  Deep  Learning  progress

Nuts  and  bolts  of  building  AI  

applications  using  Deep  Learning

Andrew  Ng

Andrew  Ng Andrew  Ng

Trend  #2:  The  rise  of  end-­to-­end  learning Major  categories  of  DL  models
Learning  with  integer  or  real-­valued  outputs:  
1. General  neural  networks  
2. Sequence  models  (1D  sequences)
• RNN,  GRU,  LSTM,  CTC,  attention  models,  ….
3. Image  models
Learning  with  complex  (e.g.,  string  valued)  outputs:  
• 2D  and  3D  convolutional  networks
4. Advanced/future  tech:
• Unsupervised  learning  (sparse  coding,  ICA,  SFA,  
…),  Reinforcement  learning,  ….  

Andrew  Ng Andrew  Ng

Andrew  Ng
End-­to-­end  learning:  Speech  recognition End-­to-­end  learning:  Autonomous  driving
Traditional  model   Traditional  model  

End-­to-­end  learning End-­to-­end  learning

This  works  well  given  enough  labeled  (audio,  transcript)  data.   Given  the  safety-­critical  requirement  of  autonomous  driving  and  thus  the  need  for  extremely  
high  levels  of  accuracy,  a  pure  end-­to-­end  approach  is  still  challenging  to  get  to  work.  End-­to-­
end  works  only  when  you  have  enough  (x,y)  data  to  learn  function  of  needed  level  of  complexity.  
Andrew  Ng Andrew  Ng

Machine  Learning  Strategy Traditional  train/dev/test  and  bias/variance

Say  you  want  to  build  a  human-­level  speech  recognition  system.  You  split  your  data  
Often  you  will  have  a  lot  of  ideas  for  how  to  improve  an   into  train/dev/test:  
AI  system,  what  do  you  do?  
Training  (60%) Dev  (20%) Test  (20%)

Good  strategy  will  help  avoid  months  of  wasted  effort.

Human  level  error  ……….  1%
“Avoidable  bias”
Training  set  error  ………...  5%
“Variance”
Dev  set  error  ……………..  8%  

Compared  to  earlier  eras,  we  still  talk  about  bias  and  variance,  but  somewhat  less  
about  the  “tradeoff”  between  them.  
Andrew  Ng Andrew  Ng

Andrew  Ng
Basic  recipe  for  machine  learning Automatic  data  synthesis  examples

• OCR
• Text  against  random  backgrounds
Machine  Learning
Bigger  model
Training  error  high? Train  longer (Bias) • Speech  recognition
Yes New  model  architecture
• Synthesize  clean  audio  against  different  background  noise
No
• NLP:  Grammar  correction
More  data
Dev  error  high? Regularization (Variance) • Synthesize  random  grammatical  errors
Yes New  model  architecture
No Sometimes  synthesized  data  that  appears  great  to  human  eyes  is  
Done! actually  very  impoverished  in  the  eyes  of  ML  algorithms,  and  covers  
only  a  minuscule  fraction  of  the  actual  distribution  of  data.  E.g.,  
images  of  cars  extracted  from  video  games.

Andrew  Ng Andrew  Ng

Different  training  and  test  set  distributions Different  training  and  test  set  distributions
Better  way:  Make  the  dev  and  test  sets  come  from  the  same  distribution.
Say  you  want  to  build  a  speech  recognition  system  for  a  new  in-­car  
Training-­Dev   Dev Test  
rearview  mirror  product.  You  have  50,000  hours  of  general  speech   Training  (~50,000h) (20h) (5h) (5h)
data,  and  10  hours  of  in-­car  data.  How  do  you  split  your  data?  This   General  speech  data In-­car  data
is  a  bad way  to  do  it:  
Human  level  error  ………...  1%
Training Dev Test “Avoidable  bias”
Training  error  ……..............  1.1%
General  speech  data  (50,000  hours) In-­car  data  
(10  hours) Overfitting  of  training  set
Training-­Dev  error  ………...  1.5%
Having  mismatched  dev  and  test  distributions  is  not  a  good  idea.   Data  mismatch
Your  team  may  spend  months  optimizing  for  dev  set  performance   Dev  set  error  ……………….  8%
only  to  find  it  doesn’t  work  well  on  the  test  set.   Overfitting  of  dev  set
Test  set  error  ……………….  8.5%  
Andrew  Ng Andrew  Ng

Andrew  Ng
New  recipe  for  machine  learning General  Human/Bias/Variance  analysis  
Bigger  model
Training  error  high? Train  longer (Bias)
Yes New  model  architecture General In-­car
speech  data   speech  data  
No (50,000  hours) (10  hours)
More  data
Train-­Dev  error  high? Regularization (Variance) Performance of   (Carry  out  human  
Yes New  model  architecture Human-­level error   evaluation  to  measure.)
humans
No “Avoidable  bias”
Make  training  data  more               Performance on  
(Insert some  in-­car  data  into  
Dev  error  high? similar  to  test  data.   (Train-­test  data   examples  you’ve  trained   Training  error training  set  to  measure.)
Yes Data  synthesis
mismatch) on
(Domain  adaptation.)
“Variance”/degree  of  
No New  model  architecture
Performance  on   overfitting
examples  you  haven’t   Training-­Dev  error Dev/Test  error
Test  error  high? More  dev  set  data (Overfit  dev   trained  on
Yes
set)
No
Data  mismatch
Done! Andrew  Ng Andrew  Ng

Human  level  performance Quiz:  Medical  imaging

You’ll  often  see  the  fastest  performance  improvements  on  a  task  while  the   Suppose  that  on  an  image  labeling  task:
ML  is  performing  worse  than  humans.  
• Human-­level  performance  is  a  proxy  for  Bayes  optimal  error,  which  we   Typical  human  ………………..…  3%  error
can  never  surpass. Typical  doctor  …………………...  1%  error
• Can  rely  on  human  intuition:  (i)  Have  humans  provide  labeled  data.             Experienced  doctor  …………….  0.7%  error
(ii)  Error  analysis  to  understand  how  humans  got  examples  right.                    
(iii)  Estimate  bias/variance.  E.g.,  On  an  image  recognition  task,  training   Team  of  experienced  doctors  ….  0.5%  error
error  =  8%,  dev  error  =  10%.  What  do  you  do?  Two  cases:  

Human  level  error  ……….  1% Human  level  error  ……….  7.5%
What  is  “human-­level  error”?
“Avoidable  bias” “Avoidable  bias”
Training  set  error  ………...  8% Training  set  error  ………...  8%
“Variance” “Variance” Answer:  For  purpose  of  driving  ML  progress,  0.5%  is  
Dev  set  error  ……………  10%  
Dev  set  error  ……………  10%  
best  answer  since  it’s  closest  to  Bayes  error.  
Focus  on  bias. Focus  on  variance.
Andrew  Ng Andrew  Ng

Andrew  Ng
AI  Product  Management AI  Product  Management
The  availability  of  new  supervised  DL  algorithms  means  we’re  rethinking  the  workflow   How  should  PMs  and  AI  teams  work  together?  Here’s  one  default  split  of  
of  how  to  have  teams  collaborate  to  build  applications  using  DL.  A  Product  Manager   responsibilities:
(PM)  can  help  an  AI  team  prioritize  the  most  fruitful  ML  tasks.  E.g.,  should  you  
improve  speech  performance  with  car  noise,  café  noise,  for  low-­bandwidth  audio,  for   Product  Manager  (PM)   AI  Scientist/Engineer
accented  speech,  or  improve  latency,  reduce  binary  size,  or  something  else?   responsibility responsibility
What  can  AI  do  today?  Some  heuristics  for  PMs:   • Provide  dev/test  sets,  ideally   • Acquire  training  data
• If  a  typical  person  can  do  a  mental  task  with  less  than  one  second  of  thought,  we   drawn  from  same  distribution.
can  probably  automate  it  using  AI  either  now  or  in  the  near  future.   • Develop  system  that  does  well  
• For  any  concrete,  repeated  event  that  we  observe  (e.g.,  whether  user  clicks  on  ad;;   • Provide  evaluation  metric  for   according  to  the  provided  
how  long  it  takes  to  deliver  a  package;;  ….),  we  can  reasonably  try  to  predict  the   learning  algorithm  (accuracy,   metric  on  the  dev/test  data.  
outcome  of  the  next  event  (whether  user  clicks  on  next  ad).   F1,  etc.)  

This  is  a  way  for  the  PM  to  express  

what  ML  task  they  think  will  make  
the  biggest  difference  to  users.  
Andrew  Ng Andrew  Ng

Machine  Learning  Yearning

Book  on  AI/ML  technical  strategy.

Thank  you  for  coming  to
Sign  up  at  http://mlyearning.org this  tutorial!  

Andrew  Ng Andrew  Ng

Andrew  Ng