New neural network architectures
![Network](https://habrastorage.org/webt/dn/iq/zn/dniqznbdtqf4dnmhqyuaa19rhrm.jpeg)
Previous article “ Neural networks. Where does it all go ? ”
This article briefly discusses some neural network architectures, mainly on the task of detecting objects , in order to find (or at least try to find) future directions in this rapidly developing field.
The article does not pretend to be comprehensive and have a good understanding of the diagonal read articles. The author is sure that while he was writing this article, many more new architectures appeared. For example, see here: https://paperswithcode.com/area/computer-vision .
Object Detection in 20 Years - A great overview of 400+ articles for detecting objects over 20 years.
The Neural Network Zoo is a zoo of neural networks whose contents are constantly changing.
An interesting video with recommendations on how to design a neural network: “ How to Design a Neural Network ”.
Efficientnet
![Efficientnet](https://habrastorage.org/webt/k_/al/kx/k_alkxw1ztkaiv3ll6w26n2dwno.png)
EfficientNet — , (, scaling) ( ) , . (compound scaling method), // . «Neural Architecture Search» (NAS, 1, 2, ) EfficientNets.
- «EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks»
- TensorFlow
- 1, 2.1, 2.2, 3
EfficientDet
![Efficientdet](https://habrastorage.org/webt/g3/i-/-t/g3i--thpw8hnbazkvc0z77qsv4w.png)
EfficientDet . . EfficientNet , BiFPN, «» / .
![EfficientDet architecture](https://habrastorage.org/webt/mu/ze/io/muzeio405gjtptbivi8y43hxz1w.png)
— EfficientDet == EfficientNet + BiFPN + /
- «EfficientDet: Scalable and Efficient Object Detection»
- TensorFlow
- PyTorch
- 1
SpineNet
![Spinenet](https://habrastorage.org/webt/gg/ms/q-/ggmsq-np6bieh3ffqllqd3lwpd0.png)
SpineNet . Google Research , state-of-the-art (SOTA) .
. , ( ). () - « » (Convolutional Encoder-Decoder Neural Network). « » . , – (). (backbone model), , , . , « » , ().
SpineNet - ( ). (Neural Architecture Search, NAS). SpineNet (Average Precision, AP). .
![Building scale-permuted network by permuting ResNet](https://habrastorage.org/webt/q0/8v/tt/q08vttva1cwtuhxkh_zyfkfbyfc.jpeg)
– ResNet (ResNet-50-FPN )
- «SpineNet: Learning Scale-Permuted Backbone for Recognition and Localization»
- «SpineNet: »
- SpineNet online demo
CenterNet
![Centernet](https://habrastorage.org/webt/rx/qq/x5/rxqqx5qi7xk9tvvd6xp93jdg67o.png)
CenterNet CornerNet-Lite 2019 (c « »).
CenterNet , , . , , 3D-, , .. (image features) . , (heatmap). . . CenterNet 3D .
3 : , .
![CenterNet diagram](https://habrastorage.org/webt/iu/1q/zc/iu1qzceryzraxuyf7juz2nhhs78.png)
– CenterNet
CornerNet CenterNet. CornerNet , : (bounding box). (anchor box), SSD YOLO, . « », .
CornerNet «corner pooling» . CenterNet «center pooling» .
![CornerNet top-left corner pooling layer](https://habrastorage.org/webt/0q/hx/ri/0qhxrirsmkplh3zgpf6ccywbphy.jpeg)
– «corner pooling» . «max-pooling» «max-pooling». (feature maps) .
- CenterNet «CenterNet: Keypoint Triplets for Object Detection» + 1 + 2
- CornerNet-Lite «CornerNet-Lite: Efficient Keypoint Based Object Detection» +
- CornerNet: «CornerNet: Detecting Objects as Paired Keypoints» + +
ThunderNet
![Thundernet](https://habrastorage.org/webt/j9/-e/ii/j9-eiiuavs9tnkzjr5neoq46zbk.png)
ThunderNet . . , , ARM ( ) 24.1 fps (frames per second, ) MobileNet-SSD.
- «ThunderNet: Towards Real-time Generic Object Detection»
- GitHub
![Thundernet architecture](https://habrastorage.org/webt/zn/jk/fs/znjkfsb-mf6kqkds7l-fu-mtyio.png)
— ThunderNet
CSPNet
![Cspnet](https://habrastorage.org/webt/hm/-s/ds/hm-sdsxwhdphwotnjv_zlaarocs.png)
CSPNet (Cross Stage Partial Network) Darknet, . , (residual neural networks, ResNet). , . , . CSPNet , . , CSPNet (feature pyramid network, FPN).
- «CSPNet: A New Backbone that can Enhance Learning Capability of CNN»
- 1, 2
![Feature pyramid network, FPN](https://habrastorage.org/webt/xn/qh/z2/xnqhz26rbgvoqqihj44bixu7ar0.png)
—
DenseNet
![A dense block with 5 layers and growth rate 4](https://habrastorage.org/webt/ye/ac/ml/yeacmll5jztgjy8h1nkfhrdoupy.jpeg)
— DenseNet c 5 k = 4. .
![DenseNet with three dense blocks](https://habrastorage.org/webt/fu/gq/fz/fugqfz_56gkeyzco8alf6qxc1tu.jpeg)
— DenseNet
DenseNet (Densely Connected Convolutional Network) 2017 . ResNet (Deep Residual Network) , CNN . (dense) , . , , ResNet, («») , (, channel-wise concatenation) . DenseNet , . , DenseNet .
- «Densely Connected Convolutional Networks»
- Keras + CoLab
- Torch ,
SAUNet
SAUNet (Shape Attentive U-Net) 2020 , .
![SAUNet architecture](https://habrastorage.org/webt/ae/rt/e6/aerte6nxnj0qhkjhtfu7pujjf8q.jpeg)
— SAUNet : (texture stream); (gated shape stream). U-Net DenseNet-121 ( DenseNet), U-Net « » (dual attention decoder block).
SAUNet : U-Net, DenseNet, Gated-SCNN Squeeze-and-Excitation Networks.
- «SAUNet: Shape Attentive U-Net for Interpretable Medical Image Segmentation»
- PyTorch
DetNASNet
![Detnasnet architecture](https://habrastorage.org/webt/bv/8g/oc/bv8gocnmykfvolrhqze5uzqgkmy.jpeg)
— DetNASNet
, , , (object detection) (image classification) . DetNASNet Neural Architecture Search (NAS) , . , NAS, .. , . 44 GPU- COCO. , ResNet-101, FLOP-.
- «DetNAS: Backbone Search for Object Detection»
- PyTorch
SM-NAS
![SM-NAS AP](https://habrastorage.org/webt/qh/qc/5y/qhqc5ysdwrao9icbxg3pur7pcfu.jpeg)
— () (mAP) COCO.
SM-NAS Structural-to-Modular NAS (SM-NAS): ; .
- .
- «SM-NAS: Structural-to-Modular Neural Architecture Search for Object Detection»
AmoebaNet
![AmoebaNet-A architecture](https://habrastorage.org/webt/b6/bw/fy/b6bwfygmsvhs_bzqldakql0p8em.jpeg)
— AmoebaNet-A. . «Normal Cell». «Reduction Cell».
AmoebaNet . AmoebaNet . AmoebaNet (search space), NASNet. TPU (Tensor Processing Units) .
- «Regularized Evolution for Image Classifier Architecture Search»
Graph Neural Network
![Graph Neural Network](https://habrastorage.org/webt/ag/2j/dx/ag2jdxr0s9sh4vcarxiyooyhaze.png)
Graph Neural Network , — , — . , .. - . . : PyTorch Geometric PyTorch, Graph Nets TensorFlow, Deep Graph .
- CoLab
- by Siraj Raval
- DGL (Deep Graph Library)
Growing Neural Cellular Automata
![Growing Neural Cellular Automata](https://habrastorage.org/webt/fc/97/6x/fc976xo9zngc8rm6as3-fstc5b4.jpeg)
—
Growing Neural Cellular Automata , «» . (, ..). . «» 16 . «» , , JPEG (), MP3 (), MPEG () ZIP ().
, , () .
- Distill
- Colab notebook
- by Yannic Kilcher
- «»
Spiking neural network . . 1952 , . , . . , , , .
- (ru, en)
- PapersWithCode.com
DPM
DPM, Deformable Part Model detector, . ( CoLab) (HOG). - 2009 , , «Object Detection in 20 Years: A Survey», Integral Channel Features (ICF), .
«» , , , . , «Deformable Part Models are Convolutional Neural Networks» DPM .
- «Deformable Part Models are Convolutional Neural Networks» + MatLab Caffe
:
- , AutoML, « », Neural Architecture Search (NAS NASNet);
- (attention mechanism), ;
- « » , (backbone) ;
- , state-of-the-art (SOTA) .
. , . , , . « » , — , .
PS I recommend the video blog ML Tokyo , in which the author explains and makes neural networks on Keras. His CNN seminar is just what a novice "neurocoder" like me needs.
Thank you for the attention!