What is the primary role of Pooling layers in CNNs?

Pooling layers downsample spatial feature maps, reducing parameter count and computational complexity while granting translation invariance.

How do Transformers solve the vanishing gradient problem in NLP?

Transformers replace sequential recurrence with Self-Attention, allowing direct connections between all word pairs regardless of position, avoiding gradient degradation.

Home ASSEB AI HubUnit 2: Deep Neural Networks

Class XIIASSEB-527Unit 2: Deep Neural Networks

Unit 2: Deep Neural Networks

Master official ASSEB Class XII AI Unit 2: Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Attention Mechanism, Transformers, and PyTorch MNIST.

Theory: 8 Marks

•

Practical: 5 Marks

•

13 min read

Subscription Gate Active

Learning Outcomes

Understand Deep Neural Network architectures and how hierarchical representations emerge across depth.
Master Convolutional Neural Networks (CNNs): kernel filters, stride, padding, and pooling layers (Max Pooling).
Understand Sequence Modeling: Recurrent Neural Networks (RNNs), hidden state persistence, and the Vanishing Gradient problem.
Explore gated recurrent architectures (LSTM and GRU) and how memory cells resolve long-term dependencies.
Analyze the Transformer architecture: Self-Attention mechanism, positional encoding, and parallel sequence processing.
Implement a CNN digit classification model on the MNIST dataset using the PyTorch framework.

Board Exam Scoring Tips

When explaining CNNs in 5-mark questions, draw the standard pipeline: Input Image -> Conv2D -> ReLU -> MaxPool2D -> Flatten -> Dense -> Softmax.
In 3-mark questions on RNN vs. Transformer, state: Transformers process all words simultaneously using Self-Attention, whereas RNNs process words sequentially.
Remember that Pooling layers do NOT contain learnable weights; they perform fixed mathematical subsampling (e.g., maximum or average).

Key Takeaways

CNNs exploit spatial locality using sliding parameter-sharing kernels and pooling layers to reduce computational complexity.
Standard RNNs process sequences step-by-step; repeated multiplication of small gradients during BPTT (Backpropagation Through Time) causes vanishing gradients.
LSTMs introduce input, forget, and output gates to preserve long-range dependencies.
Transformers eliminate recurrence entirely, using Self-Attention to process entire sequences in parallel.
PyTorch is the official deep learning framework highlighted in the ASSEB Class XII syllabus for hands-on tensor training.

Practice Unit MCQ Quiz

Back to ASSEB AI Hub

# Python Errata Fix Sample...

Premium ASSEB Syllabus

Unlock Unit 2: Deep Neural Networks

Access verified code errata manual, 50 external viva questions, period lesson scripts, and practical examination rubrics.

Student Pass₹149/mo

Full Class XIIunits, non-math question bank & Doubt Desk.

Teacher Pass₹299/mo

All XI & XII units, 45-min lesson scripts & 50-mark rubrics.

Single Chapter₹49 once

Instant permanent unlock for this specific chapter and errata code.

Your Email (for watermarking)

WhatsApp Phone

256-Bit Encrypted Payment • Supports UPI, Cards & NetBanking

Previous UnitUnit 1: Artificial Neural Networks Next UnitUnit 3: Generative AI