Synetic College Logo
NotesB.Sc IT (AI)Deep Learning & Computer Vision
Year 2 · Semester 1

Deep Learning & Computer Vision Notes

Program:B.Sc IT (AI)
Semester:Semester 3
Hours:15
Units:8 Units

Subject Overview

Year 2 (Deep Learning & Language) moves into neural networks — how modern AI actually works under the hood, from a single neuron to transformers. This semester's goal is to understand neural networks deeply enough to design, train and debug them for vision tasks: fundamentals and PyTorch, training technique, convolutional networks, transfer learning, and advanced computer vision.

Subject at a glance

8

Units

33

Topics

15

Hours

8

Infographics

Unit-wise Syllabus

8 units — click WhatsApp below to get the full notes for each

1

Unit 1: Neural Network Fundamentals (3 hours)

  • Perceptrons
  • activation functions
  • forward and backward propagation
  • building a neural network from scratch in NumPy
ProcessHow a neural network learns
  1. 1Forward pass

    Inputs × weights + bias, then activation

  2. 2Compute loss

    Compare prediction with truth

  3. 3Backward pass

    Gradients via the chain rule

  4. 4Update weights

    Optimizer step

  5. 5Repeat for many epochs

Study guide

This module builds the foundation of deep learning: perceptrons, activation functions, forward and backward propagation and building a neural network from scratch in NumPy. Building it yourself once makes every framework make sense.

2

Unit 2: PyTorch Essentials (3 hours)

  • Tensors
  • autograd
  • nn.Module
  • datasets and dataloaders
  • training loops
  • why PyTorch dominates research
CycleThe PyTorch training loop
The PyTorch training loop
1Load a batch
2Forward
3Loss
4Backward
5Step
6Reset
  1. 1. Load a batch: DataLoader
  2. 2. Forward: outputs = model(x)
  3. 3. Loss: loss = criterion(...)
  4. 4. Backward: loss.backward()
  5. 5. Step: optimizer.step()
  6. 6. Reset: optimizer.zero_grad()

Study guide

You learn PyTorch: tensors, autograd, nn.Module, datasets and dataloaders, training loops and why PyTorch dominates research. Write a full training loop by hand before using any helper library.

3

Unit 3: Training Deep Networks (3 hours)

  • Optimizers (SGD, Adam)
  • learning rate schedules
  • regularization
  • batch norm
  • dropout
  • debugging training
ComparisonFixing common training problems
Symptom
Try

Overfitting

Train good, validation poor

Dropout, weight decay, augmentation, more data

Underfitting

Both poor

Bigger model, train longer, better features

Unstable loss

Loss jumps or becomes NaN

Lower learning rate, batch norm, gradient clipping

Slow progress

Loss falls very slowly

Adam, learning rate schedule

Study guide

Training deep networks is part craft: optimisers (SGD, Adam), learning rate schedules, regularisation, batch norm, dropout and debugging training. Learn to read loss curves to diagnose overfitting and underfitting.

4

Unit 4: Convolutional Neural Networks (4 hours)

  • Convolutions
  • pooling
  • classic architectures (LeNet, ResNet, EfficientNet)
  • image classification end-to-end
ProcessInside a CNN
  1. 1

    Input image

    e.g. 224 × 224 × 3

  2. 2

    Convolution

    Filters detect edges and textures

  3. 3

    Activation

    ReLU

  4. 4

    Pooling

    Downsample, keep strongest signals

  5. 5

    Deeper conv blocks

    Shapes, then objects

  6. 6

    Fully connected + softmax

    Class probabilities

Study guide

Convolutional neural networks power image understanding: convolutions, pooling, classic architectures (LeNet, ResNet, EfficientNet) and end-to-end image classification. Understand why convolutions suit images.

5

Unit 5: Transfer Learning & Fine-tuning (3 hours)

Using pretrained models from Hugging Face and torchvision, feature extraction versus fine-tuning

ComparisonFeature extraction vs fine-tuning
Feature extraction
Fine-tuning

Pretrained layers

Frozen

Some or all unfrozen

Data needed

Little

More

Compute

Low

Higher

Best when

New task is similar to the original

New domain differs, e.g. medical images

Study guide

Transfer learning reuses what pretrained models have learned: using models from Hugging Face and torchvision, and feature extraction versus fine-tuning. Fine-tuning a pretrained model usually beats training from scratch on small data.

6

Unit 6: Advanced CV Tasks (4 hours)

  • Object detection (YOLO)
  • segmentation (U-Net, SAM)
  • image generation basics and diffusion intuition
ComparisonAdvanced vision tasks
Output
Popular model

Classification

One label per image

ResNet, EfficientNet

Object detection

Boxes + labels

YOLO

Segmentation

A label for every pixel

U-Net, SAM

Generation

New images

Diffusion models

Study guide

Advanced computer vision covers object detection (YOLO), segmentation (U-Net, SAM), image generation basics and diffusion intuition. Run a detector on your own photos.

7

Unit 7: Tools introduced

  • PyTorch
  • Hugging Face transformers and datasets
  • Google Colab Pro / Kaggle GPUs for training
  • Roboflow for dataset preparation
  • Cursor with Claude in agentic mode for refactoring training code
ClassificationDeep learning toolkit
Tools

  • PyTorch

    Build and train networks

  • Hugging Face

    Pretrained models and datasets

  • Colab and Kaggle GPUs

    Free or cheap compute

  • Roboflow

    Label and prepare image data

  • AI coding agents

    Refactor training code

Study guide

The tools are PyTorch, Hugging Face transformers and datasets, Colab Pro and Kaggle GPUs, Roboflow for dataset preparation and Cursor with Claude in agentic mode for refactoring training code. Learn to prepare and label data well, since data quality drives results.

8

Unit 8: Mini-projects

  • Build a CNN from scratch to classify CIFAR-10
  • fine-tune a ResNet for a custom domain (e.g. plant disease detection)
  • deploy an object detector to a webcam feed using YOLOv8
ProcessPlant disease detector project
  1. 1Collect and label leaf images
  2. 2Augment

    Flips, crops, colour jitter

  3. 3Fine-tune a pretrained ResNet
  4. 4Evaluate

    Confusion matrix per disease

  5. 5Deploy

    Webcam or phone demo

Study guide

The mini-projects are a CNN from scratch for CIFAR-10, a fine-tuned ResNet for a custom domain such as plant disease detection and a YOLOv8 object detector on a webcam feed. A working webcam demo is a striking portfolio piece.

Why Deep Learning & Computer Vision Matters

Deep learning drives modern computer vision, speech and language systems. Skills in PyTorch, transfer learning and detection are in high demand in AI engineering, robotics, healthcare imaging and agriculture technology.

How to Prepare for Deep Learning & Computer Vision

  • Build a network from scratch once, then move to PyTorch.
  • Monitor training and validation loss to catch overfitting early.
  • Fine-tune pretrained models on a dataset you collect yourself.

Deep Learning & Computer Vision: Frequently Asked Questions

What is transfer learning?

Transfer learning reuses a model trained on a large dataset as the starting point for a new task, which saves data and time.

Why is PyTorch popular?

PyTorch is flexible, easy to debug and widely used in research, which makes new methods available in it first.

Want the Complete Notes & Past Papers?

WhatsApp us and we'll send you the full notes, question banks, and previous year papers for Deep Learning & Computer Vision.