Data Science, Machine Learning and Neural Networks
•
30 likes•9,018 views
Lecture briefly overviewing state of the art of Data Science, Machine Learning and Neural Networks. Covers main Artificial Intelligence technologies, Data Science algorithms, Neural network architectures and cloud computing facilities enabling the whole stack.
Recommended
More Related Content
What's hot
What's hot (20)
Viewers also liked
Similar to Data Science, Machine Learning and Neural Networks
Similar to Data Science, Machine Learning and Neural Networks (20)
Recently uploaded
Recently uploaded (20)
Data Science, Machine Learning and Neural Networks
- 1. DATA SCIENCE, MACHINE LEARNING, NEURAL NETWORKS Maxim Orlovsky, PhD, MD CloudBusinessCity, Mentor (cloudbusinesscity.com) GRPIIIQ, CEO (qoderoom.com, banqsystems.com) BICA Labs, Head (bicalabs.org)
- 4. BIG DATA • Volume • Velocity • Variety • Variability • Veracity • analysis • capture • data curation • search • sharing • storage • transfer • visualization • querying • updating • information privacy
- 5. DATA MINING computational process of discovering patterns in large data sets involving methods at the intersection of artificial intelligence, machine learning, statistics, and database systems Pre-Data Science Buzzword J
- 6. DATA SCIENCE • Part of Computer Science • Interdisciplinary field • Data -> Knowledge • Predictive analytics
- 7. CLOUD COMPUTING shared computer processing resources and data to computers and other devices on demand Cloud computing reduces cost of Data Science research and lowers entering threshold for startups
- 9. “ ” MACHINE LEARNING GIVES COMPUTERS THE ABILITY TO LEARN WITHOUT BEING EXPLICITLY PROGRAMMED Arthur Samuel, 1959
- 10. THE DIFFERENCE BETWEEN ML AND PROGRAMMING Programming Machine Learning Result of program Deterministic Non-deterministic Program Code Architecture Data storage External Embedded Changeability By human By machine
- 11. MACHINE LEARNING IS MORE THEN AI • Clustering • Regression • Dimensionality reduction • Decision trees • Genetic and evolutionary algorithms Machine learning is when computer updates it’s own algorithm depending on the data or its result
- 12. TYPES OF MACHINE LEARNING • Supervised: when you know what’s right and wrong (i.e. have labelled training sets) • Non-supervised: when you don’t know right answers/there is no labelled training sets • Reinforced: combination of supervised and unsupervised learning; similar to human’s learning
- 13. K-METHODS • k-Means Clustering: partition n observations into k clusters • k-Nearest Neighbors: assign class according to the environment
- 14. GENETIC AND EVOLUTIONARY ALGORITHMS Classical Algorithm Genetic Algorithm Generates a single point at each iteration. The sequence of points approaches an optimal solution. Generates a population of points at each iteration. The best point in the population approaches an optimal solution. Selects the next point in the sequence by a deterministic computation. Selects the next population by computation which uses random number generators.
- 16. COGNITIVE SCIENCE examines the nature, the tasks, and the functions of cognition • language • perception • memory • attention • reasoning • emotion
- 18. AI: TYPES • Specialized: performs only one task or subset of tasks, usually better then humans (compare to dogs, that smell better then we do) • Generic (human level and super-human)
- 20. WHEN GENERIC AI WILL APPEAR?
- 21. WHEN GENERIC AI WILL APPEAR?
- 23. BRAIN VS AI Brain • Massive parallelism: 100 000 000 000 “cores” • Extreme “bandwidth”: 700 000 000 000 000 connections between “cores” • ~10^18 “transistors” • Asynchronous • Adaptive hardware: neuroplasticity • “Analog”, but suitable for differential and integral computations Present day computer • Non-parallel architecture • Low bandwidth • ~10^9 transistors • Synchronous (clock rate) • Static hardware • Digital, but linear computations
- 24. SPECIALIZED AI CREATES MORE RISKS THEN GENERIC
- 25. NEURAL NETWORKS
- 30. UNDERSTANDING NEURAL NETWORKS Neural network as a graph of gateways * +w b
- 31. UNDERSTANDING NEURAL NETWORKS: HERE COMES TENSORS
- 33. WEIGHTS AND BIASES: HOW DOES THIS WORK
- 36. AI DISRUPTION 2016: KEY FACTORS 1. Machine Power and Cloud Computing 2. Big Data and its availability 3. Frameworks and ready-to-go cloud APIs
- 37. STARTUP TODO Design a product with USP and then 1. Look for the source of data 2. Find what you can personalize 3. Use cloud computing power 4. Use ready-to-go APIs when available 5. Don’t be afraid of creating and training own neural nets 6. Always use a proper ready-to-go framework for that purpose
- 40. ARCHITECTURES • Linear / recurrent • Non-deep / deep • Deterministic / probability • Supervised / unsupervised / reinforced
- 41. APPLICATIONS • Computer vision • NLP • Translation • Text-to-speech and vice verse • Generative methods • Personalization and adaptive methods • Complex solutions implementing different types of AI to obtain a cohesive result
- 46. THE NEXT REMBRANDT https://www.nextrembrandt.com “We now had a digital file true to Rembrandt’s style in content, shapes, and lighting. But paintings aren’t just 2D — they have a remarkable three- dimensionality that comes from brushstrokes and layers of paint. To recreate this texture, we had to study 3D scans of Rembrandt’s paintings and analyze the intricate layers on top of the canvas.”
- 47. CONTACTS Maxim Orlovsky About.me profile (all social networks): BICA Labs Scientific enquiries: Qoderoom Business enquiries:
- 48. CTO PART
- 49. MODERN NEURAL NETWORK ARCHITECTURES AND HOW THEY WORK
- 51. NN MECHANICS
- 56. HOW CONVOLUTION HAPPENS INSIDE NEURAL NETWORK
- 64. GENERATING HOUSE NUMBERS WITH RNN Credits: Fei-Fei Li & Andrej Karpathy & Justin Johnson, Stanford University
- 65. LONG SHORT-TERM MEMORY (LSTM): BEST RECURRENT ARCHITECTURE
- 66. LSTM: SIMPLIFICATION “Memory” New data Previous result Output “Forget and remember” Correct by “recalling”
- 67. DESIGNING NEURAL NET WITH YOUR DATA 1. Find a way to embed data 2. Understand what you’d like to receive from network 3. Design proper network architecture: 1. Use recurrent networks for time-based data 2. Use LSTM networks if time intervals between the data are large or non-even 3. Select number of layers according to data dimensionality 4. Has training set? Use supervised learning. Otherwise – reinforced. 4. Visualize and re-iterate hundreds of times 5. PROFIT!
- 68. FRAMEWORKS • TensorFlow • Teano • Torch • CNTK • Caffe
- 69. CAFFE • http://caffe.berkeleyvision.org • From Berkley University • Written in C++ • Create networks in Protocol Buffers: no need to write code • Has Python and MATLAB bindings • Good for feedforward networks • Good for finetuning existing networks • Not good for recurrent networks layer { name: "ip1" type: "InnerProduct" param { lr_mult: 1 } param { lr_mult: 2 } inner_product_param { num_output: 500 weight_filler { type: "xavier" } bias_filler { type: "constant" } } bottom: "pool2" top: "ip1" } Credits: Fei-Fei Li & Andrej Karpathy & Justin Johnson, Stanford University
- 70. TORCH • http://torch.ch • From New York University • Written in C and Lua • Used a lot a Facebook, DeepMind • Create networks in Lua • You usually write your own training code • Lots of modular pieces that are easy to combine • Less plug-and-play than Caffe • Easy to write your own layer types and run on GPU • Not good for recurrent networks Credits: Fei-Fei Li & Andrej Karpathy & Justin Johnson, Stanford University
- 71. LUA • High level scripting language, easy to interface with C • Similar to Javascript: • One data structure: table == JS object • Prototypical inheritance metatable == JS prototype • First-class functions • Downsides: • 1-indexed – bad for tensors =( • Variables global by default =( • Small standard library Credits: Fei-Fei Li & Andrej Karpathy & Justin Johnson, Stanford University
- 72. TEANO • http://deeplearning.net/software/theano • From University of Montreal • Python + numpy • Embracing computation graphs, symbolic computation • RNNs fit nicely in computational graph • Raw Theano is somewhat low-level • High level wrappers (Keras, Lasagne) ease the pain • Large models can have long compile times • Much “fatter” than Torch; more magic Credits: Fei-Fei Li & Andrej Karpathy & Justin Johnson, Stanford University
- 73. CNTK – THE MICROSOFT COGNITIVE TOOLKIT • https://www.cntk.ai • From Microsoft • Written in C++ • Programmed in Python and C++ • BrainScript: powerful abstraction • Good for both recurrent and convolution nets
- 74. TENSORFLOW • https://www.tensorflow.org • From Google • Python + numpy • Computational graph abstraction, like Theano; great for RNNs • Easy visualizations (TensorBoard) Multi-GPU and mzlti-node training • Data AND model parallelism; best of all frameworks • Slower than other frameworks right now • Much “fatter” than Torch; more magic Credits: Fei-Fei Li & Andrej Karpathy & Justin Johnson, Stanford University
- 75. OVERVIEW Credits: Fei-Fei Li & Andrej Karpathy & Justin Johnson, Stanford University
- 76. DATA SCIENCE IN NEURAL NETWORKS Dimensionality reduction
- 83. DATA SCIENCE IN NEURAL NETWORKS Inceptionism
- 84. INCEPTIONISM
- 89. MICROSOFT AZURE MACHINE LEARNING
- 95. EXPLORING AZURE COGNITIVE SERVICES Demo
- 96. USING AZURE ML TEXT ANALYTICS API Demo
- 97. USING AZURE DEEP LEARNING INSTANCES Demo
- 98. FURTHER READING Christopher Olah Ex Google Brain project member Andrej Karpathy DeepMind, Open AI, Stanford Stanford CS231n Neural networks & computer vision
- 99. OTHER MATERIALS (IN RUSSIAN) AI and our future Интервью «Platfor.ma» Dangers of AI Интервью «Радио Аристократы» AI & Blockchain Доклад на конференции Blockchaincof