Day #191 - Engineering and Reverse-Engineering Intelligence Using Probabilistic Programs, Program Induction, and Deep Learning

Key Lessons

• Intelligence is modelling the world
• Imagine unseen things
• Refine, Synthesis and model new things

• Decades away from bridging it

• Basic Idea motivated by Cognitive Science
• Probablistic Programs - Bayesian and other graphical models
• Programming Languages - Code Driven
• Neural Networks - CNN, RNN
• Representing Data as a Fine grained graphical model

Commonsense scene understanding

• Detect the dangers from the scene
• Understand the perspective
• Judge the stability of stacked objects
• Sense the activity that happens in scene

Origins of Common Sense

• How kids starts to learn and understand things
• Building the base intelligence
• Problem - Goals - Subgoals 
• Awareness of physical objects in the world

Reverse Engineer Common Sense

• Define Casual models
• Modelling the actions
• Update States with interactions
• Cost benefit trade-off actions
• Find Inferences
• Common Sense Scene Understanding

The below slide is a nice representation of different modes in learning

• Perception, Thinking, Learning, Development, Evolution
• The same applies to understand ourself and our situations. From my experiences how I evolved to be the next version. One is the psychological aspect another is these were the processes I went through.

• Meta-Inference Programs
• Speed - Reliability - Flexibility
• MonteCarlo Methods, CNN, RNN based methods

Physics Engine

• Accounting for Gravity, Senses
• Mental Simulation / Pattern Recognition Approach

• Develop perception based on the object
• Probablistic IPE

• Look at the data, test results
• Observe the distribution
• Understand the actions

Learning based / Pattern Recognition Approach

• Block arrangement simulated using CNN
• PhysNet
• Train on 200K images

Neural Scene De-Rendering

• Monte Carlo Inference
• Actions, Goals, Constraints need to be recognized

• Build program learning programs
• Model based learning
• One shot learning
• Probablistic high level program -> Token Generators 
• Capture Objects, Come up with Experimental, Learning model, Unsupervised Learning model

Part II

• Model represented by code, variables represent stochastic choices (random choices)
• Model using Graphical Models

Code for it

• Outliers
• Gaussian Distribution
• Evaluate the polynomial
• Degree Prior - Probabaility Assignment
• Stochastic Choice Degree (Random)

• Adding Hyper Parameters
• Query Operation - Probablistic program
• Set of Observations

• Need for custom inference strategies
• AI assisted Data Analysis
• Replicate basic intuitive capability
• Learning as a form of program induction
• Learning programs in domain specific language
• Probablistic program that generates probablistic code
• Represent Models as Programs

Scene Understanding

• Rapidly Exploring Random Tree + Path Planner
• Generative probabilistic graphics programming

Summary - Way forward is - Mix of NN, Probablistic Approach, Random Explorations of problem space to learn quickly

Generating 3D shapes using a Deep Network

• Technology to Scale up this
• Meta languages

Engineering and Reverse-Engineering Intelligence Using Probabilistic Programs, Program Induction, and Deep Learning from TechTalksTV on Vimeo.

Next List

How Computers See the Earth: A ML Approach to Understanding Satellite Imagery (Cloud Next '18)
Crop Yield Estimation from Satellite for Tropical Agriculture
dotAI 2017 - Mathias Ortner - Deep learning for Space imagery applications
Agriculture: The Next Machine-Learning Frontier | Data Dialogs 2016
Forecasting: Using Machine Learning Techniques in Energy, Environment and Agriculture
Machine Learning: Farm-to-Table Keynote II: Ranveer Chandra, Microsoft Research
NEW VIDEOS ] Automation, Robotics & Machine Learning in Agriculture
Zia Ahmed: Application of machine learning in agricultural research .
Machine Learning in the Agricultural Context Presentation
The Future of Farming with AI: Truly Organic at Scale
Generating Faces with Torch
Stochastic Video Generation with a Learned Prior

Happy Mastering DL!!!