El título "Leer artículos para usted". Marzo de 2020. Parte 2

Hola Habr!

Continuamos publicando reseñas de artículos científicos de miembros de la comunidad Open Data Science del canal #article_essense. Si quieres recibirlos antes que los demás, ¡únete a la comunidad ! La primera parte de la asamblea de revisión de marzo se publicó anteriormente .

:

1. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis (UC Berkeley, Google Research, UC San Diego, 2020)
2. Scene Text Recognition via Transformer (China, 2020)
3. PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization (Imperial College London, Google Research, 2019)
4. Lagrangian Neural Networks (Princeton, Oregon, Google, Flatiron, 2020)
5. Deformable Style Transfer (Chicago, USA, 2020)
6. Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need? (MIT, Google, 2020)
7. Attentive CutMix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification (Carnegie Mellon University, USA, 2020)

1. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis

: Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, Ren Ng (UC Berkeley, Google Research, UC San Diego, 2020)
:: GitHub project :: Video :: Blog
: ( belskikh)

3 ( 25 ) , - .

:

1. 3 2 .
2. 3 (x, y, z) , MLP, (r, g, b) .
3. , 2 2 .

, MLP, L2 3 .

, :

1. , x, y, z. : MLP x, y, z 8 (ReLU, 256 ) 256- . 4 MLP (ReLU, 128 ), RGB .
2. Positional Encoding . , , . ( ) positional encoding MLP
3. , , "", "". Hierarchical volume sampling.

. L2 GT "" "" . 1-2 1xV100.

2. Scene Text Recognition via Transformer

: Xinjie Feng, Hongxun Yao, Yuankai Qi, Jun Zhang, Shengping Zhang (China, 2020)
:: GitHub project
: ( belskikh)

: , .. .

Optical Character Recognition (OCR), ResNet Transformer. , ( ..).

:

2. 9696
3. 4 ResNet ( 1/16, )
4. HxWxC (6x6x1024) H*WxC (36x1024), FC , ( 36x256)
5. Word embedding ( 256)
6. .

, "" , , . .

4 ResNet-101 ( torchvision). :

• 4 , -;
• multi-head self attention ;
• position-wise ;
• 4 , -, - — multi-head attention ;
• - residual layer norm;
• ( , ).

SynthText RRC-ArT. Tesla P40 ( ).

"" ( -) / .

(. ):

• , , ;
• .

:

1. -.
2. .
3. - / .

3. PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization

: (Imperial College London, Google Research, 2019)

: ( yorko)

extractive abstractive. – , – , , , .

PEGASUS ( – Imperial College, – Google Research) abstractive summarization Gap Sentences Generation objective . Masked Language Modeling, BERT & Co. , – . , abstractive self-supervised objective, , . extractive- , – .

. , reverse-engineering. : 3 , [MASK1], [MASK2]. “-” – . , Gap Sentences Generation objective MLM BERT, , GSG , -MLM . -MLM :).

() , “” . 3 : (Random), (Lead) Principal, ROUGE1-F1. ( 30%) , .

Principal , Lead, .. bias, .

: C4 (Colossal and Cleaned version of Common Crawl, 750 Gb), HugeNews – 3.8 Tb ( , ).

12 ( , , , reddit ..), SOTA ( , . 2019). , , 1000 .

Appendix E , 10 SOTA ( , 12 ):

– , “” , , Chelsea “Jose Mourinho’s side” “The Blues”. , . - – .

: , - – , . . “big data” , .

– abstractive . ROUGE – , . ROUGE , / - . , . , SOTA .

4. Lagrangian Neural Networks

: Miles Cranmer, Sam Greydanus, Stephan Hoyer, Peter Battaglia, David Spergel, Shirley Ho (Princeton, Oregon, Google, Flatiron, 2020)
:: GitHub project :: Blog
: ( graviton)

, "" ( ) . , . "" , , , , 2- .

:
— . , (Hamiltonian Neural Networks Deep Lagrangian Networks), . :

• Hamiltonian Neural Networks — "" , -, ;
• Deep Lagrangian Networks — , .. , , .

( ). .. , , , , , . .

:

• , , .
• , — ( , , ).
• , , (), .

(-) , jax.

4- 500 . .

ReLU , .. 2- . ReLU^2, ReLU^3, tanh, sigmoid softplus. softplus .

• .
• . , : 0.4% 8% (. ).

5. Deformable Style Transfer

: Sunnie S. Y. Kim, Nicholas Kolkin, Jason Salavon, Gregory Shakhnarovich (Chicago, USA, 2020)

: ( digitman)

, . , . , .

DST(deformable style transfer) "" "", : , - (, ). — , . , ( ).

( ), - , - . NBB (neural best-buddies): CNN 2- , , "" , , k (80), k 2- , .

, . dense flow field 'I' warped W(I, θ). thin-plate spline interpolation. , , θ — .

: — , f — , g — , h — .

, : Gatys ( Gramm matrix vgg , — ) STROTSS ( 3- , ). , . DST , :

$$

$$L_{style}(I_s,X) + L_{style}(I_s,W(X,\theta)).$$

I_s X ( ) I_s W(X, θ). .

$$

$$L_{warp}(P,P',\theta)=\frac{1}{k}\sum^k_{i=1}||(p_i + \theta_i) - p'_i||_2$$

— deformation loss k , p_i p_i' — source target . θ. , p_i' p_i, , , - source target . "total variation norm of the 2D warp field". .

$$

$$R_{TV}(f)=\frac{1}{W H}\sum^W_{i=1}\sum^H_{j=1}||f_{i+1, j} - f_{i,j}||_1 + ||f_{i, j+1} - f_{i,j}||_1$$

:

$$

$$L(X, \theta, I_c, I_s, P) = \alpha L_{content}(I_c, X) + L_{style}(I_s,X) + L_{style}(I_s,W(X,\theta)) \\+ \beta L_{warp}(P, P', \theta) + \gamma R_{TV}(f_{theta}),$$

α β , γ — . ( ) — warped stylized image W(I, θ).

, , , . : , , STROTSS, STROTSS, DST, , DST.

user study AMT , "" , "" .

6. Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need?

: Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B. Tenenbaum, Phillip Isola (MIT, Google, 2020)
:: GitHub project
: ( belskikh)

few-show learning, . , 3%, , meta-learning .

-, , - - . . - , - - . Few-shot learning -.

, :

1. ( , - , ).
2. - .
3. - (1-5) , .

self-distillation Born-again . , - . - GT , KL — .

, . , (, , ) -.

ResNet-12 SeResNet-12 miniImageNet, tiered- ImageNet, CIFAR-FS, FC100.

kNN, , .

7. Attentive CutMix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification

: Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, Marios Savvides (Carnegie Mellon University, USA, 2020)

: ( artgor)

cutmix. , , attention maps, . , . CIFAR-10, CIFAR-100, ImageNet (!) ResNet, DenseNet, EfficientNet. + 1.5% .

, , :

• , , ;
• feature map, ;
• , .

: , CAM CutMix, Chris Kaggle::Bengali.AI Handwritten Grapheme Classification.

pytorch:

• CIFAR-10 — 80 , batch size 32, learning rate 1e-3, weight decay 1e-5;
• CIFAR-100 — 120 , batch size 32, learning rate 1e-3, weight decay 1e-5;
• ImageNet — 100 ResNet DenseNet, 180 EfficientNet, batch size 64, learning rate 1e-3

Ablation study

, — , . 1 15. 6 .