Day #183 - Interesting Talks from London TensorFlow Meetup

Talk #1 - Convolutional LSTMs for video prediction : self-driving cars & medical image processing

Key Lessons

• Recursivity is a powerful procedure to solve complex function in simple steps
• Scale Invariance picture generated from recursivity

RNN

• Elmen Structure for RNN
• Context units (Information about past) keeps looping
• Update present state based on past
• Gradient Explosion

Convolutional LSTM

• LSM + FCN
• Video is special temporal process
• Segment video / Extract features using Convolutional LSTM
• RNN layer 
• Recurrent convolutional LSTM blocks forward outputs and inner states to future predictions
• Train them in Titan X GPU 8 hrs to train
• For one hour of video @ 20 frames per second

Keras and Tensorflow

• Code to check https://coxlab.github.io/prednet/

Weather forecasting

• LSTM Approach
• Server Maps
• Predict the next maps for precipitation / pressure 

Talk #2 - Predicting congestion on London's roads with TensorFlow - #LTM

Key Lessons
Data Collection

• 3 months data from 14K sensors
• 5 TB of Data
• 120 billion lines

Congestion prediction

• Avoid grid locks
• Predict ahead of time
• Preprocessing - beam (ETL)
• Tensorflow - NN

Beam Code

• Hello world of ETL tool
• Syntax is not Pythonic
• Data Aggregation  by 5 mins

Congestion

• Occupancy of Sensor (for five minutes)
• Streams occupied
• Metrics on how congested road is
• Derive Speed
• Max Speed Cars attained / Current Speed (Compute Relative Speed)

Machine Learning

• TF - High level APIs
• ML Kit
• Library for numerical computation
• Predict values 40 minutes ahead
• Using RNN
• Look for ripple patterns
• Congestion Alert and notify
• Traffic in this window, Hidden States -> Output Layer 

• RNN Vanishing gradient problem so LSTM came into picture

• Sample Code

• Model GPU 

Next Videos
VideoLSTM: Convolves, attends, and flows for action recognition
CVPR18: Tutorial: Part 4: Generative Adversarial Networks
CVPR18: Tutorial: Part 3: Human Activity Recognition
CVPR18: Tutorial: Part 2: Visual Recognition and Beyond
CVPR18: Tutorial: Part 1: Visual Recognition and Beyond
CVPR18: Tutorial: Part 2: Unsupervised Visual Learning
CVPR18: Tutorial: Part 2: Interpretable Machine Learning for Computer Vision
CVPR18: Session 2-1A: Object Recognition & Scene Understanding II
WACV18: Predicting Facial Attributes in Video using Temporal Coherence and Motion-Attention
WACV8: Driving Scene Perception Network: Real-time Joint Detection, Depth Estimation
WACV18: A Semi-Supervised Two-Stage Approach to Learning from Noisy Labels
An Introduction to LSTMs in Tensorflow
Tutorial: Probabilistic Programming

Happy Mastering DL!!!