Download Unit 3 PDF Notes
Download the complete handwritten-style PDF notes for CS601 Machine Learning Unit 3. These notes are prepared for RGPV students for fast revision, PYQ-based preparation and 7-mark/14-mark answer writing.
Read Document Notes Important Questions PYQ AnalysisUnit 3 Official Syllabus
Convolutional Neural Network, flattening, subsampling, padding, stride, convolution layer, pooling layer, loss layer, dense layer, 1x1 convolution, inception network, input channels, transfer learning, one-shot learning, dimension reduction, implementation of CNN like TensorFlow and Keras.
Why Unit 3 is Important?
Unit 3 is one of the most scoring units of CS601 Machine Learning because it focuses on Convolutional Neural Networks, also known as CNNs. CNNs are widely used in image recognition, face detection, object detection, medical image analysis, self-driving cars and computer vision.
In RGPV exams, CNN architecture, convolution layer, pooling layer, padding, stride, transfer learning and TensorFlow/Keras implementation are frequently asked topics. If students prepare CNN architecture with diagrams, they can easily write strong 7-mark and 14-mark answers.
1. Convolutional Neural Network
Convolutional Neural Network is a deep learning neural network mainly designed for processing image data. It automatically extracts important features from images using convolution and pooling operations.
Exam Definition
A Convolutional Neural Network is a deep neural network that uses convolution layers, pooling layers and dense layers to extract spatial features from images and perform classification or recognition tasks.
Input Image ↓ Convolution Layer ↓ Activation Function ↓ Pooling Layer ↓ Flattening ↓ Dense Layer ↓ Output
Why CNN is Needed?
- Traditional neural networks cannot efficiently handle large image data.
- CNN automatically extracts features from images.
- CNN reduces manual feature engineering.
- CNN is highly effective for computer vision tasks.
2. Convolution Layer
The convolution layer is the core layer of CNN. It applies filters on the input image to extract useful features such as edges, corners, shapes and textures.
Input Image ↓ Filter / Kernel ↓ Feature Map
Important Terms
- Filter: Small matrix used to detect features.
- Kernel: Another name for filter.
- Feature Map: Output produced after applying filter.
Example
In an image of a face, convolution layer can detect eyes, nose, edges and facial boundaries.
3. Pooling Layer and Subsampling
Pooling layer reduces the size of feature maps while preserving important information. It helps reduce computation and controls overfitting.
Types of Pooling
| Pooling Type | Meaning |
|---|---|
| Max Pooling | Selects maximum value from selected region. |
| Average Pooling | Takes average value from selected region. |
Feature Map ↓ Pooling ↓ Reduced Feature Map
4. Padding
Padding means adding extra pixels around the border of an image before applying convolution. It helps preserve image size and prevents loss of border information.
Original Image ↓ Add Zero Border ↓ Padded Image ↓ Convolution
Types of Padding
- Valid Padding: No padding is added.
- Same Padding: Padding is added to keep output size same as input.
Importance
- Protects edge information.
- Controls output size.
- Improves feature extraction.
5. Stride
Stride defines how many pixels the filter moves at each step during convolution. Larger stride reduces output size, while smaller stride gives more detailed feature maps.
Stride = 1 → Filter moves 1 pixel Stride = 2 → Filter moves 2 pixels
Effect of Stride
| Stride | Effect |
|---|---|
| Small Stride | More detailed output |
| Large Stride | Smaller output size |
6. Flattening
Flattening converts a multidimensional feature map into a one-dimensional vector. This vector is passed to dense layers for classification.
2D Feature Maps ↓ Flattening ↓ 1D Vector ↓ Dense Layer
Why Flattening is Needed?
- Dense layers accept one-dimensional input.
- It connects convolutional part with classification part.
- It prepares extracted features for final prediction.
7. Dense Layer and Loss Layer
Dense layer is a fully connected layer where each neuron is connected to every neuron of the previous layer. It performs final classification based on extracted features.
Loss Layer
Loss layer calculates the difference between predicted output and actual output. The model uses this error to improve training.
Flattened Features ↓ Dense Layer ↓ Prediction ↓ Loss Layer ↓ Error Calculation
8. Input Channels
Input channels represent the number of color channels in an image. A grayscale image has one channel, while a colored RGB image has three channels: Red, Green and Blue.
| Image Type | Channels |
|---|---|
| Grayscale Image | 1 Channel |
| RGB Image | 3 Channels |
RGB Image ↓ Red Channel Green Channel Blue Channel
9. 1x1 Convolution
1x1 convolution uses a filter of size 1 by 1. It is mainly used to reduce or increase the number of channels without changing the spatial size of feature maps.
Uses of 1x1 Convolution
- Dimension reduction
- Channel mixing
- Computational efficiency
- Used in Inception Network
10. Inception Network
Inception Network is a CNN architecture that applies multiple filter sizes in parallel, such as 1x1, 3x3 and 5x5 filters. It allows the network to learn features at different scales.
Input ↓ 1x1 Convolution 3x3 Convolution 5x5 Convolution Pooling ↓ Concatenate ↓ Output
Advantages
- Extracts multi-scale features.
- Improves accuracy.
- Uses 1x1 convolution for efficiency.
- Reduces computational cost.
11. Transfer Learning
Transfer Learning is a technique where a pre-trained model is reused for a new but related task. Instead of training a model from scratch, we use knowledge learned from a large dataset.
Pre-trained Model ↓ Reuse Learned Features ↓ Fine-Tune on New Dataset ↓ New Prediction Task
Example
A CNN trained on ImageNet can be reused for classifying medical images, animals or handwritten digits with small modifications.
Advantages
- Requires less data.
- Saves training time.
- Improves accuracy.
- Useful when dataset is small.
12. One-Shot Learning
One-Shot Learning is a machine learning technique where a model learns from only one or very few examples. It is useful when large datasets are not available.
Example
Face recognition systems can identify a person after seeing only one image of that person.
One Example ↓ Feature Comparison ↓ Similarity Matching ↓ Prediction
Applications
- Face recognition
- Signature verification
- Medical diagnosis
- Security systems
13. Dimension Reduction
Dimension reduction is the process of reducing the number of input features while keeping important information. It helps reduce computation and improves model performance.
Need of Dimension Reduction
- High-dimensional data is difficult to process.
- It reduces memory usage.
- It improves training speed.
- It helps remove unnecessary features.
High-Dimensional Data ↓ Dimension Reduction ↓ Important Features ↓ Efficient Model
Common Method
PCA, or Principal Component Analysis, is one of the most common dimension reduction techniques.
14. CNN Implementation using TensorFlow and Keras
TensorFlow and Keras are popular deep learning libraries used to build and train CNN models. Keras provides a simple high-level interface, while TensorFlow handles numerical computation.
Basic CNN Implementation Flow
- Import TensorFlow and Keras libraries.
- Load dataset.
- Preprocess images.
- Create CNN model.
- Add convolution and pooling layers.
- Add flatten and dense layers.
- Compile model.
- Train model.
- Evaluate accuracy.
model = Sequential() model.add(Conv2D(32, (3,3), activation='relu')) model.add(MaxPooling2D(pool_size=(2,2))) model.add(Flatten()) model.add(Dense(128, activation='relu')) model.add(Dense(10, activation='softmax'))
Unit 3 PYQ Trend Analysis
| Topic | PYQ Frequency | Importance |
|---|---|---|
| CNN Architecture | Repeated | ★★★★★ |
| Convolution and Pooling Layers | Repeated | ★★★★★ |
| Padding and Stride | Frequently Asked | ★★★★ |
| Transfer Learning | Often Asked | ★★★★ |
| Inception Network | Moderate | ★★★ |
| TensorFlow and Keras | Recently Asked | ★★★★ |
| Dimension Reduction | Repeated in ML papers | ★★★★★ |
Top 20 Important Questions for Unit 3
- Explain Convolutional Neural Network with neat diagram.
- Explain CNN architecture in detail.
- Explain convolution layer and feature map.
- Explain pooling layer and its types.
- Explain padding and stride in CNN.
- Explain flattening in CNN.
- Explain dense layer and loss layer.
- Explain input channels in image data.
- Explain 1x1 convolution and its uses.
- Explain Inception Network.
- Explain transfer learning with example.
- Explain one-shot learning.
- Explain dimension reduction in Machine Learning.
- Explain CNN implementation using TensorFlow and Keras.
- Differentiate convolution layer and pooling layer.
- Differentiate padding and stride.
- Explain how CNN extracts hierarchical features.
- Explain why pooling reduces spatial dimensions.
- Explain role of CNN in computer vision.
- Write applications of CNN.
Most Expected Questions 2026
- 95%: Explain CNN architecture with neat diagram.
- 90%: Explain convolution layer, pooling layer and dense layer.
- 90%: Explain padding, stride and flattening.
- 85%: Explain transfer learning and one-shot learning.
- 85%: Explain implementation of CNN using TensorFlow and Keras.
- 80%: Explain 1x1 convolution and Inception Network.
24-Hour Revision Box
CNN ↓ Image Learning Network Convolution Layer ↓ Feature Extraction Filter / Kernel ↓ Detects Edges and Patterns Pooling Layer ↓ Reduces Size Padding ↓ Protects Border Information Stride ↓ Filter Movement Flattening ↓ 2D to 1D Vector Dense Layer ↓ Final Classification 1x1 Convolution ↓ Channel Reduction Inception Network ↓ Multiple Filters in Parallel Transfer Learning ↓ Reuse Pre-trained Model One-Shot Learning ↓ Learn from Few Examples Dimension Reduction ↓ Reduce Features TensorFlow / Keras ↓ CNN Implementation
How to Write 7-Mark Answer?
For a 7-mark Unit 3 answer, write definition, diagram, working, advantages and applications. Keep answer around 2 to 3 pages in exam copy. Diagrams are very important for CNN answers.
How to Write 14-Mark Answer?
For a 14-mark answer, write introduction, definition, full architecture diagram, detailed working, layers explanation, advantages, disadvantages, applications and conclusion. CNN architecture and auto feature extraction should be clearly explained.
Related Machine Learning Units
FAQs
What is CNN in Machine Learning?
CNN is a deep learning model used mainly for image processing and computer vision tasks.
Which is the most important topic in Unit 3?
CNN architecture with convolution layer, pooling layer, flattening and dense layer is the most important topic.
Why is pooling used in CNN?
Pooling reduces the size of feature maps and decreases computational cost while preserving important features.
What is transfer learning?
Transfer learning reuses a pre-trained model for a new related task, reducing training time and data requirement.
Is TensorFlow and Keras important for RGPV exams?
Yes, implementation steps of CNN using TensorFlow and Keras can be asked as a 7-mark question.