ME644: Machine Learning for Engineers Please go through the FCH carefully
(uploaded in pingala)
© Malay K. Das, 210 Southern Lab, mkdas@[Link]
Course Website: [Link]
Course Policy Syllabus: [Link]/ME/ME644
Machine Learning: Introduction, examples
k-nearest neighbor (kNN) ME644 is an introductory ML
course. If you have already done
(or now doing) similar courses
The course DOES NOT require any previous (such as CS771 or EE656), please
knowledge of machine learning drop this course immediately to
avoid deregistration
Lecture notes are copyrighted; do not share notes
without written permission from the instructor
© Malay K. Das, ME, IIT Kanpur, mkdas@[Link]
Course objective: To cover FUNDAMENTAL concepts of machine learning (ML) and
SOME applications of ML in mechanical engineering
- The focus is on the FUNDAMENTALS, thus, the course is MATHEMATICAL
- The course is NOT application-focused; applications come as EXAMPLES only
- The GOAL of the course:
- To impart the MINIMAL ML SKILLS a mechanical engineer needs
- To impart a SOLID FOUNDATION in ML for a mechanical engineer who wants to
delve deeper into ML theory/application
Course is outcome oriented; carefully read the course objective and outcome
2
Course outcome: At the end of the course, the students should be able to
1. explain scope, limitations, theory, and computational techniques of (a) regression,
(b) classification, (c) clustering, (d) dimensionality reduction, and (e) feedforward-
backpropagation shallow neural network, physically and mathematically,
2. write computer programs from scratch for the above, and
3. solve simple problems of engineering interest using the above programs.
- In this course, you will write programs sometimes from first principles and other times
using existing ML libraries
- Sound knowledge of a programming language is essential; knowledge of Python could
be useful, but not essential
© Malay K. Das, ME, IIT Kanpur, mkdas@[Link]
MATHEMATICAL MODELS IN ENGINEERING
- Physical phenomenon described in mathematical equations
- flow model, heat transfer model, failure model, combustion model etc.
- Used for prediction: weather prediction, cooling prediction, propulsion prediction
- Knowledge of theory (behind the physical phenomenon) is essential
How do we know the correct theory, how do we solve the model, what happened
when prediction contradicts experiments ……??
Despite many questions, predictive models remain a major tool for engineers
4
we always use models for
input data model output (prediction)
prediction
Example prediction
input
model
Example: wind turbine model
input numerical methods
prediction
Often, we do not have a well-accepted mechanistic
model, we only have some (mostly experimental) data
Machine Learning (ML) could be useful in such cases
© Malay K. Das, ME, IIT Kanpur, mkdas@[Link]
Machine Learning Definition (T. M. Mitchell, Machine Learning, 1997)
A computer program is said to learn from
prediction improves experience E with respect to some class of task T
with experience and performance measure P, if its performance
computer program at task T, as measured by P, improves with
experience E
Machine learning may be useful when We need data to
We don’t have a model to describe a process compensate for
these lacunae;
We have an incomplete model machine learning
We have a model but it is too complicated to solve comes as a useful
tool in such cases
We have a model but it is ill-posed
6
Prediction for tomorrow’s temperature: Traditional Modeling strategy
Initial, and boundary model
Inputs Outputs Tomorrow’s
conditions set of PDEs temperature
various properties
- Requires massive computation Machine
Learning could be
- Difficult to prescribe accurate initial/boundary conditions helpful here
- Requires highly skilled manpower
- Cannot tweak the model much, even if the predictions are not very good
© Malay K. Das, ME, IIT Kanpur, mkdas@[Link]
Prediction for tomorrow’s temperature: ML-based (data-driven) strategy
Ignores the PDEs called Training Data
Temperature (output) is called the label Collect lots of feature + label
(from past measurements)
Label is a function of inputs, called feature
model creation
(could be past temperature, humidity etc.) (called Training)
Selecting correct features requires domain feature label
??
knowledge
Once the Training is over, the Model is ready to accept the Test Data
Model output is the label of the test data,
test data has the feature but no label
which is the intended outcome of the
entire exercise
8
Prediction from traditional modeling Prediction from Machine Learning
Models are mechanistic (developed knowing Models are empirical (data-driven), may
the mechanism behind the phenomenon) have some physical basis (bias)
Problem (we are trying to solve) is the Problem and known problems/solutions
input; prediction is the output are the inputs; prediction is the output
Models are rigid; solution methods and Models are flexible; change with
certain input parameters may vary experience (bad predictions)
Mostly Deterministic Mostly Probabilistic
Purpose: prediction Purpose: prediction/generation
© Malay K. Das, ME, IIT Kanpur, mkdas@[Link]
Machine Learning is closely tied to mathematics
Linear algebra, vector calculus
Programming is essential
Optimization
In this course Probability
Mathematics of machine learning
Supervised Learning: finding label of test data, knowing features +
labels of training data
Unsupervised Learning: finding structure in unlabeled data
Artificial Neural Network (ANN): a tool used in ML
This course DOES NOT include deep learning, reinforcement learning (you will get the
ideas though)
10
