ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Machine Learning Techniques

ML Component Framework

Dr. Ashish Tendulkar

IIT Madras

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 1 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

1 ML Component Framework

2 Training Data

3 Model

4 Loss Function

5 Optimization

6 Evaluation

7 Summary

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 2 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

1 ML Component Framework

2 Training Data

3 Model

4 Loss Function

5 Optimization

6 Evaluation

7 Summary

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 3 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Machine Learning Process

Loss Label

Loss

TP vs FP rate at
a decision
threshold
w
Predicted Label

TP vs FP rate at
Metric
START HERE another decision
threshold
Negative
Starting Gradient Evaluation
Point

Final
Point

Loss

Model
w

Optimization Inference
Data
Training

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 4 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

ML Component Framework

• Training data
• Model
• Loss function
• Optimization procedure
• Evaluation criteria

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 5 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

ML Component Framework

• Training data
• Model
• Loss function
• Optimization procedure
• Evaluation criteria

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 5 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

ML Component Framework

• Training data
• Model
• Loss function
• Optimization procedure
• Evaluation criteria

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 5 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

ML Component Framework

• Training data
• Model
• Loss function
• Optimization procedure
• Evaluation criteria

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 5 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

ML Component Framework

• Training data
• Model
• Loss function
• Optimization procedure
• Evaluation criteria

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 5 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

ML Component Framework

• Training data
• Model
• Loss function
• Optimization procedure
• Evaluation criteria

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 5 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

1 ML Component Framework

2 Training Data

3 Model

4 Loss Function

5 Optimization

6 Evaluation

7 Summary

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 6 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• No data, no ML.
• In supervised learning, training data consists of input and
output pairs.
• Each input is represented by a bunch of numbers called
features or attributes.
• Apply certain transformations to convert input into a bunch of
numbers.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 7 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• No data, no ML.
• In supervised learning, training data consists of input and
output pairs.
• Each input is represented by a bunch of numbers called
features or attributes.
• Apply certain transformations to convert input into a bunch of
numbers.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 7 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• No data, no ML.
• In supervised learning, training data consists of input and
output pairs.
• Each input is represented by a bunch of numbers called
features or attributes.
• Apply certain transformations to convert input into a bunch of
numbers.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 7 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• No data, no ML.
• In supervised learning, training data consists of input and
output pairs.
• Each input is represented by a bunch of numbers called
features or attributes.
• Apply certain transformations to convert input into a bunch of
numbers.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 7 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• No data, no ML.
• In supervised learning, training data consists of input and
output pairs.
• Each input is represented by a bunch of numbers called
features or attributes.
• Apply certain transformations to convert input into a bunch of
numbers.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 7 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

For example,
• Input may be an image or a piece of text.
• Certain pre-processing steps or transformations have to be
applied to obtain feature representation for such input data.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 8 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

For example,
• Input may be an image or a piece of text.
• Certain pre-processing steps or transformations have to be
applied to obtain feature representation for such input data.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 8 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

For example,
• Input may be an image or a piece of text.
• Certain pre-processing steps or transformations have to be
applied to obtain feature representation for such input data.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 8 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• Features: are provided by the domain experts.

• For example, while predicting price of a house:
• The expert would tell us which features are most important in
determining the price.
• The expert would also provide us with a dataset of houses with
their features and prices (which is what we are interested in
predicting.)
• Output: a real number or a discrete value from a predefined
set.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 9 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• Features: are provided by the domain experts.

• For example, while predicting price of a house:
• The expert would tell us which features are most important in
determining the price.
• The expert would also provide us with a dataset of houses with
their features and prices (which is what we are interested in
predicting.)
• Output: a real number or a discrete value from a predefined
set.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 9 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• Features: are provided by the domain experts.

• For example, while predicting price of a house:
• The expert would tell us which features are most important in
determining the price.
• The expert would also provide us with a dataset of houses with
their features and prices (which is what we are interested in
predicting.)
• Output: a real number or a discrete value from a predefined
set.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 9 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• Features: are provided by the domain experts.

• For example, while predicting price of a house:
• The expert would tell us which features are most important in
determining the price.
• The expert would also provide us with a dataset of houses with
their features and prices (which is what we are interested in
predicting.)
• Output: a real number or a discrete value from a predefined
set.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 9 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• Features: are provided by the domain experts.

• For example, while predicting price of a house:
• The expert would tell us which features are most important in
determining the price.
• The expert would also provide us with a dataset of houses with
their features and prices (which is what we are interested in
predicting.)
• Output: a real number or a discrete value from a predefined
set.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 9 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• Features: are provided by the domain experts.

• For example, while predicting price of a house:
• The expert would tell us which features are most important in
determining the price.
• The expert would also provide us with a dataset of houses with
their features and prices (which is what we are interested in
predicting.)
• Output: a real number or a discrete value from a predefined
set.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 9 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

Example of a real valued label:

• The housing price prediction problem.
• Output label is price, a real number.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 10 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

Example of a real valued label:

• The housing price prediction problem.
• Output label is price, a real number.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 10 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

Example of a real valued label:

• The housing price prediction problem.
• Output label is price, a real number.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 10 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

Examples of a discrete valued label:

• The handwritten digit prediction problem. 10 different values
as class labels - digit 0 to 9.
• Loan sanctioning. Binary labels - yes or no corresponding to
application being sanctioned or rejected.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 11 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

Examples of a discrete valued label:

• The handwritten digit prediction problem. 10 different values
as class labels - digit 0 to 9.
• Loan sanctioning. Binary labels - yes or no corresponding to
application being sanctioned or rejected.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 11 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

Examples of a discrete valued label:

• The handwritten digit prediction problem. 10 different values
as class labels - digit 0 to 9.
• Loan sanctioning. Binary labels - yes or no corresponding to
application being sanctioned or rejected.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 11 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Training Data

• Once we have obtained training data, we get an idea about

input and output, which helps us in defining suitable ML
problems and choosing appropriate components like model,
loss and optimization procedure for training.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 12 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

1 ML Component Framework

2 Training Data

3 Model

4 Loss Function

5 Optimization

6 Evaluation

7 Summary

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 13 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

• Model provides a mathematical form of mapping between

input and output.
• The input is represented with a bunch of features and output
can be a real number or a discrete value from some finite set.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 14 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

• Model provides a mathematical form of mapping between

input and output.
• The input is represented with a bunch of features and output
can be a real number or a discrete value from some finite set.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 14 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

• Model provides a mathematical form of mapping between

input and output.
• The input is represented with a bunch of features and output
can be a real number or a discrete value from some finite set.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 14 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

Example of a linear model :

• Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• The key problem here is to estimate values of weights.
• All weights together form an entity called weight vector.
• We estimate the weight vector by training the model on the
training data.
• The ideal weights are the ones that when used in the model,
produce output that is close to the actual output for all
training data points.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 15 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

Example of a linear model :

• Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• The key problem here is to estimate values of weights.
• All weights together form an entity called weight vector.
• We estimate the weight vector by training the model on the
training data.
• The ideal weights are the ones that when used in the model,
produce output that is close to the actual output for all
training data points.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 15 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

Example of a linear model :

• Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• The key problem here is to estimate values of weights.
• All weights together form an entity called weight vector.
• We estimate the weight vector by training the model on the
training data.
• The ideal weights are the ones that when used in the model,
produce output that is close to the actual output for all
training data points.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 15 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

Example of a linear model :

• Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• The key problem here is to estimate values of weights.
• All weights together form an entity called weight vector.
• We estimate the weight vector by training the model on the
training data.
• The ideal weights are the ones that when used in the model,
produce output that is close to the actual output for all
training data points.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 15 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

Example of a linear model :

• Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• The key problem here is to estimate values of weights.
• All weights together form an entity called weight vector.
• We estimate the weight vector by training the model on the
training data.
• The ideal weights are the ones that when used in the model,
produce output that is close to the actual output for all
training data points.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 15 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

Example of a linear model :

• Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• The key problem here is to estimate values of weights.
• All weights together form an entity called weight vector.
• We estimate the weight vector by training the model on the
training data.
• The ideal weights are the ones that when used in the model,
produce output that is close to the actual output for all
training data points.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 15 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

• Depending on the nature of the output, we choose our model.

• When the output is a real number, we choose models of
regression - which are capable of producing a real valued
output.
• When the output is a discrete value, we choose models of
classification - which produce a discrete quantity as an output.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 16 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

• Depending on the nature of the output, we choose our model.

• When the output is a real number, we choose models of
regression - which are capable of producing a real valued
output.
• When the output is a discrete value, we choose models of
classification - which produce a discrete quantity as an output.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 16 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Model

• Depending on the nature of the output, we choose our model.

• When the output is a real number, we choose models of
regression - which are capable of producing a real valued
output.
• When the output is a discrete value, we choose models of
classification - which produce a discrete quantity as an output.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 16 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

1 ML Component Framework

2 Training Data

3 Model

4 Loss Function

5 Optimization

6 Evaluation

7 Summary

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 17 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• The key objective of training a model is to estimate the weight

vector and we need a principled way of doing that.
• We need a suitable method for measuring the difference
between predicted and actual output.
• Loss function provides that measure.
• Loss function is a function of weight vector - as we change the
weight vector, we obtain a new model, which will have
different a loss.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 18 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• The key objective of training a model is to estimate the weight

vector and we need a principled way of doing that.
• We need a suitable method for measuring the difference
between predicted and actual output.
• Loss function provides that measure.
• Loss function is a function of weight vector - as we change the
weight vector, we obtain a new model, which will have
different a loss.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 18 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• The key objective of training a model is to estimate the weight

vector and we need a principled way of doing that.
• We need a suitable method for measuring the difference
between predicted and actual output.
• Loss function provides that measure.
• Loss function is a function of weight vector - as we change the
weight vector, we obtain a new model, which will have
different a loss.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 18 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• The key objective of training a model is to estimate the weight

vector and we need a principled way of doing that.
• We need a suitable method for measuring the difference
between predicted and actual output.
• Loss function provides that measure.
• Loss function is a function of weight vector - as we change the
weight vector, we obtain a new model, which will have
different a loss.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 18 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Denote loss with letter J.

• J : W −→ R
• J(W ) = Difference between actual and predicted output for
all training samples.
• In the following example model:
Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• Everything except weight vector is fixed on the right side of
the equation- features are specified as part of the training
data, which is fixed and weights are variables.
• The weights can be changed to obtain a different result or
output.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 19 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Denote loss with letter J.

• J : W −→ R
• J(W ) = Difference between actual and predicted output for
all training samples.
• In the following example model:
Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• Everything except weight vector is fixed on the right side of
the equation- features are specified as part of the training
data, which is fixed and weights are variables.
• The weights can be changed to obtain a different result or
output.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 19 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Denote loss with letter J.

• J : W −→ R
• J(W ) = Difference between actual and predicted output for
all training samples.
• In the following example model:
Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• Everything except weight vector is fixed on the right side of
the equation- features are specified as part of the training
data, which is fixed and weights are variables.
• The weights can be changed to obtain a different result or
output.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 19 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Denote loss with letter J.

• J : W −→ R
• J(W ) = Difference between actual and predicted output for
all training samples.
• In the following example model:
Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• Everything except weight vector is fixed on the right side of
the equation- features are specified as part of the training
data, which is fixed and weights are variables.
• The weights can be changed to obtain a different result or
output.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 19 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Denote loss with letter J.

• J : W −→ R
• J(W ) = Difference between actual and predicted output for
all training samples.
• In the following example model:
Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• Everything except weight vector is fixed on the right side of
the equation- features are specified as part of the training
data, which is fixed and weights are variables.
• The weights can be changed to obtain a different result or
output.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 19 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Denote loss with letter J.

• J : W −→ R
• J(W ) = Difference between actual and predicted output for
all training samples.
• In the following example model:
Output = weight0 + weight1 · feature1 + weight2 · feature2 +
· · · + weightm · featurem
• Everything except weight vector is fixed on the right side of
the equation- features are specified as part of the training
data, which is fixed and weights are variables.
• The weights can be changed to obtain a different result or
output.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 19 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• An example of loss function is squared loss function: It is

calculated as a sum of square of differences between the actual
and predicted values.
n
X
J(W ) = [predicted (i) − actual (i) ]2
i=1

• Above equation can be simplified to:

J(W ) = ni=1 [{weight0 + weight1 · feature1 (i) + weight2 ·
P
feature2 (i) + · · · + weightm · featurem (i) } − actual (i) ]2

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 20 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• An example of loss function is squared loss function: It is

calculated as a sum of square of differences between the actual
and predicted values.
n
X
J(W ) = [predicted (i) − actual (i) ]2
i=1

• Above equation can be simplified to:

J(W ) = ni=1 [{weight0 + weight1 · feature1 (i) + weight2 ·
P
feature2 (i) + · · · + weightm · featurem (i) } − actual (i) ]2

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 20 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• An example of loss function is squared loss function: It is

calculated as a sum of square of differences between the actual
and predicted values.
n
X
J(W ) = [predicted (i) − actual (i) ]2
i=1

• Above equation can be simplified to:

J(W ) = ni=1 [{weight0 + weight1 · feature1 (i) + weight2 ·
P
feature2 (i) + · · · + weightm · featurem (i) } − actual (i) ]2

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 20 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Our job is to find the weight vector that results in the lowest
loss as per the defined loss function.
• How to estimate such a weight vector?
• Brute force search for the optimal weight vector, but that is
not at all efficient.
• What is the best way to estimate it?
• Well that’s what is the purpose of our next component:
optimization procedure.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 21 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Our job is to find the weight vector that results in the lowest
loss as per the defined loss function.
• How to estimate such a weight vector?
• Brute force search for the optimal weight vector, but that is
not at all efficient.
• What is the best way to estimate it?
• Well that’s what is the purpose of our next component:
optimization procedure.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 21 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Our job is to find the weight vector that results in the lowest
loss as per the defined loss function.
• How to estimate such a weight vector?
• Brute force search for the optimal weight vector, but that is
not at all efficient.
• What is the best way to estimate it?
• Well that’s what is the purpose of our next component:
optimization procedure.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 21 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Our job is to find the weight vector that results in the lowest
loss as per the defined loss function.
• How to estimate such a weight vector?
• Brute force search for the optimal weight vector, but that is
not at all efficient.
• What is the best way to estimate it?
• Well that’s what is the purpose of our next component:
optimization procedure.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 21 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Loss Function

• Our job is to find the weight vector that results in the lowest
loss as per the defined loss function.
• How to estimate such a weight vector?
• Brute force search for the optimal weight vector, but that is
not at all efficient.
• What is the best way to estimate it?
• Well that’s what is the purpose of our next component:
optimization procedure.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 21 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

1 ML Component Framework

2 Training Data

3 Model

4 Loss Function

5 Optimization

6 Evaluation

7 Summary

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 22 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Finding Weight Vector

• Obtain a weight vector that minimizes the loss function.

• Formally,
W = argminW J(W )

• Applications of derivatives are typically in 12th calculus in the

context of such problems.
• Take a derivative of loss function w.r.t. the weight vector and
set it to 0.
d
J(W ) = 0
dW

• Solve this equation directly or with analytical methods to

obtain the optimal weight vector.
• The optimization procedure is a cornerstone of model training.
Dr. Ashish Tendulkar Online Degree Program, IIT Madras
Machine Learning Techniques 23 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Finding Weight Vector

• Obtain a weight vector that minimizes the loss function.

• Formally,
W = argminW J(W )

• Applications of derivatives are typically in 12th calculus in the

context of such problems.
• Take a derivative of loss function w.r.t. the weight vector and
set it to 0.
d
J(W ) = 0
dW

• Solve this equation directly or with analytical methods to

obtain the optimal weight vector.
• The optimization procedure is a cornerstone of model training.
Dr. Ashish Tendulkar Online Degree Program, IIT Madras
Machine Learning Techniques 23 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Finding Weight Vector

• Obtain a weight vector that minimizes the loss function.

• Formally,
W = argminW J(W )

• Applications of derivatives are typically in 12th calculus in the

context of such problems.
• Take a derivative of loss function w.r.t. the weight vector and
set it to 0.
d
J(W ) = 0
dW

• Solve this equation directly or with analytical methods to

obtain the optimal weight vector.
• The optimization procedure is a cornerstone of model training.
Dr. Ashish Tendulkar Online Degree Program, IIT Madras
Machine Learning Techniques 23 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Finding Weight Vector

• Obtain a weight vector that minimizes the loss function.

• Formally,
W = argminW J(W )

• Applications of derivatives are typically in 12th calculus in the

context of such problems.
• Take a derivative of loss function w.r.t. the weight vector and
set it to 0.
d
J(W ) = 0
dW

• Solve this equation directly or with analytical methods to

obtain the optimal weight vector.
• The optimization procedure is a cornerstone of model training.
Dr. Ashish Tendulkar Online Degree Program, IIT Madras
Machine Learning Techniques 23 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Finding Weight Vector

• Obtain a weight vector that minimizes the loss function.

• Formally,
W = argminW J(W )

• Applications of derivatives are typically in 12th calculus in the

context of such problems.
• Take a derivative of loss function w.r.t. the weight vector and
set it to 0.
d
J(W ) = 0
dW

• Solve this equation directly or with analytical methods to

obtain the optimal weight vector.
• The optimization procedure is a cornerstone of model training.
Dr. Ashish Tendulkar Online Degree Program, IIT Madras
Machine Learning Techniques 23 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Finding Weight Vector

• Obtain a weight vector that minimizes the loss function.

• Formally,
W = argminW J(W )

• Applications of derivatives are typically in 12th calculus in the

context of such problems.
• Take a derivative of loss function w.r.t. the weight vector and
set it to 0.
d
J(W ) = 0
dW

• Solve this equation directly or with analytical methods to

obtain the optimal weight vector.
• The optimization procedure is a cornerstone of model training.
Dr. Ashish Tendulkar Online Degree Program, IIT Madras
Machine Learning Techniques 23 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Finding Weight Vector

• Obtain a weight vector that minimizes the loss function.

• Formally,
W = argminW J(W )

• Applications of derivatives are typically in 12th calculus in the

context of such problems.
• Take a derivative of loss function w.r.t. the weight vector and
set it to 0.
d
J(W ) = 0
dW

• Solve this equation directly or with analytical methods to

obtain the optimal weight vector.
• The optimization procedure is a cornerstone of model training.
Dr. Ashish Tendulkar Online Degree Program, IIT Madras
Machine Learning Techniques 23 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

1 ML Component Framework

2 Training Data

3 Model

4 Loss Function

5 Optimization

6 Evaluation

7 Summary

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 24 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Evaluation

After optimization:
• How to evaluate the model on unseen data?
• What metric to use?
• The evaluation metric changes as per the problem - different
metric for regression and classification problems.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 25 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Evaluation

After optimization:
• How to evaluate the model on unseen data?
• What metric to use?
• The evaluation metric changes as per the problem - different
metric for regression and classification problems.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 25 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Evaluation

After optimization:
• How to evaluate the model on unseen data?
• What metric to use?
• The evaluation metric changes as per the problem - different
metric for regression and classification problems.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 25 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Evaluation

After optimization:
• How to evaluate the model on unseen data?
• What metric to use?
• The evaluation metric changes as per the problem - different
metric for regression and classification problems.

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 25 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Evaluation

• What if the model does not perform as per the expectation as

found via evaluation metric?
• Reiterate the ML pipeline.
• Try to add superior
• features
• models
• loss function
• optimization procedure

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 26 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Evaluation

• What if the model does not perform as per the expectation as

found via evaluation metric?
• Reiterate the ML pipeline.
• Try to add superior
• features
• models
• loss function
• optimization procedure

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 26 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Evaluation

• What if the model does not perform as per the expectation as

found via evaluation metric?
• Reiterate the ML pipeline.
• Try to add superior
• features
• models
• loss function
• optimization procedure

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 26 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

1 ML Component Framework

2 Training Data

3 Model

4 Loss Function

5 Optimization

6 Evaluation

7 Summary

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 27 / 28
ML Component Framework Training Data Model Loss Function Optimization Evaluation Summary

Summary

Machine Learning Process

Loss Label

Loss

TP vs FP rate at
a decision
threshold
w
Predicted Label

TP vs FP rate at
Metric
START HERE another decision
threshold
Negative
Starting Gradient Evaluation
Point

Final
Point

Loss

Model
w

Optimization Inference
Data
Training

Dr. Ashish Tendulkar Online Degree Program, IIT Madras

Machine Learning Techniques 28 / 28