Brave machine's tomorrow nazli temur
1 like228 views
The document discusses the advancements in computer vision and the development of unmanned vehicles, highlighting the desire for machines to mimic human capabilities and consciousness. It outlines significant milestones in the evolution of combat robots, explores the impact of deep learning, and emphasizes the importance of data quality in computer vision tasks. Furthermore, it mentions the ethical considerations surrounding AI and advocates for innovation in technology to drive development.
Brave machine's tomorrow nazli temur
- 1. COMPUTER VISION Presented by Nazli Temur TODAY'S TECH BRAVE MACHINES'TOMORROW
- 2. COVERED TODAY A BRIEF OUTLINE The State of Today's Tech Creation of Unmanned Vehicles The Philosophy of the Future The Digital Age with Deep Learning Recent Developments-problems on AI,DL Innovation Index of Turkey Conclusion
- 3. THE STATE OF TODAY'S TECH WHERE WE ARE TODAY We are at the sky, at the space and more. There is no undiscovered part of the world land and we touch every point either by our steps or our machines. We have control over the world! We have control over health, public identity and roads that we have never passed before. we want our machines to be mobile, sensitive and self-aware. As we create them, implicitly we want them at least identical to us.
- 4. CREATION OF UNMANNED VEHICLES IMPOSSIBLE THINGS TO DO LIST OF D.A.R.P.A. - Mimicking human body - Seeing like human - Climbing Stairs - Opening Doors - Climbing like human - Distinguishing enemy and friend - Being Dispensable - As agile as a soldier
- 5. Creation of Unmanned Vehicles CRAWLING Can not Think Can not Climb Can See FLYING Exploration and suicide bombing from the sky CLIMBING Being able to go anywhere that human being and animals can reach
- 6. TIMELINE OF FIGHTING ROBOTS BEFORE 2008 GOLIATH (1942) Wired, crawling, suicide tank, Used @ II.World War With 65 Kilos of bomb 1 time usage PREDATOR (1995) Remote controlld UAV Used @Bosnia Kosova So called watching purpose 500 kg a commander,a camera ARMED PREDATOR (2000) Remote controlld UAV Used @Afghanistan equiped with Hellfire missiles for Ladin, @10000 feet with sound speed,silent SWORDS (2007) Remote controlled land vehicle with camera, Used @Iraq M249 Gun, Killer Droid MAARS (2008) Hardly Armed Crawler, 3 cameras, Strong communcate 360 degree head rotation in 1 second. ARMED PREDATOR (2008) Used @Uraq Sadr City Named as SHADOW Milis discovery and shoot the target via missiles
- 7. WHERE DO WE GO NEXT? HOW DO WE GET THERE?
- 8. Creation of Unmanned Vehicles HUMANOID ROBOTS Better capabilities of human such as mobilities and behaviours FULLY AUTONOMOUS Independent and quick decision making and reasoning. REACTIVE Not only making decisions but also taking actions that complies with human ethics.
- 9. TIMELINE OF FIGHTING ROBOTS AFTER 2005 BIG DOG Mobility and Agility as animals, No injury risk, can walk and climb robust stance and hold, 150 kg load, GECKO TYPE (2004) Can go anywhere, vertical climbing on trees, buildings can stay on any surface purpose:discovery, observation,spy STICKY BOT inspired from lizard,localised even at ceiling and waits for victim. Can stay there without moving for days even weeks HULC CYBORG Wearable Load Carrier Skeleton. Mimics each human moves and loads all, Saves %15 O2 consumption, titanium material SARKOS External skeleton that can take command direct from human brain. This can create a more powerful, super fast robot soldier FULL AUTONOMY Independent decisio making and learning free will
- 10. D.A.R.P.A. wonders when to authorise robots with killing decision They have planned to have autonomous war robots by 2025 For DARPA, Achieving Artificial Intelligence is Crutial...
- 11. READYtoMARKET
- 12. NON-VIOLENT JAPANESE ROBOTS NEURAL NETS ON THE STAGE This robot can learn activities same as human do. Control system is achieved by neural networks, This network can dynamically be updated and it contains memories. It can organise its network structure and learn like a baby. Artificial intelligence is mostly about learning.
- 13. LEARN LIKE HUMAN Behaving like a human requires basic learning on human actions.But to interact with a human a machine needs more.. ACT LIKE HUMAN Machine should know that human has its own plan, purpose and action. To understand and react accordingly, they should be able to know what is in human mind. This brought social intelligence robots, Ex: Robisuke It can observe emotions and identitiy.
- 14. OPEN QUESTION? COGNITION What is the moment that a robot advances itself and becomes self aware!! 1) https://lexfridman.com/ai/ 2) Christof Koch
- 15. ROBOT PHILOSOPHY VS HUMAN PHILOSOPHY It is funny that we placed unarmed robots where there is no human, and armed robots where there is human, then we claim that we do not want more people to die... Quoting with Mark Coeckelbergh - Ethics argue
- 17. HELLO DEEP LEARNING TRENDING Classification Error Over Years Data 40 30 20 10 02010 2011 2012 2013 2014 2015 150 100 50 0 2010 2011 2012 2013 2014 2014 2015 H um an 30 20 10 0 Number of Layers Deep Learning Trend vs Others
- 18. DEEP LEARNING OVERVIEW PRINCIPLE We receıve set of known ınputs and try to learn their features by the help of known outputs. To be able to capture the features we learn weights that map input to the known output.
- 19. HOW IT WORKS Learn weights for convolutional filters and fully connected layers. Backpropagation cross-entropy loss
- 20. DETAILS OF DEEP LEARNING We receıve set of known ınputs and try to learn their features by the help of known outputs. To be able to capture the features we learn weights that map input to the known output.
- 21. FEATURE LEARNING,DEEP LEARNING EDGE DETECTION
- 22. Viewpoint Variation Illumination Deformation DATASET PREPERATION Occlusion Background Clutter Intra-Class Variation Scale Invariance
- 23. We try to increase number of samples from same set of images via some operations. DATA AUGMENTATION
- 24. RECENT DEVELOPMENTS ON DEEP LEARNING OPTIMIZATION AND ROBUSTNESS - Adding some noise close to gibbon features to a panda image can help the detection of gibbon with good precision from a panda image even panda image does not look like gibbon. PROCESSING POWER - We want models to run on embedded devices, Snapchat ex: weight sampling to reduce memory usage by %50.Allowing 5x3 (15) weights be condensed to 7 weights. - Training networks with binary weights and activations at run time instead of 32 bit floating point. REINFORCEMENT LEARNING - Learning sequential tasks from human input, so interpreting both reward and punishment. INTERPRETABLITY OF AI - Narrow AI for single task on single domain. - Broad AI for mastering multiple task on multiple domain - General AI for thinking and learning on multiple domains. - Right answer with high likelihood. identifying learned patterns and detecting errors
- 25. REINFORCE UNFAIR BIAS Correctly identify a person’s gender from photographs 99 % of the time, but only for white men. For dark skinned-women the accuracy dropped to just 35 %. Reason is to be under-represented AI ACCOUNTABLE TO PEOPLE AI systems should be designed to provide appropriate opportunities for feedback, relevant explanations, and appeal. They should also be subject to human direction and control. PRIVACY DESIGN PRINCIPLES Giving opportunity for notice and consent, encouraging architectures with privacy safeguards Deep Learning Open Problems SAME-DIFFERENT Succesfull at finding same types of objects but in difficulty to capture how different thing are.
- 26. The Future with AI AI GOALS - Logical Reasoning - Knowledge Representation - Planning and Navigation - Natural Language Processing - Perception - Emergent Intelligence AI FIELDS - Machine Learning - Search and Optimization - Constraint Satisfaction - Logical Reasoning - Probabilistic Reasoning - Control Theory Deep Learning Machine Learning AI Artificial intelligence mean incorporating human intelligence to machines. Machine learning mean empowering computer systems with the ability to “learn”. Deep Learning technique for realizing machine learning, the next evolution of machine learning.
- 27. WHERE AI TECH LEADS US - BAD NEWS Perceive - Decide - Act But due to the Non Deterministic World, uncertainties are present. Decisions can not be made purely from the facts that are given!!
- 28. WHERE AI TECH LEADS US - GOOD NEWS Perceive - Decide - Act But due to the Non Deterministic World, uncertainties are present. Decisions can not be made purely from the facts that are given!!
- 29. INPUT QUALITY Computer Vision is heavily dependent on the quality of images, the factors like which camera was used, what time of the day was the image/video taken, and if the camera was stable. KNOWLEDGE OF THE MODEL If an object or image which wasn’t present in the training set, the model will only show incorrect results. A new type of gun may not known, if it's shape has never seen. PRIVACY DESIGN PRINCIPLES Giving opportunity for notice and consent, encouraging architectures with privacy safeguards SCENE UNDERSTANDING Object recognition gets most of the attention, because it is easy to define and understand.Given an image of the world, what is going on, what are the visual and structural elements and how do they relate each other Challenges in Computer Vision
- 30. PROJECT EXAMPLES - COMPUTER VISION RELATED Onesoil to determine what are the grown agricultural product types from satellite DeepMimic to mimic human behaviour. Waymo to predict car behaviour. Google Predicting cardiovascular risk from retina,also the gender with %97. https://map.onesoil.ai/2018#2/44.35/-43.66
- 31. YOU WILL BE THE PARADIGM SHIFTER! Technological innovations are very important for country development. Data is the new oil of the digital economy. We should have control over data - information - knowledge transformation. Presentations have different purposes. INFLUENCING CHANGE IN TURKEY
- 32. %27,6% HAD SOME INNOVATION 38,5% NO INNOVATION STARTUP INNOVATION IN 2014-2016
- 33. 82,2% OF STARTUPS INNOVATION IN TURKEY DO NOT MAKE INNOVATION Because there is no force for making innovation. %17.8 has blocking factors on innovation process.
- 34. TYPES OF INNOVATION The pie of product innovation in percent is %31 IMPROVEMENT Presentations have different purposes. INFLUENCING CHANGE IN TURKEY
- 35. Hug The World! PERSONAL USE COMMUNITY USE GLOBAL USE
- 36. Conclusion
- 37. BE INSPIRED ALBERT EINSTEIN Imagination is more important than knowledge.
- 38. Additional Reading MEDIUM Any Related Assays are fine COURSERA Mathematics for Machine Learning Deep Learning Specialisation Application Level Python Google Could Machine Learning MIT DEEP LEARNING Lex Fridman Podcasts of Lex Fridman Ava Soliemany UDEMY Computer Vision INDEPENDENT Adrian Rosebrock STANFORD CS231 Machine Learning 101
- 39. NICE TO READ The deepest problem with deep learning https://medium.com/@GaryMarcus/the-deepest-problem-with-deep-learning-91c5991f5695 Computer Vision: A study on different CNN architectures and their applications https://medium.com/alumnaiacademy/introduction-to-computer-vision-4fc2a2ba9dc Neural Networks seem to follow a puzzlingly simple strategy to classify images https://medium.com/bethgelab/neural-networks-seem-to-follow-a-puzzlingly-simple-strategy-to- classify-images-f4229317261f Benchmarking Hardware for CNN Inference in 2018 https://towardsdatascience.com/benchmarking-hardware-for-cnn-inference-in-2018-1d58268de12a Deep Reinforcement Learning Doesn't Work Yet https://www.alexirpan.com/2018/02/14/rl-hard.html A Gentle Introduction to Graph Neural Networks (Basics, DeepWalk, and GraphSage) https://towardsdatascience.com/a-gentle-introduction-to-graph-neural-network-basics-deepwalk-and- graphsage-db5d540d50b3
- 40. NICE TO READ Should we ban fully autonomous weapons? https://medienportal.univie.ac.at/uniview/wissenschaft-gesellschaft/detailansicht/artikel/should- we-ban-fully-autonomous-weapons/ Top 2018 Machine Learning Trends https://medium.com/the-official-integrate-ai-blog/top-2018-machine-learning-trends-and-our-2019- preview-9a6c82e6afba Current Trends in Deep Learning https://knowitlabs.no/current-trends-in-deep-learning-85e378dc813 Innovation https://www.dogrulukpayi.com/bulten/yenilik-arastirmasi Terminology of AI,ML,DL https://towardsdatascience.com/clearing-the-confusion-ai-vs-machine-learning-vs-deep-learning- differences-fce69b21d5eb Waymo https://medium.com/waymo/learning-to-drive-beyond-pure-imitation-465499f8bcb2
- 41. NAZLI TEMUR COMPUTER VISION R&D ENGINEER ADDRESS https://www.linkedin.com/in/nazlitemur/ EMAIL temur.nazlı@gmail.comGET IN TOUCH I'D LOVE TO HEAR YOUR THOUGHTS
- 42. WORLD CLOUD