2. Self-intro
• B.S. (2008-2012), Shanghai Jiao Tong University
• Electrical and Computer Engineering
• PHD. (2012-2017), University of Southern California
• Biomedical Engineering, USC-Viterbi PhD fellowship recipient
• Research focuses Computational Neuroscience, AI and Neural Networks
• Work(2017-present), Decision Engines Inc.
• AI & DS Team Lead, AI Architect
3. Agenda
• Intro to AI
• Convolutional Neural Networks (CNN)
• Recurrent Neural Networks (RNN)
• Deep Reinforcement Learning (DRL)
• Generative Adversarial Networks (GAN)
• Deep Learning Challenges
• Career Advice
4.
5. “systems that have been taught or learned how to carry out specific tasks”
6. We are still very far away from
“Artificial General Intelligence”
• General Intelligence is the type of adaptable intellect found in humans
• A flexible form of intelligence capable of learning how to carry out
vastly different tasks
• Anything from haircutting, driving, building spreadsheets
7. Sophia Scam??
In October 2017, Sophia became the
first robot to receive citizenship of any
country.
Sophia the Robot live on Jimmy Kimmel Show, 2018
11. What makes Deep Learning Successful now?
• Massively parallel computing with GPUs
• Appearance of large, high-quality labeled datasets
• Software platforms
• New architecture and techniques
12. Deep Learning Fundamental Architectures
• Deep Restricted Boltzmann Machine (Pioneer of deep learning)
• Convolutional Neural Networks (CNN)
• Deep Recurrent Neural Networks (LSTM, GRU)
• Deep Reinforcement Learning (DRL)
• Generative Adversarial Networks (GAN)
15. CNN Applications
• Self Driving Cars (Object Detection)
• Face Detection and Recognition (Face ID)
• Medical Image Diagnosis (Image classification and localization)
• Human Gesture and Pose recognition
• Optical Character Recognition
• Natural Language Processing
• Robotic and Manufacturing
34. Challenges of Deep Learning Models
• 1. Lack of transparency
• Lack of interpretations
• very hard to debug
• 2. Required a lot of training data, especially annotated data by human
• ImageNet has 14 million images, Coco data set has more than 100, 000 images
• Transfer learning could help
• 3. Not very robust, and easy to be attacked
• Change a single pixel of image could lead to a misclassification
• 4.Very shallow
• Most of the DL models are now only good at perception levels
• Cannot deduct and infer like human
• Cannot make use of prior knowledge
• Bad at hierarchical representations of knowledge
36. Data Science + AI Jobs
• “Old-fashioned” Data Science Positions (Process Structured Data)
• Data Analyst
• Business Analyst
• Data Engineer
• Data Scientist (Process structured data like excel, database)
• New DS & AI Related Jobs (Process Unstructured Data)
• Deep learning / Machine Learning Data Scientist
• Computer Vision Engineer
• Natural Language Processing (NLP) Engineer
• ML, DL, CV, NLP Research Scientist
37. How to become an AI Expert?
• Foundations:
• Math, Statistics, Linear Algebra, Signal Processing, Image Processing
• Knowledge of Machine Learning and Deep Learning
• Basic skills of Linux, Bash, Docker
• Basic Web Techs such as HTML, CSS, JS, API, etc
• Algorithm, OOP, system design
• Tools:
• Python/C, C++ (not recommend R)
• Keras + Tensorflow / Pytorch
• Advanced Projects and Skills:
• Domain Expert in Computer Vision / NLP / Speech Recognition / Speech Syntheses / OCR / Video Analysis /…
• Research skills, reading and writing papers, presentation, etc.
• Advanced Projects that showed your ability to complete an AI project from end to end
• Advanced Projects showed your capability in research, problem solving, or innovations
38. Recommended Learning Resources for Beginner
• Andrew Ng, Machine Learning and Deep Learning Courses on Coursera
• Feifei Li, Stanford CS231n, focus on computer vision
• Chris Manning, Stanford CS224n, focus on NLP
• Book: Hands-On Machine Learning with Scikit-Learn and TensorFlow
• Code Examples: https://github.com/keras-team/keras/tree/master/examples
39. Recommended Projects for Beginners
• 1. Image Classification
• Crawl data from google, build a classifier from scratch
• 2. Object Detection
• Collect and annotate the data by yourself
• Use Tensorflow object detection api to fine-tune the model
• 3. Document Classification
• Collect documents with different categories
• Build a text classifier using NLP models
• 4. Pick a problem that you are really intrigued and want to solve
• Data collection, annotation,
• model selection and training
• Build a demo, show your results