Deep Learning
meetup
Protobuf/Lua (Caffe/Torch)
Python (TensorFlow)
Java (DeepLearning4J)
Programming vs ML
AI Landscape
Stage of AI
Deep Learning history
What is (not) DL
● Deep Learning is not GOFAI
● Deep Learning is different than ML
● Deep Learning doesn't mimic Brain
● Deep Learning is not Statistics
● Deep Learning is not Big Data
● Deep Learning is not made by Data Scientists
● Deep Learning is not ANN or MLP
ML algorithms
Training DL model
ConvNets
32
32
3
Convolution Layer
32x32x3 image
5x5x3 filter
convolve (slide) over all
spatial locations
activation
map
1
28
28
ConvNets
32
32
3
Convolution Layer
activation maps
6
28
For example, if we had 6 5x5 filters, we’ll get 6 separate activation maps:
We processed [32x32x3] volume into [28x28x6] volume.
Q: how many parameters are used ?
A: (5*5*3)*6 = 450 parameters, (5*5*3)*(28*28*6) = ~350K multiplies
28
● InputLayer
● ConvLayer
● ReLuLayer
● PoolLayer
● FullyConnLayer
● SoftMaxLayer
Models
● LeNet - 1998
● AlexNet - ILSVRC 2012
● ZFNet - ILSVRC 2013
● GoogLeNet - ILSVRC 2014
● VGGNet - ILSVRC 2014
● ResNet - ILSVRC 2015
● ILSVRC 2016
Pre-trained Models
Karpathy
● Matlab is so 2012
● Caffe is so 2013
● Theano is so 2014
● Torch is so 2015
● TensorFlow is so 2016
● *PyTorch is so 2017
Links
● DL4J-examples
● Deep Learning Papers
● ConvNetJava
● NVIDIA DIGITS
● Darknet
● Awesome Deep Learning Papers
● Learning for Self-Driving Cars
Links v2
● Graphcore Poplar
● Dlib
● Udacity's Self-Driving Car Simulator
● DL4J Model Zoo
● Trained models for Keras
● DIGITS Object Detection
● DIGITS Semantic Segmentation
Data to train
● Auto Pilot
● Traffic Light Recognition
● Traffic Signs Recognition
● Lane Finding
● Object Recognition
Autonomous car
● Autopilot TensorFlow
● Udacity self-driving car
● CommaAI openpilot
● FUBAR Labs
● Nvidia Autopilot Keras GTA V
● Udacity self-driving car nanodegree
● Traffic Light Recognition
● Traffic Signs Recognition
● Quadcopter Navigation

Hubba Deep Learning