SLIDE 1
Interpretability and Visualization
- f Deep Neural Networks
Interpretability and Visualization of Deep Neural Networks Au Aude - - PowerPoint PPT Presentation
Interpretability and Visualization of Deep Neural Networks Au Aude - - PowerPoint PPT Presentation
Interpretability and Visualization of Deep Neural Networks Au Aude Oliva MI MIT Convoluti tional Neura ral Netw twork rks convolution max-pooling normalization full connected Alexnet Ea Each ch laye yer learns prog ogressive
SLIDE 2
SLIDE 3
SLIDE 4
Convoluti tional Neura ral Netw twork rks
convolution max-pooling normalization full connected
Krizhevsky et al (2012).NIPS
Ea Each ch laye yer learns prog
- gressive
vely y mor
- re com
complex x features Alexnet
SLIDE 5
pl places2.csail.mi mit.edu du
What t did th the netw twork rk learn rn ?
SLIDE 6
Zhou, Khosla, at al (2015). ICLR
Compari ring Object t and Scenes CNNs
SLIDE 7
Data driven approach inspired by Neuroscience: Empirical receptive field
SLIDE 8
Pipeline for r esti timati ting th the Recepti tive Fields:
Goal is to identify which regions of the image lead to the high unit activations.
5000 occluded versions Di Discr crepancy cy map per unit
Zhou, Khosla, at al (2015). ICLR
SLIDE 9
Pipeline for r esti timati ting th the Recepti tive Fields
To To consolidate the information from several images, we ce center the discr crepancy cy map arou
- und the spatial loca
- cation
- n of
- f the
uni unit t tha hat c caus used ed t the m he maximum um a activation f n for t the g he given i en image. e. Th Then we average the re-ce centered discr crepancy cy maps to
- ge
generate t the f final r recept ptive f field o d of e each gi given u unit. .
Zhou, Khosla, at al (2015). ICLR
SLIDE 10
Annota tati ting th the Semanti tics of Units ts
Pool5, unit 76; Label: ocean; Type: scene; Prec ecision: n: 93% 93%
SLIDE 11
Pool5, unit 13; Label: Lamps; Type: object; Pr Precisi sion: 84%
Annota tati ting th the Semanti tics of Units ts
SLIDE 12
Pool5, unit 77; Label: legs; Type: object part; Prec ecision: n: 96% 96%
Annota tati ting th the Semanti tics of Units ts
SLIDE 13
La Layer 1
SLIDE 14
La Layer 2
SLIDE 15
La Layer 4
SLIDE 16
La Layer 5
SLIDE 17
SLIDE 18
SLIDE 19
SLIDE 20
SLIDE 21
SLIDE 22
SLIDE 23
SLIDE 24
SLIDE 25
SLIDE 26
SLIDE 27
SLIDE 28
SLIDE 29
Visualizing Units & Connections
http://people.csail.mit.edu/torralba/research/drawCNN/drawNet.html
SLIDE 30
http://netdissect.csail.mit.edu/
SLIDE 31
Corr Corresp spon
- ndence
ce betw tween deep mod
- dels
an and human man brai ain ?
SLIDE 32
voxe voxel
Vo Voxels within se searchlight
La Layer 3
vs vs.
fMRI fMRI vo voxe xel
Sp Spea earman n Co Correla latio ion
Algorithmic-specific fMRI searchlight analysis
A A spati atial ally unbias ased view of th the relati ations in similari arity ty stru tructu ture be betwe ween m mode dels a and f d fMRI I vs vs.
Cichy, Khosla, Pantazis, Torralba & Oliva, A. (2016). Scientific Reports.
SLIDE 33
Spatiotemporal maps of co correlations be between huma man br brain and d mode model layers
Layers 1-2 Layers 2-4 Layers 5-8
Pari rieta tal Ventra tral
Cichy, Khosla, Pantazis, Torralba & Oliva, A. (2016). Scientific Reports.
SLIDE 34
Co Compari ring Natu tura ral and Arti rtifici cial De Deep Neural al Netwo works
- New f
New fiel elds of ds of exper expertise: se: Cognitive / Clinical / Social / Perceptual Comput Computat ational
- nal Exper
Experiment mental alist st
- St
Stud udyi ying ng t the i he imp mplement ementat ation
- n that works best for
performing specific tasks
- Char
Charact acter erizi zing t ng the net he networ
- rk behavi
k behavior
- r when it is
adapting, compromised or enhanced
- Exp
Explor
- ring
ng t the al he alter ernat natives ves that have not been taken by biological systems
Ra Rados doslaw Cic Cichy Ad Adit itya Kh Khosla An Antonio io To Torralba Bo Bolei Zho Zhou Da Davi vid Ba Bau