5. Yearbook
• prior 2012
• Machine learning was more about SVM, Graphical Models, non-parametric baysian, and simple hacks, e.g., decision tree, naïve bayesian,
regressions, etc
•2012~2013
• Deep Neural Network, big data, mature distributed computing architectures
• Refreshing accuracy record in image recognition tasks
• Feed forward Neural Networks, CNN, RBM
•2014~2015
• Bridging CV, NLP and expanding to other domains
• New architectures and new ML tasks
• RNN, LSTM, RL
•2016~future
• Larger models on CV, NLP, keep expanding to other domains
• Larger systems, larger models on CNN, LSTM, etc
9. Computer Vision (crowded market)
Generic Algorithms & API providers
•Special object recognition (face, etc)
•General object recognition (image search, etc)
•Moving object detection & recognition (pedestrian detection, etc)
•Image understanding (visual QA, artify, etc)
•Video understanding (video search, etc)
Verticals
•Satellite image analysis (understanding civil developments)
•Home & office place security & surveillance
•User interest analysis and high precision targeting (ads)
•ADAS & Autonomous Driving
•Robotics and drones
•The list goes on and on
10. Natural Language Understanding (crowded market)
Generic Algorithms & API providers
•Personal assistance (x.ai, api.ai, etc)
•Chat bots (facebook ecosystem, viv.ai, etc)
•Knowledge understanding (IBM watson, etc)
Verticals
•Customer service
•Travel management
•Financial service
•Smart homes
•Connected cars
•The list goes on and on
12. Distributed Architectures
Generic solvers
•Stochastic Gradient Descent
•Coordinate Descent
•MCMC
•ADMM
•…
Design Choices
•sync or async
•CPU or GPU cluster
•Online training or offline training
•...
Examples
•Parameter Server
•DistBelief
•Tensorflow
Flexible solutions
•CoreOS, etcd, Docker
•Kubernetes
•Pachyderm
•Mesos
•etc
Big data ecosystem
•Spark & tachyon
•yarn
•hdfs
•etc
Deep learning tools
•Tensorflow
•Torch-IPC
•DistBelief
•Petuum
•GraphLab
13. Standalone Toolkits
CPU with GPU speedup
•Torch
•Caffe
•Theano
•Tensorflow
Good for research or small applications
Embedded Support
•Tiny-CNN
•Tensorflow-embedded
Good for phone apps, raspberry Pi, cars, drones,
robotics
What deep learning can do
Who are working on that.
How are they working on that
What are they working on
Pre 2012: NN are powerful, but unstable. A great student that no one can teach effectively.
2012, when I started with deep learning.
Unified ML model or at least part of the unified model
Tools like caffe
2014
Merging effort from vision and language.
2016
Expanding to other domains, robotics, control, recommendation systems
Since DL requires a lot of data & computation, the capability of computing is a competitive advantage