Int Introductio ion t n to Deep Deep Lea earn rning Prof. Leal-Taixé and Prof. Niessner 1
The The Te Team Lecturers Prof. Dr. Laura Prof. Dr. Matthias Leal-Taixé Niessner Tutors Patrick Andreas Dendorfer Rössler Prof. Leal-Taixé and Prof. Niessner 2
Wh What at is Com omputer er Vi Vision on? First defined in the 60s in artificial intelligence groups • “Mimic the human visual system” • Center block of robotic intelligence • Prof. Leal-Taixé and Prof. Niessner 3
Prof. Leal-Taixé and Prof. Niessner 4
So Some de decade des l later… Computer Vision Prof. Leal-Taixé and Prof. Niessner 5
Engineering Computer Mathematics science Artificial Robotics Intelligence ML NLP Algorithms Speech Optimization Computer Optics Vision Image Neuroscience processing Physics Biology Psychology Prof. Leal-Taixé and Prof. Niessner 6
Engineering Computer Mathematics science Artificial Robotics Intelligence ML NLP Algorithms Speech Optimization Computer Optics Vision Image Neuroscience processing Physics Biology Psychology Prof. Leal-Taixé and Prof. Niessner 7
Engineering Computer Mathematics science Artificial Robotics Intelligence ML NLP Algorithms Speech Optimization Computer Optics Vision Image Neuroscience processing Physics Biology Psychology Prof. Leal-Taixé and Prof. Niessner 8
Engineering Computer Mathematics science Artificial Robotics Intelligence ML NLP Algorithms Speech Optimization Computer Optics Vision Image Neuroscience processing Physics Biology Psychology Prof. Leal-Taixé and Prof. Niessner 9
Im Image classification Pre 2012 Prof. Leal-Taixé and Prof. Niessner 10
Im Image classification Awesome A magic box Open the box Become magicians Post 2012 Prof. Leal-Taixé and Prof. Niessner 11
Why Why Deep Le Lear arning ning? Prof. Leal-Taixé and Prof. Niessner 12
De Deep p Le Learning g Histo tory Prof. Leal-Taixé and Prof. Niessner 13
The The empire str trike kes back ck Prof. Leal-Taixé and Prof. Niessner 14
Wh What at has as chan anged? ed? • MNIST digit 1988 recognition dataset LeCun 10 7 pixels used in • et al. training 2012 ImageNet image • Krizhevsky recognition dataset 10 14 pixels used in et al. • training Prof. Leal-Taixé and Prof. Niessner 15
What Wh at made ade this pos ossible? e? Big Data Hardware Deep Models know Models are Models are where to learn from trainable complex Prof. Leal-Taixé and Prof. Niessner 16
De Deep p Le Learning g nowadays ys Emoticon suggestion AlphaGo Machine translation Prof. Leal-Taixé and Prof. Niessner 18
De Deep p Le Learning g nowadays ys Self-driving cars Prof. Leal-Taixé and Prof. Niessner 19
De Deep p Le Learning g nowadays ys Healthcare, cancer detection Prof. Leal-Taixé and Prof. Niessner 20
De Deep p Le Learning g market […]market research report Deep Learning Market […] Global • Forecasts to 2022", the deep learning market is expected to be worth USD D 1,7 ,722.9 Million by y 2022. Prof. Leal-Taixé and Prof. Niessner 26
De Deep p Le Learning g at t TU TUM S. Caelles, K.K. Maninis, J. Pont-Tuset, L. Leal-Taixé, D. Cremers, and L. Van Gool. One-Shot Video Object Segmentation, CVPR 2017. Prof. Leal-Taixé and Prof. Niessner 27
De Deep p Le Learning g at t TU TUM Prof. Leal-Taixé and Prof. Niessner 28
De Deep p Le Learning g at t TU TUM CC3 CC2 CC1 Reshape Conv+BN+ReLU Pooling Upsample Concat Score DDFF Prof. Leal-Taixé and Prof. Niessner 29
Comp Computer er Vis ision ion at TUM ScanNet Stats: -Kinect-style RGB-D sensors -1513 scans of 3D environments -2.5 Mio RGB-D frames -Dense 3D, crowd-source MTurk labels -Annotations projected to 2D frames ScanNet: Dai, Chang, Savva, Halber, Funkhouser, Niessner. , CVPR 2017. Prof. Leal-Taixé and Prof. Niessner 30
De Deep p Le Learning g at t TU TUM Map Photo Prof. Leal-Taixé and Prof. Niessner 31
Int Introductio ion t n to D Deep Le Lear arning ning Prof. Leal-Taixé and Prof. Niessner 32
Ab About t the the lectu ture Theory: 11 lectures • Every Thursday 18-20h (MI HS 1) • Practice: 4 exercises, practical sessions • Every Tuesday 18-20h (Interim HS1) • January 31 st : guest lecture by tba • https://dvl.in.tum.de/lectures/i2dl-ws18.html Prof. Leal-Taixé and Prof. Niessner 33
Grading syst Gra g system Exam: tba tba • Review: 2 review sessions • Important: no retake exam • Practice: 4 exercises (Tuesdays) • Bonus 0.3 + questions in the final exam • https://dvl.in.tum.de/lectures/i2dl-ws18.html Prof. Leal-Taixé and Prof. Niessner 34
Ex Exer ercis ise e lec ectures es Tuesday lecture 1: Exercise submission system will • be explained, no not to be missed ! Tuesday lecture 2: DL math background • Tuesday lecture 3: Python introduction • Prof. Leal-Taixé and Prof. Niessner 35
Machine Learning Optimization basics Introduction Introduction to to Deep NN Learning CNN Back- propagation RNN Prof. Leal-Taixé and Prof. Niessner 36
Sli Slides • All material will be uploaded on Moodle • Questions regarding the syllabus, exercises or contents of the lecture, use Moodle! • Questions regarding organization of the course: i2dl@dvl.in.tum.de • Emails to our individual addresses will not be answered. Prof. Leal-Taixé and Prof. Niessner 37
De Deep p Le Learning g at t TU TUM DL for DL in Medical Robotics Applicat. (Bä Bäuml) (Me Menze) Intro to Machine Deep Learning Learning (Gü Günnema mann) DL for DL for Vision Physics (Ni Niessner, , (Th Thuerey) Le Leal al-Ta Taixe) Prof. Leal-Taixé and Prof. Niessner 38
Ma Machine chine Le Learning ning Prof. Leal-Taixé and Prof. Niessner 39
Mac Machine e lear earning Task Prof. Leal-Taixé and Prof. Niessner 40
Im Image classification Prof. Leal-Taixé and Prof. Niessner 41
Pose Appearance Illumination Prof. Leal-Taixé and Prof. Niessner 42
Im Image classification Occlusions Prof. Leal-Taixé and Prof. Niessner 43
Im Image classification Background clutter Prof. Leal-Taixé and Prof. Niessner 44
Image classification Im Representation Prof. Leal-Taixé and Prof. Niessner 45
Machine Mac e lear earning • How can we learn to perform image classification? Task Experience Image Data classification Prof. Leal-Taixé and Prof. Niessner 46
Mac Machine e lear earning Unsupervised learning Supervised learning No label or target class • • Find out properties of the structure of the data Clustering (k-means, • PCA) Prof. Leal-Taixé and Prof. Niessner 47
Mac Machine e lear earning Unsupervised learning Supervised learning Prof. Leal-Taixé and Prof. Niessner 48
Mac Machine e lear earning Unsupervised learning Supervised learning • Labels or target classes Prof. Leal-Taixé and Prof. Niessner 49
Mac Machine e lear earning Unsupervised learning Supervised learning CAT DOG DOG CAT CAT DOG Prof. Leal-Taixé and Prof. Niessner 50
Machine Mac e lear earning • How can we learn to perform image classification? Experience Underlying assumption that train and test data come Training data Test data Data from the same distribution Prof. Leal-Taixé and Prof. Niessner 51
Mac Machine e lear earning Unsupervised learning Supervised learning Reinforcement learning interaction Agents Environment Prof. Leal-Taixé and Prof. Niessner 52
Mac Machine e lear earning Unsupervised learning Supervised learning Reinforcement learning reward Agents Environment Prof. Leal-Taixé and Prof. Niessner 53
Mac Machine e lear earning • How can we learn to perform image classification? Task Experience Performance Image Data measure classification Accuracy Prof. Leal-Taixé and Prof. Niessner 54
A simple le cla lassifier Prof. Leal-Taixé and Prof. Niessner 55
Ne Neare rest st Ne Neigh ghbo bor ? Prof. Leal-Taixé and Prof. Niessner 56
Ne Neare rest st Ne Neigh ghbo bor NN classifier = dog distance Prof. Leal-Taixé and Prof. Niessner 57
Ne Neare rest st Ne Neigh ghbo bor k-NN classifier = cat distance Prof. Leal-Taixé and Prof. Niessner 58
Ne Neare rest st Ne Neigh ghbo bor What is the performance on training data for NN classifier? What classifier is more likely to perform best on test data? Courtesy of Stanford course cs231n Prof. Leal-Taixé and Prof. Niessner 59
Ne Neare rest st Ne Neigh ghbo bor Distance (L1, L2) • Hyperpar aram ameters k (number of neighbors) • These parameters are problem dependent. • How do we choose these hyperparameters? Prof. Leal-Taixé and Prof. Niessner 60
Cr Cros oss valid idation ion train Run 1 validation Run 2 Run 3 Run 4 Run 5 Split the trai aini ning ng dat ata a into N folds Prof. Leal-Taixé and Prof. Niessner 61
Recommend
More recommend
