1
COURSE OBJECTIVES
• To understand the theoretical foundations,
algorithms and methodologies of Neural
Network
• To design and develop an application using
specific deep learning models
• To provide the practical knowledge in handling
and analysing real world applications.
COURSE OUTCOMES
• Recognize the characteristics of deep learning models
that are useful to solve real-world problems.
• Understand different methodologies to create application
using deep nets.
• Identify and apply appropriate deep learning algorithms
for analyzing the data for variety of problems.
• Implement different deep learning algorithms
• Design the test procedures to assess the efficacy of the
developed model.
• Combine several models in to gain better results
CONTENT
Topics to be Covered under these units
1
I
Introduction to Machine Learning and its categories
2
Learning Algorithms – Overview and classification
3
Maximum Likelihood Estimation (MLE) – Concept and derivation
4
Application of MLE in classification problems
5
Basics of building Machine Learning algorithms
6
Introduction to Neural Networks
7
Structure and function of Multilayer Perceptron (MLP)
8
Forward Propagation in MLP
9
Backpropagation Algorithm – Concepts
10
Backpropagation Algorithm – Mathematical formulation
11
Stochastic Gradient Descent – Working and updates
12
Curse of Dimensionality – Causes and implications
REFERENCE
• Kevin P. Murphy "Machine Learning: A Probabilistic
Perspective", The MIT Press, 2012.
• EthemAlpaydin,"Introduction to Machine Learning”, MIT
Press, Prentice Hall of India, Third Edition 2014.
• Giancarlo Zaccone, Md. RezaulKarim, Ahmed Menshawy
"Deep Learning with TensorFlow: Explore neural networks
with Python", Packt Publisher, 2017.
• Antonio Gulli, Sujit Pal "Deep Learning with Keras", Packt
Publishers, 2017.
• Francois Chollet "Deep Learning with Python", Manning
Publications, 2017.
Advance Neural Network &
Deep Learning(BCAAIML505)
Advance Neural Network & Deep
Learning
Unit-1
What is Machine Learning (ML)?
• Machine Learning (ML) is a subfield of Artificial
Intelligence (AI) that allows computer systems to
learn from data without being explicitly programmed.
Rather than using predefined instructions for every
task, ML systems identify patterns in data and make
decisions based on experience. The term was coined
by Arthur Samuel in 1959. The success of ML today is
driven by big data, increased computational power,
and advanced algorithms.
Importance of Machine Learning
• ML is crucial in various domains:
• - Automates complex tasks like image recognition and
language translation.
• - Enhances data-driven decision-making.
• - Enables personalization in marketing, entertainment, and
healthcare.
• - Used in predictive analytics to forecast trends and
behaviors.
• - Helps in scientific research and solving ill-posed problems.
Learning Process in ML
• The ML process includes:
• - Training Data: Labeled dataset to teach the model.
• - Features: Inputs or variables that describe the data.
• - Labels: Outputs or target values.
• - Model: Learns relationships between features and labels.
• - Training: Adjusting model parameters.
• - Evaluation: Using metrics (accuracy, precision, etc.) to
assess performance.
• - Prediction: Applying the model to new data.
Categories of ML: Supervised Learning
• Supervised Learning uses labeled data to learn
mapping from inputs to outputs.
• - Classification: Predicting categories (e.g., spam
detection).
• - Regression: Predicting continuous values (e.g.,
house prices).
• Algorithms include Logistic Regression, SVM,
Decision Trees, KNN, and Neural Networks.
Unsupervised Learning
• Unsupervised Learning deals with unlabeled data.
• - Clustering: Grouping similar data points (e.g., customer
segmentation).
• - Dimensionality Reduction: Reducing input features
(e.g., PCA).
• - Association Rules: Finding relationships in data (e.g.,
market basket analysis).
• Common algorithms include K-Means, Hierarchical
Clustering, DBSCAN, PCA, and Apriori.
Reinforcement Learning
• Reinforcement Learning (RL) trains an agent to make
sequences of decisions by rewarding desired behavior.
• - Components: Agent, Environment, States, Actions,
Rewards, and Policies.
• - Goal: Learn a policy that maximizes total rewards.
• Applications include game playing, robotics, self-driving
cars.
• Algorithms: Q-Learning, SARSA, DQN, Policy Gradients.
Semi- & Self-Supervised Learning
• Semi-Supervised Learning uses both labeled and
unlabeled data. It's useful when labels are scarce.
• Self-Supervised Learning generates its own labels
from data. Common in NLP and CV.
• - Examples: BERT (NLP), SimCLR (Vision).
• - Uses pretext tasks like predicting missing data.
• Benefits: Effective use of large unlabeled datasets.
Key ML Considerations
• Key factors affecting ML success:
• - Data Quality: Accurate, relevant, and sufficient data.
• - Feature Engineering: Transforming raw data into useful
features.
• - Overfitting/Underfitting: Avoid too complex or too simple
models.
• - Bias-Variance Tradeoff: Balance between simplicity and
flexibility.
• - Metrics: Choose appropriate metrics (F1, AUC, RMSE, etc.).
ML Workflow
• Steps in ML project:
• 1. Define problem clearly.
• 2. Collect and preprocess data.
• 3. Split data (train/test/validation).
• 4. Choose suitable model.
• 5. Train the model.
• 6. Evaluate using test data.
• 7. Tune hyperparameters.
• 8. Deploy and monitor the model.
Learning Algorithms Overview
• Learning algorithms are the backbone of ML.
• - Linear Regression: Models linear relationships.
• - Logistic Regression: Classifies using probability.
• - Decision Trees: Tree structure for
classification/regression.
• - SVM: Separates data with optimal hyperplanes.
• - KNN: Predicts based on neighbors.
• - Naive Bayes: Uses Bayes theorem.
• - Neural Networks: Multiple layers of connected neurons.
Introduction to Machine Learning –
Definitions and Goals
• ML enables computers to learn from data.
• Definitions by Arthur Samuel (1959) and Tom
Mitchell (1997).
• ML focuses on prediction, pattern discovery,
automation, and adaptation.
Well-defined Learning Problems
• Components: Task (T), Experience (E),
Performance Measure (P).
• Clear definition ensures reproducibility,
measurement, and success.
• Examples: Spam filtering, house price
prediction, chess playing.
Designing a Learning System – Components
and Process
• Steps: Data Collection → Preprocessing →
Model Selection → Training → Evaluation →
Deployment.
• Model Evaluation uses test data and
performance metrics.
• Iterative process with monitoring and
retraining.
Issues in Machine Learning – Noise, Overfitting, Bias-Variance
• Noise: Irrelevant/misleading data, affects
generalization.
• Overfitting: High training accuracy but poor
test performance.
• Bias-Variance: Trade-off to minimize total
error.
MCQs 1 & 2
• 1. Who defined ML as 'ability to learn without being explicitly
programmed'?
A. Tom Mitchell
B. Andrew Ng
C. Arthur Samuel
D. Geoffrey Hinton [Ans: C]
• 2. Which component in ML represents the task being solved?
A. Experience
B. Performance
C. Task
D. Evaluation [Ans: C]
MCQs 3 & 4
• 3. Which of these is NOT a goal of ML?
A. Prediction
B. Automation
C. Code writing
D. Pattern Recognition [Ans: C]
• 4. What does the P in T-E-P stand for?
A. Performance
B. Processing
C. Prediction
D. Probability [Ans: A]
MCQs 5 & 6
• 5. Which is an example of a classification task?
A. Predicting temperature
B. Recognizing spam emails
C. Forecasting stock price
D. Predicting height [Ans: B]
• 6. What is bias in ML?
A. Model’s sensitivity to training data
B. Simplification error
C. Random error
D. Noise [Ans: B]
MCQs 7 & 8
• 7. Which technique reduces model complexity?
A. Gradient Descent
B. Regularization
C. Dropout
D. Backpropagation [Ans: B]
• 8. What is the goal of Find-S algorithm?
A. Find all consistent hypotheses
B. Eliminate inconsistent examples
C. Find most general hypothesis
D. Find most specific hypothesis [Ans: D]
MCQs 9 & 10
• 9. Which algorithm ignores negative examples?
A. Candidate-Elimination
B. List-Then-Eliminate
C. Find-S
D. Decision Trees [Ans: C]
• 10. What does '?' symbol mean in hypotheses?
A. No match
B. Match anything
C. Error
D. Specific value [Ans: B]
MCQs 11 & 12
• 11. Which boundary is most general in Candidate-Elimination?
A. S
B. T
C. G
D. H [Ans: C]
• 12. What is inductive bias?
A. Noise in data
B. Assumptions guiding learning
C. Incorrect output
D. Inference process [Ans: B]
MCQs 13 & 14
• 13. Which algorithm explicitly maintains hypothesis space?
A. Find-S
B. Candidate-Elimination
C. List-Then-Eliminate
D. Linear Regression [Ans: C]
• 14. What happens to G set when a positive example fails?
A. Shrink
B. Expand
C. Stay same
D. Empty [Ans: A]
MCQs 15 & 16
• 15. Which strategy involves refining S and G sets?
A. List-Then-Eliminate
B. Find-S
C. Candidate-Elimination
D. Clustering [Ans: C]
• 16. What is overfitting?
A. High test accuracy
B. Good generalization
C. Fits noise too well
D. Undertrained model [Ans: C]
MCQs 17 & 18
• 17. Which error can never be reduced?
A. Bias
B. Variance
C. Irreducible Error
D. Prediction Error [Ans: C]
• 18. Which is a core concept of concept learning?
A. Feature extraction
B. Hypothesis search
C. Error analysis
D. Training accuracy [Ans: B]
MCQs 19 & 20
• 19. Which algorithm is computationally intensive?
A. Find-S
B. Candidate-Elimination
C. List-Then-Eliminate
D. SVM [Ans: C]
• 20. Which task involves boolean-valued functions?
A. Regression
B. Clustering
C. Concept Learning
D. Prediction [Ans: C]
THANK YOU

BCAAIML505 - PPT.pptx FOR DATA 7 ANALYICS SCIENCE

  • 1.
    1 COURSE OBJECTIVES • Tounderstand the theoretical foundations, algorithms and methodologies of Neural Network • To design and develop an application using specific deep learning models • To provide the practical knowledge in handling and analysing real world applications.
  • 2.
    COURSE OUTCOMES • Recognizethe characteristics of deep learning models that are useful to solve real-world problems. • Understand different methodologies to create application using deep nets. • Identify and apply appropriate deep learning algorithms for analyzing the data for variety of problems. • Implement different deep learning algorithms • Design the test procedures to assess the efficacy of the developed model. • Combine several models in to gain better results
  • 3.
    CONTENT Topics to beCovered under these units 1 I Introduction to Machine Learning and its categories 2 Learning Algorithms – Overview and classification 3 Maximum Likelihood Estimation (MLE) – Concept and derivation 4 Application of MLE in classification problems 5 Basics of building Machine Learning algorithms 6 Introduction to Neural Networks 7 Structure and function of Multilayer Perceptron (MLP) 8 Forward Propagation in MLP 9 Backpropagation Algorithm – Concepts 10 Backpropagation Algorithm – Mathematical formulation 11 Stochastic Gradient Descent – Working and updates 12 Curse of Dimensionality – Causes and implications
  • 4.
    REFERENCE • Kevin P.Murphy "Machine Learning: A Probabilistic Perspective", The MIT Press, 2012. • EthemAlpaydin,"Introduction to Machine Learning”, MIT Press, Prentice Hall of India, Third Edition 2014. • Giancarlo Zaccone, Md. RezaulKarim, Ahmed Menshawy "Deep Learning with TensorFlow: Explore neural networks with Python", Packt Publisher, 2017. • Antonio Gulli, Sujit Pal "Deep Learning with Keras", Packt Publishers, 2017. • Francois Chollet "Deep Learning with Python", Manning Publications, 2017.
  • 5.
    Advance Neural Network& Deep Learning(BCAAIML505)
  • 6.
    Advance Neural Network& Deep Learning Unit-1
  • 7.
    What is MachineLearning (ML)? • Machine Learning (ML) is a subfield of Artificial Intelligence (AI) that allows computer systems to learn from data without being explicitly programmed. Rather than using predefined instructions for every task, ML systems identify patterns in data and make decisions based on experience. The term was coined by Arthur Samuel in 1959. The success of ML today is driven by big data, increased computational power, and advanced algorithms.
  • 8.
    Importance of MachineLearning • ML is crucial in various domains: • - Automates complex tasks like image recognition and language translation. • - Enhances data-driven decision-making. • - Enables personalization in marketing, entertainment, and healthcare. • - Used in predictive analytics to forecast trends and behaviors. • - Helps in scientific research and solving ill-posed problems.
  • 9.
    Learning Process inML • The ML process includes: • - Training Data: Labeled dataset to teach the model. • - Features: Inputs or variables that describe the data. • - Labels: Outputs or target values. • - Model: Learns relationships between features and labels. • - Training: Adjusting model parameters. • - Evaluation: Using metrics (accuracy, precision, etc.) to assess performance. • - Prediction: Applying the model to new data.
  • 10.
    Categories of ML:Supervised Learning • Supervised Learning uses labeled data to learn mapping from inputs to outputs. • - Classification: Predicting categories (e.g., spam detection). • - Regression: Predicting continuous values (e.g., house prices). • Algorithms include Logistic Regression, SVM, Decision Trees, KNN, and Neural Networks.
  • 11.
    Unsupervised Learning • UnsupervisedLearning deals with unlabeled data. • - Clustering: Grouping similar data points (e.g., customer segmentation). • - Dimensionality Reduction: Reducing input features (e.g., PCA). • - Association Rules: Finding relationships in data (e.g., market basket analysis). • Common algorithms include K-Means, Hierarchical Clustering, DBSCAN, PCA, and Apriori.
  • 12.
    Reinforcement Learning • ReinforcementLearning (RL) trains an agent to make sequences of decisions by rewarding desired behavior. • - Components: Agent, Environment, States, Actions, Rewards, and Policies. • - Goal: Learn a policy that maximizes total rewards. • Applications include game playing, robotics, self-driving cars. • Algorithms: Q-Learning, SARSA, DQN, Policy Gradients.
  • 13.
    Semi- & Self-SupervisedLearning • Semi-Supervised Learning uses both labeled and unlabeled data. It's useful when labels are scarce. • Self-Supervised Learning generates its own labels from data. Common in NLP and CV. • - Examples: BERT (NLP), SimCLR (Vision). • - Uses pretext tasks like predicting missing data. • Benefits: Effective use of large unlabeled datasets.
  • 14.
    Key ML Considerations •Key factors affecting ML success: • - Data Quality: Accurate, relevant, and sufficient data. • - Feature Engineering: Transforming raw data into useful features. • - Overfitting/Underfitting: Avoid too complex or too simple models. • - Bias-Variance Tradeoff: Balance between simplicity and flexibility. • - Metrics: Choose appropriate metrics (F1, AUC, RMSE, etc.).
  • 15.
    ML Workflow • Stepsin ML project: • 1. Define problem clearly. • 2. Collect and preprocess data. • 3. Split data (train/test/validation). • 4. Choose suitable model. • 5. Train the model. • 6. Evaluate using test data. • 7. Tune hyperparameters. • 8. Deploy and monitor the model.
  • 16.
    Learning Algorithms Overview •Learning algorithms are the backbone of ML. • - Linear Regression: Models linear relationships. • - Logistic Regression: Classifies using probability. • - Decision Trees: Tree structure for classification/regression. • - SVM: Separates data with optimal hyperplanes. • - KNN: Predicts based on neighbors. • - Naive Bayes: Uses Bayes theorem. • - Neural Networks: Multiple layers of connected neurons.
  • 17.
    Introduction to MachineLearning – Definitions and Goals • ML enables computers to learn from data. • Definitions by Arthur Samuel (1959) and Tom Mitchell (1997). • ML focuses on prediction, pattern discovery, automation, and adaptation.
  • 18.
    Well-defined Learning Problems •Components: Task (T), Experience (E), Performance Measure (P). • Clear definition ensures reproducibility, measurement, and success. • Examples: Spam filtering, house price prediction, chess playing.
  • 19.
    Designing a LearningSystem – Components and Process • Steps: Data Collection → Preprocessing → Model Selection → Training → Evaluation → Deployment. • Model Evaluation uses test data and performance metrics. • Iterative process with monitoring and retraining.
  • 20.
    Issues in MachineLearning – Noise, Overfitting, Bias-Variance • Noise: Irrelevant/misleading data, affects generalization. • Overfitting: High training accuracy but poor test performance. • Bias-Variance: Trade-off to minimize total error.
  • 21.
    MCQs 1 &2 • 1. Who defined ML as 'ability to learn without being explicitly programmed'? A. Tom Mitchell B. Andrew Ng C. Arthur Samuel D. Geoffrey Hinton [Ans: C] • 2. Which component in ML represents the task being solved? A. Experience B. Performance C. Task D. Evaluation [Ans: C]
  • 22.
    MCQs 3 &4 • 3. Which of these is NOT a goal of ML? A. Prediction B. Automation C. Code writing D. Pattern Recognition [Ans: C] • 4. What does the P in T-E-P stand for? A. Performance B. Processing C. Prediction D. Probability [Ans: A]
  • 23.
    MCQs 5 &6 • 5. Which is an example of a classification task? A. Predicting temperature B. Recognizing spam emails C. Forecasting stock price D. Predicting height [Ans: B] • 6. What is bias in ML? A. Model’s sensitivity to training data B. Simplification error C. Random error D. Noise [Ans: B]
  • 24.
    MCQs 7 &8 • 7. Which technique reduces model complexity? A. Gradient Descent B. Regularization C. Dropout D. Backpropagation [Ans: B] • 8. What is the goal of Find-S algorithm? A. Find all consistent hypotheses B. Eliminate inconsistent examples C. Find most general hypothesis D. Find most specific hypothesis [Ans: D]
  • 25.
    MCQs 9 &10 • 9. Which algorithm ignores negative examples? A. Candidate-Elimination B. List-Then-Eliminate C. Find-S D. Decision Trees [Ans: C] • 10. What does '?' symbol mean in hypotheses? A. No match B. Match anything C. Error D. Specific value [Ans: B]
  • 26.
    MCQs 11 &12 • 11. Which boundary is most general in Candidate-Elimination? A. S B. T C. G D. H [Ans: C] • 12. What is inductive bias? A. Noise in data B. Assumptions guiding learning C. Incorrect output D. Inference process [Ans: B]
  • 27.
    MCQs 13 &14 • 13. Which algorithm explicitly maintains hypothesis space? A. Find-S B. Candidate-Elimination C. List-Then-Eliminate D. Linear Regression [Ans: C] • 14. What happens to G set when a positive example fails? A. Shrink B. Expand C. Stay same D. Empty [Ans: A]
  • 28.
    MCQs 15 &16 • 15. Which strategy involves refining S and G sets? A. List-Then-Eliminate B. Find-S C. Candidate-Elimination D. Clustering [Ans: C] • 16. What is overfitting? A. High test accuracy B. Good generalization C. Fits noise too well D. Undertrained model [Ans: C]
  • 29.
    MCQs 17 &18 • 17. Which error can never be reduced? A. Bias B. Variance C. Irreducible Error D. Prediction Error [Ans: C] • 18. Which is a core concept of concept learning? A. Feature extraction B. Hypothesis search C. Error analysis D. Training accuracy [Ans: B]
  • 30.
    MCQs 19 &20 • 19. Which algorithm is computationally intensive? A. Find-S B. Candidate-Elimination C. List-Then-Eliminate D. SVM [Ans: C] • 20. Which task involves boolean-valued functions? A. Regression B. Clustering C. Concept Learning D. Prediction [Ans: C]
  • 31.