Visual Artificial Intelligence

Course Introduction

Dongwann Kang
Associate Professor
Department of Computer Science and Engineering
Course Introduction
• Lecturer
– Dongwann Kang
• PhD in Computer Graphics
• dongwann@[Link]
• 02-970-6719
• Room #328, Mirae Hall (미래관)

• Class web-page
– e-Class: [Link] (or HelloLMS App.)
– Course information including notices, course notes, any materials are accessible via e-class
Course overview
• Level
– Upper-level undergraduate and entry-level graduate students in Computer Science & Engineering

• Course Topics
– Deep Learning Fundamentals (Optimization, Neural Networks, CNNs, RNNs, Transformers)
– Core Vision Tasks (Image Classification, Object Detection, Facial Expression Recognition)
– End-to-End Model Training and Architectural Design
– Recent Research Papers and Hands-on Experiments in Visual AI

• Course Objectives
– Understand the implementation and training of artificial neural networks
– Learn modern deep learning architectures for visual computing
– Develop the ability to critically analyze recent research in computer vision
– Gain hands-on experience through small-scale experiments and model adaptations
Visual Artificial Intelligence
• Course Coverage
– This course sits at the intersection of Deep
Learning and Computer Vision, within
the broader field of Artificial Intelligence.
Artificial Intelligence

– Focus is placed on applying modern deep

Computer Vision Machine Learning learning architectures to visual tasks,
Deep Learning bridging theory and practice.

– The diagram illustrates how this class

explores the overlapping area:
• Artificial Intelligence (outer field)
• Machine Learning as a subset
• Deep Learning within machine learning
• Computer Vision as an application domain
• Visual AI represents the overlapping region of
deep learning and computer vision.
Camera Obscura: The Desire to Capture Images
• Core Idea
– Humans have always pursued ways to
capture and preserve visual reality.
• Key Points
– Camera Obscura: Latin for “Dark Chamber”
– Inverted image projection, dating back to
antiquity
– Early step toward photography and modern
imaging
• Why It Matters Here 1st published illustration of camera obscura in Gemma Frisius' book
“De Radio Astronomica et Geometrica,” 1545
– Represents the first milestone in humanity’s
journey from capturing light
→ to recording images
→ to understanding them computationally.

From Da Vinci’s Notebook A 19c illustration of a camera obscura box with mirror
From Film to Digital Cameras: Capturing Images at Scale
• Invention of Photography
– Transition from light projection to recording
images on film
• Digital Sensors
– Conversion of light into pixels, enabling
large-scale storage and reproduction
• Key Shift
– Images became data that could be The first Daguerreotype was made by Louis-Jacques-Mandé Daguerre
processed, transmitted, and analyzed by
computers
• Impact
– Opened the door for computational methods
and laid the groundwork for computer vision

Steven J. Sasson, inventor of the first digital camera,

comparing his device with today's digital cameras.
Early Computer Vision: Processing Images
• First Algorithms
– Focused on low-level image processing
(edge detection, filtering, feature extraction)
• Goal
– Enable computers to process visual input,
not just store it
• Limitations
– Rule-based and handcrafted features
struggled with complex, real-world images
• Significance
– Established the foundation of computer
vision research and highlighted the need for
more powerful approaches

The Sobel operator developed in the late 1960s

Deep Learning Era: Learning from Data
• Breakthrough
– Introduction of deep neural networks,
especially Convolutional Neural Networks
(CNNs)
• Key Advantage
– Learned hierarchical features directly from
raw pixels, reducing reliance on handcrafted
rules
Network architecture of AlexNet
• Applications
– Image classification, object detection, facial
recognition, and beyond
• Impact
– Marked the transition from processing
images to learning representations,
dramatically improving performance in visual
tasks
Image classification results
Visual AI Today: Towards Understanding
• Beyond Recognition
– From identifying objects to interpreting
context, relationships, and meaning
• Modern Architectures
– Transformers, multimodal models, and
generative AI expanding visual capabilities
• Applications
– Autonomous driving, medical imaging, facial
expression recognition, content generation
• Key Idea
– Visual AI is evolving from seeing to
understanding, bridging perception and
intelligence
Weekly schedule
• Week 1 • Week 7–8
– Course Introduction – Training Deep Neural Networks
– Recurrent Neural Networks
• Week 2–4
– Deep Learning Fundamentals • Week 9–11
• Image & Linear Classification – Transformers in Vision
• Optimization Basics – Object Detection
• Neural Networks & Backpropagation
– Facial Expression Recognition

• Week 5–6 • Week 12–15

– Convolutional Neural Networks
– Student Paper Presentations
• Basics of CNNs
• Modern CNN Architectures
Evaluation system
• Evaluation Components & Details
– Attendance (20%)
• 2% deducted per unexcused absence. More than 1/3 absence results in automatic fail (F).
– Paper Presentation (40%)
• Each student selects a recent visual AI paper from a given pool and presents its motivation, model
architecture, experimental methods, and contributions.
• Must also propose at least one follow-up experiment.
– Experimental Report (40%)
• Implement and adapt the chosen model to a new but related domain (e.g., applying an image classification
network to facial expression recognition).
• Report includes adaptation process, results, and at least one ablation or comparative study. Optional
improvements to model or training are encouraged.

• Important Note
– Completion of both paper presentation and experimental report is mandatory.
– Failure to submit either will result in a failing grade (F).
Important Notice & Recommendations
• Course Enrollment Policy
– Students are not allowed to take both Visual Artificial Intelligence and Advanced Machine Learning
(Prof. Han) in the same semester.
– These two courses overlap significantly in content, assignments, and evaluation methods (e.g.,
paper presentations).
– Therefore, students must choose one course only.
– If it is later discovered that a student has enrolled in both courses simultaneously, the lowest
possible grade will be assigned.

• Recommendation for Students

– If you already have prior experience in deep learning for vision, I personally recommend taking
Advanced Machine Learning (Prof. Han).
– If you are new to the field or lack foundational knowledge, I recommend taking this course.
• Any question?