Talk on Artificial Intelligence & Machine Learning in Defense Systems at ‘Tutorial cum workshop on AI&ML’ organized by IEEE Bombay Section in collaboration with the India Council during August 10-11, 2018.
3. Mature field
Text & Audio – Recurrent Neural Network (RNN)
Image & Video – Convolutional Neural Network (CNN)
Powerful frameworks such Tensorflow & Theano
Pre-trained models reaching Human Level performance and
approaching “Bayes Optimal Error” in certain applications
Powerful CPUs & GPUs and frameworks such as TensorRT for
optimization on target platforms
Mature to the extent that it’s as simple as building using LEGO
blocks!
4. Lego Blocks
Azure Machine Learning Studio – “Fully-
managed cloud services that enables you to
easily build, deploy and share predictive
analytics solutions”
Amazon SageMaker – “Enables data scientists
and developers to quickly and easily build,
train, and deploy machine learning models …
and one-click training, tuning, and inference.”
Google Cloud AI – “Cloud AI provides modern
machine learning services, with pre-trained
models and a service to generate your own
tailored models.
5. New Domains - Art
Mature
pre-trained models
components based solutions,
with UI based drag & drop
Art
neurons/activation function and
SGD
low level plumbing in “ML” assembly
programming in frameworks such as
Tensorflow or Theano
The pre-trained models, the LEGO
blocks, do not play well together
The father of modern neural networks, Geoffrey Hinton, who got neural
network in mainstream by introducing “back propagation” in 1986, has
recently been focusing on the fact that neurons are not the right
representation, and has been talking about a new high level concept
called “capsules”
7. Data
Data
What is deep
learning?
Input
Hand-designed
programs
Output
Summary
From: Deep Learning – Ian
Goodfellow et al.
Input
Hand-
designed
features
Mapping
from
features
Output
Input Features
Mapping
from
features
Output
Input
Simple
Features
Abstract
Features
Mapping
from
Features
Output
Rule based Systems
Classic Machine Learning
Machine Learning
Deep Learning
8. Challenges in Defence ML
Not Enough
Current advances are built on
of Data”
Unsupervised
Learning
Enormous Strides in ML have been
confined to Supervised Learning
Not enough
research in
Defence System
Current Research have been driven by
commercial applications
9. Generative
Adversarial Nets
(GAN)
Generative model G -
captures the data
distribution
Discriminative model D –
estimates the probability that
a sample came from the
training data rather than G
The training procedure for G
is to maximize the probability
of D making a mistake
Finally, G recovers the
training data distribution
Estimating generative models
via an adversarial process, in
which we simultaneously train
two models
Generative Adversarial Nets –
Ian Goodfellow et al, 2014.
10. Case Studies
Soft Error Predictions in Mission Critical
Systems
Tracking for Heterogeneous Sensors in a
Distributed Network
(Current Research at Tata Power SED)
12. Soft Error Prediction using ML
Aim
• To predict errors in real-time so as to take pro-active
“maintenance” actions
Scope
• Software runtime errors, and “soft errors” in associated
components
Target
• Software Intensive System, with limited deployments
15. Tracking
1.
Alignm
ent
2.
Gating
3.
Associa
tion
4.
Filterin
g
5. Life
Cycle
Mgmt
6.
Present
ation
Key steps are Filtering & Association
Filtering is based on Kalman Filtering –
best linear state estimator
Association – Optimisation Problem
State Estimation depends on target
motion modeling, and often non linear
Attribute fusion requires non-Bayesian
MLE approaches
No of sensor & their nature exploding,
so is data
Real time tracking a challenge
Example: CERN challenges on Kaggle
16. Approach
Deep Learning involves discriminative models, that map a high-
dimensional, rich sensory input to a class label
They are based on the backpropagation and dropout algorithms,
using piecewise linear units
Difficult in approximating many intractable probabilistic
computations that arise in Maximum Likelihood Estimation and non
linear Minimum Mean Square Estimation (MMSE), and due to
difficulty of leveraging the benefits of piecewise linear units in the
generative context.
Generative Network ► Simulator
Adversarial Network ► 4D CNN
18. Reading Recommendation
Learning representations by back-propagation errors – Rumelhart &
Hinton (1986 – Nature)
ADAM: A method for stochastic optimization – Diederik & Jimmy Lei Ba
(https: // arxiv.org/ pdf/1412.6980.pdf)
Generative Adversarial Nets – Goodfellow et al (https: // arxiv.org/
pdf/1406.2661.pdf)
Geoffrey Hinton talk "What is wrong with convolutional neural nets ?“
(https: //www. youtube.com/ watch?v=rTawFwUvnLE)
Matrix capsules with EM routing – Hinton (https: // openreview.net/
pdf?id=HJWLfGWRb)
Deep Learning – Ian Goodfellow, Yoshua Bengio & Aaron Courville
19. Sunil Chomal
Head Software
Tata Power Strategic Engineering Division
(proposed part of Tata Aerospace & Defence)
IEEE Computer Chapter IEEE Computational Intelligence
Society Agile Alliance Open Geospatial Consortium
(OGC)
www.linkedin.com/in/sunil-chomal
sunilchomal@gmail.com
Editor's Notes
Deep Learning -> Learning multiple levels of compositions. Solves problem of “Representational Learning” by introduction Rep. that are expressed in terms of other simpler Reps.
Pixels -> Edges -> Corners/Contours -> Object Parts -> Identification
DL is approach to AI that enables computers to improve with Experience & Data.
Algorithms being same, skills required to extract / model features reduces as amount of data increases
DL millions of neurons with 100’s of connections per neuron as of today
1. We live in a Digital Society and in the age of big data
2. Supervised Learning – Dataset containing many features associated with label / target
Unsupervised Learning – Experiences a dataset containing many features, but without labels. We need to learn the entire PD that generated the data set. PCA & K-means clustering.
3. Generative Adversarial Nets – Ian Goodfellow et al, 2014.
Estimating generative models via an adversarial process, in which we simultaneously train two models
Very similar to “Anomaly Detection” problems like credit card frauds.
System models your purchase habits
Thief’s purchases come from a different PD that yours.
What separates ML from Optimisation is that we want generalization error / test error to be low as well and not just training error.