Upcoming SlideShare
Loading in...5
×
 

Like this? Share it with your network

Share

Deep Learning and its Applications - Computer Vision

on

  • 419 views

Deep Learning and its Applications - Computer Vision Zipfian Academy Meetup ...

Deep Learning and its Applications - Computer Vision Zipfian Academy Meetup

Deep-learning is useful in detecting anomalies like fraud, spam and money laundering; identifying similarities to augment search and text analytics; predicting customer lifetime value and churn; recognizing faces and voices.

The framework's neural nets include restricted Boltzmann machines, deep-belief networks, deep autoencoders, convolutional nets and recursive neural tensor networks.

Statistics

Views

Total Views
419
Views on SlideShare
413
Embed Views
6

Actions

Likes
0
Downloads
15
Comments
0

2 Embeds 6

https://www.linkedin.com 5
https://twitter.com 1

Accessibility

Categories

Upload Details

Uploaded via as Microsoft PowerPoint

Usage Rights

CC Attribution License

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Processing…
Post Comment
Edit your comment

Deep Learning and its Applications - Computer Vision Presentation Transcript

  • 1. The image part with relationship ID rId14 was not found in the file. { Deep  Learning And  Its  Applications:  Computer  Vision Adam  Gibson {  deeplearning4j.org  //  skymind.io  //  zipfian  academy
  • 2. The image part with relationship ID rId14 was not found in the file. •  Object  Recognition •  Image  Categorization •  Scene  Parsing •  Face  Recognition Computer  Vision:  A  Primer
  • 3. The image part with relationship ID rId14 was not found in the file. •  OpenCV   •  SIFT •  Filters/Edge  Detection •  Feature  Extraction What’s  currently  done?
  • 4. The image part with relationship ID rId14 was not found in the file. •  Representation  Learning   •  More  precise  than  hand-­‐‑done   features •  Non-­‐‑linearities  and  higher-­‐‑order   trends •  Pretrain  and  Hessian  Free This  is  manual!
  • 5. The image part with relationship ID rId14 was not found in the file. •  Representation  Learning •  Position  Invariance  with  convolutions •  Semantic  Hashing   Deep  Learning  and  Images
  • 6. The image part with relationship ID rId14 was not found in the file. •  Normal  pixels  –  0-­‐‑255  –   normalization •  Sparse  –  binarization  (depending  on   pixel  presence) Different  kinds  of  images
  • 7. The image part with relationship ID rId14 was not found in the file. •  Faces  =  a  collection  of  images. •  With  persistent  pa_erns  of  pixels. •  Pixel  pa_erns  =  features. •  Nets  learn  to  identify  features  in  data,  to   classify  faces  as  faces  and  label  them:  John  or   Sarah. •  Nets  train  by  reconstructing  faces  from  features   many  times. •  Measuring  their  work  against  a  benchmark. Facial  recognition
  • 8. The image part with relationship ID rId14 was not found in the file. DL4J’s  Facial  Reconstructions
  • 9. The image part with relationship ID rId14 was not found in the file. •  Slices  of  a  feature  space  (Max  pooling) •  Learns  different  portions  for  easily  scalable   and  robust  feature  engineering. Position  Invariance  -­‐‑  Convolutions
  • 10. The image part with relationship ID rId14 was not found in the file. Visual  Example  -­‐‑  Convolutions
  • 11. The image part with relationship ID rId14 was not found in the file. Pen  Strokes
  • 12. The image part with relationship ID rId14 was not found in the file. •  Facebook  uses  facial  recognition  to  make   itself  stickier  and  know  more  about  us. •  Government  agencies  use  it  to  secure   national  borders. •  Video  game  makers  use  it  to  construct  more   realistic  worlds. •  Stores  use  it  to  identify  customers  and  track   behavior. What  are  faces  for?
  • 13. The image part with relationship ID rId14 was not found in the file. •  2  layers  of  neuron-­‐‑like  nodes. •  The  1st  is  the  visible,  or  input,  layer •  The  2nd  is  “hidden.”  It  identifies  features  in  input •  Symmetrically  connected. •  “Restricted”  =  no  visible-­‐‑visible  or  hidden-­‐‑hidden   ties •  All  connections  happen  between  layers. Restricted  Bolgmann   Machines  (RBMs)
  • 14. The image part with relationship ID rId14 was not found in the file. •  A  stack  of  RBMs. •  Each  RBM’s  hidden  layer  à  Next  RBM’s  visible/input   layer.   •  DBNs  learn  more  &  more  complex  features •  Example:   •  1)  Pixels  =  input;   •  2)  H1  learns  an  edge  or  line;   •  3)  H2  learns  a  corner  or  set  of  lines;   •  4)  H3  learns  two  groups  of  lines  forming  an  object   -­‐‑-­‐‑  a  face! •  Final  layer  classifies  feature  groups:  sunset,  elephant,   flower,  John,  Sarah. Deep-­‐‑Belief  Net  (DBN)
  • 15. The image part with relationship ID rId14 was not found in the file. •  2  DBNs. •  1st  DBN  *encodes*  data  into  vector  of  10-­‐‑30   numbers  =  Pre-­‐‑training. •  2nd  DBN  decodes  data  into  original  state. •  Backprop  only  happens  on  2nd  DBN •  2nd  is  the  fine-­‐‑tuning  stage  (reconstruction  entropy). •  Reduces  documents  or  images  to  compact  vectors  . •  Useful  in  search,  QA  and  information  retrieval. Deep  Autoencoder
  • 16. The image part with relationship ID rId14 was not found in the file. Deep  Autoencoder  Architecture
  • 17. The image part with relationship ID rId14 was not found in the file. Image  Search  Results
  • 18. The image part with relationship ID rId14 was not found in the file. •  Top-­‐‑down  &  hierarchical  rather  than  feed-­‐‑forward  (DBNs). •  Handles  sequence-­‐‑based  classification,  windows  of  several   events,  entire  scenes  (multiple  objects). •  Features  themselves  are  vectors.   •  A  tensor  =  a  multi-­‐‑dimensional  matrix,  or  multiple  matrices  of   the  same  size. Recursive  Neural  Tensor  Net
  • 19. The image part with relationship ID rId14 was not found in the file. RNTNs  &  Scene  Composition