Final presentation and demo
1Record a demo video before the calss
2Mon(Apr.29) and Wed(May 1)
30min for each group.
3 submit code and presentation and report
Review of Previous class
Humand vs deep learning
Pedestrian Detection:behavior
...
Outline
Semantic segmentation/annotation, GAN image synthesis, train CNN with synthetic data, Imitation learing.
1 Milestones
1943-Neural networks;
1957-1998;
2006-2018: dbn, IMAGE NET, ALEXNET, DROPOUT, GANs, deepface, bert.
Tools:
CUDA,caffe 2014(C++) facebook, Tensorflow, Pytorch
Obeject detection
R-CNN--Overfeat-multibox,fast R-CNN-yolo-SSD-FEATURE PYRAMID NET(fpn)=reTINAnET-mASK R-CNN-FefineDet-M2Det
M2Det:https://arxiv.org/pdf/1811.04533.pdf
DeepLab family
fully convolutional networks(fcns) feplace fully-connected layers with convolutional layers.
conditional random fields(CRFs) to caputre local and long-range dependencies within an image to refine the prediction map.
Astrous rate:percepiton of view
Semantic segmentation
Annotation with Polygon-RNN++: interactive annotation tool, annotate your datasets much faster
semi supvised,
GAN:generative adversarial network ,from supervised--semi-unsupervised training.
see:https://skymind.ai/wiki/generative-adversarial-network-gan
NVIDIA 2018. synthetic views
GAN image synthesis: synthesis fro label to image
https://tcwang0509.github.io/vid2vid/
https://tcwang0509.github.io/vid2vid/images/teaser.gif
Train CNN with Synthetic Data
domain randomization;
flying distractors:
https://arxiv.org/pdf/1804.06516.pdf
DAWNBench benchmark suite: increase the learning rate and decrease moment.
Reinforcement Learning (RL) https://en.wikipedia.org/wiki/Reinforcement_learning
compared with different machine learning:https://towardsdatascience.com/types-of-machine-learning-algorithms-you-should-know-953a08248861
cart pole problem:https://towardsdatascience.com/cartpole-introduction-to-reinforcement-learning-ed0eb5b58288
https://groups.csail.mit.edu/locomotion/
DL: general purpose framework for representation learning
gieven an objective, learn representation,directly from raw input
RL:for decision making
RL:defines the objective
DL:gives the mechanism
RL+dl =GENERAL INTELLIGENCE
RL fundamentals:
problems: policy leraning, reawrad value function learning; environment model learing(know current sate and action, predict next sate and reward)
charactersitics:close system with limited staes and actions;no direct rules/instrucations to action, you don't apply any,,
Problem formulation:
experiment(observations, actions, rewards)
sate is decied by experience
in a fully observed environment: markov assumption st=f(ot)
https://en.wikipedia.org/wiki/Markov_decision_process
value based RL
https://web.stanford.edu/~anayebi/projects/CS_239_Final_Project_Writeup.pdf
https://selfdrivingcars.mit.edu/deeptraffic/
https://www.tesla.com/autopilot