Deep Learning for Big Data Applications Improvement and

Deep Learning for Big Data Applications – Improvement and Future Directions by Dr. Tarik Alafif Date: October 2, 2018 1

Outlines • Importance and history of deep learning • Why deep learning? • Deep learning models • CNN • Face detection and face recognition applications using CNN with big data • Experimental results • Conclusion and future directions

In “Nature” 27 January 2016: • “Deep. Mind’s program Alpha. Go beat Fan Hui, the European Go champion, five times out of five in tournament conditions. . . ” • “Alpha. Go was not preprogrammed to play Go: rather, it learned using a general -purpose algorithm that allowed it to interpret the game’s patterns. ” • “…Alpha. Go program applied deep learning in neural networks (convolutional NN) — brain-inspired programs in which connections between layers of simulated neurons are strengthened through examples and experience. ”

Where Deep Learning come from? - How invented

Deep Learning

The Mammalian Visual Cortex Inspires CNN Convolutiona l Neural Net output input

Why Deep Learning • Biological Plausibility – e. g. Visual Cortex • Regular machine learning techniques can’t handle big data learning and may handle to overfitting. • Highly varying functions (Non linear) can be efficiently represented with deep architectures • Less weights/parameters to update than a less efficient shallow representation • Sub-features created in deep architecture can potentially be shared between multiple tasks • Type of Transfer/Multi-task learning

Deep Learning Models • Convolutional Neural Networks (CNN) • Stacked Auto-encoder

Deep Learning Models • Auto-encoder example

Deep Learning Models • Auto-encoder M. Aslan, Z. Hailat, T. Alafif, and X. Chen. “Multi-Channel Multi-Model Feature Learning for Face Recognition”. Pattern Recognition Letters. 2017.

Deep Learning Models (Cont. ) • Generative Adversial Network (GAN)

CNN Introduction The idea of a CNN was brought from the so-called Neocognitron proposed by Fukushima (1980). The Neocognitron makes use of receptive fields, i. e. each neuron is only connected to a certain number of neighboring neurons, in the preceding layer. This idea has been inspired by the discovery of the cat’s visual system by Hubel and Wiesel (1962). David Hubel and Torsten Wiesel were awarded 1981's Nobel Prize. [8]

CNN Architecture

CNN Architecture [5] Marc Aurelio Ranzato, “NEURAL NETS FOR VISION, ” CVPR 2012 Tutorial on Deep Learning Part III.

CNN Architecture Figure: Pooling Figure: Subsampling by factor of 2

CNN Architecture

Example Using CNN: Hand-written Digit Recognition • Input:

MNIST

CNN Architecture -1 -1 +1 2 -1 -1

Current CNN Models

Current CNN Models (Cont. )

Deep Learning Computation Requirements • Central Processing Unit (CPU): example RAM >= 256 • Graphical Processing Unit (GPU): q Tesla k 40 or better q NIVIDIA DGX-1 (Most recent) • Tensor Processing Unit (TPU)

Some Deep Learning Open Source Libraries Name Languag e Link Note Pylearn 2 Python http: //deeplearning. net/software/pyl earn 2/ A machine learning library built on Theano Python http: //deeplearning. net/software/the ano/ A python deep learning library Caffe C++ http: //caffe. berkeleyvision. org/ A deep learning framework by Berkeley Torch Lua http: //torch. ch/ An open source machine learning framework Overfeat Lua http: //cilvr. nyu. edu/doku. php? id=cod e: start A convolutional network image processor Deeplearning 4 j Java http: //deeplearning 4 j. org/ A commercial grade deep learning library Word 2 vec C https: //code. google. com/p/word 2 ve c/ Word embedding framework Glo. Ve C http: //nlp. stanford. edu/projects/glov e/ Word embedding framework Doc 2 vec C https: //radimrehurek. com/gensim/m odels/doc 2 vec. html Language model for paragraphs and documents Stanford. NLP Java http: //nlp. stanford. edu/ A deep learning-based NLP package Tensor. Flow Python http: //www. tensorflow. org A deep learning based python library

My Face Detector and Facial Race Classifier

Solution: Our Datasets Collections and Architecture – LSLF and LSLNF Datasets (Contd. )

Solution: Our Datasets Collections and Architecture – Crowed. Faces and Crowed. Non. Faces Datasets (contd. )

Solution: LSDL Face Detector – LSDL Architecture and Training (contd. )

Solution: LSDL Face Detector – LSDL Face Detection

Solution: LSDL Face Detector – LSDL Face Detection (contd. )

Solution: LSDL Face Detector – LSDL Experimental Evaluation (contd. ) • Our face detection achieves the best performance on AFW dataset and a reasonable performance on FDBB dataset without involving any square detection bounding boxes manual extension or square detection bounding boxes adjustment

Solution: LSDL Face Detector – LSDL Experimental Evaluation (contd. ) • Some of the reported methods are not completely fair since they extend their square detection BBs vertically 20%-40% to fit FDDB elliptical face annotations • Some they predict and adjust the difference between the BBs and ground truth using trained regressors • Our LSDL method tunes for AFW but unfairly panelized in FDDB since it uses elliptical face annotations and the overlap of square and ellipses region is usually smaller than 50% Io. U ratio and consider them as false positives

Solution: LSDL Face Detector – LSDL Qualitative Results

What are Human Races? • We provide different facial characteristics for the four major races in table below:

Introduction: Human Performance on Facial Race Classification (Cont. ) • Challenging facial examples from our CIMN dataset are shown in Figure 1 below

Solution: Our Proposed Model • The winning neuron must receive the value 1 while the other neurons receive the value 0

Solution: Our Proposed Model (Cont. )

Solution: Experiments • Constrained evaluation: • We used the FERET dataset • 2, 695 grayscaled and colored frontal face images (1749 Caucasian, 296 Indian, 429 Mongolian, and 221 Negroid) • Confusion matrix:

Solution: Experiments (Cont. )

Solution: Experiments (Cont. ) • Unconstrained evaluation (Cont. ): • Confusion matrix

Solution: Experiments (Cont. ) • Unconstrained evaluation (Cont. ):

Conclusion • Vision using deep learning technologies mimic human vision intelligently • Deep learning plays a big part in the AI and computer vision areas • Deep learning can be extended to solve many problem in other different areas

Future Computer Vision Research Works and Directions • Deep learning theories and applications • Object detection and tracking • Biometrics • Crowed scenes

Future Computer Vision Research Works and Directions