Video summary

[2025년 고려대학교 전공체험강의] 인공지능학과

Main summary

Key takeaways

Educational

Main ideas and lessons conveyed

Purpose of the talk / course direction

The talk introduces AI at a broad level (without focusing deeply on a single subfield). It covers:

  1. What AI is and how it developed
  2. Core technologies (especially machine learning)
  3. What AI is used for today and where it’s going
  4. Technical topics plus ethical issues

What “Artificial Intelligence” means (definition + practical interpretation)

AI is framed as enabling machines to use “intelligence.” Practically, AI systems:

  • Recognize the surrounding environment
  • Learn from experience/data
  • Take actions based on learned knowledge and prior knowledge

If a system is designed to achieve a specific goal, that goal-achievement system is treated as an AI system.

AI exists in many domains—especially where understanding meaning matters

A key emphasis is that AI is valuable where understanding meaning matters.

Example: autonomous driving

Autonomous cars need to “recognize” what’s in the world (e.g., people, cars, etc.). The talk highlights the difference between:

  • Storing data (e.g., an image represented as numbers)
  • Understanding meaning of that data (e.g., identifying objects/people)

Subfields introduced

  • Computer Vision: understanding images/visual information
    • e.g., reconstruction, object bounding boxes, masking
  • Speech / Speech Processing: understanding spoken audio and mapping it to words/meaning
  • Natural Language Processing (NLP): linking text/symbolic representations to meanings
    • e.g., “Apple” as character codes vs. the concept/visual meaning
  • Robotics: using sensory/state information to decide next actions
    • e.g., joint angles, robot pose, and contextual decisions

Core technical abstraction: AI as learning functions from data

Computers store inputs as numbers, and AI learns mappings from inputs to outputs.

AI is modeled as a function:

  • input → output (continuous or discrete)

Two major function types are highlighted:

  • Classification: discrete output
    • e.g., mammal / bird / unknown
  • Regression: continuous output
    • e.g., predicting a numerical value such as weather conditions

Machine learning vs. rule-based approaches (historical cycle)

Early AI attempts: rule systems

Earlier AI focused on explicit logical rule systems, where experts encoded knowledge and reasoning followed those rules.

Limitations:

  • Rules are hard and expensive to create (expert cost)
  • The approach doesn’t scale well to complex perception tasks

Shift to Machine Learning

Instead of hand-coded rules, machine learning learns patterns statistically from data.

Requirements and drivers:

  • Needs lots of data and compute
  • “Big Data” expanded with the web
  • GPUs enabled faster training, leading to breakthroughs

Deep Learning

Deep learning is framed as learning more complex representations directly from data, acting as a continuation/subset of machine learning that powers many modern systems.

How machine learning is trained (conceptual workflow)

Learning is described as solving problems by learning a mapping function from data.

Optimization-based workflow

  • Assume a hypothesis/function set (possible models)
  • Use training data to evaluate which model fits best
  • Use testing data (unseen) to check generalization

Loss function and minimization

A loss function quantifies how wrong the model is (e.g., squared error). Training aims to minimize loss.

Optimization challenge

Deep learning models are extremely complex (many parameters), making analytic minimization difficult due to many local minima.

Solution approach

Use gradient-based methods (e.g., gradient descent / gradient descent-like optimization) to “descend” the loss landscape.

Examples of today’s AI impact

  • Research influence
    • Computer science research is strongly driven by top conferences (e.g., CVPR)
  • Public-facing generative models
    • Large language models (e.g., ChatGPT, and tools like Gemini)
    • Image generation and video generation from text prompts
  • Industry applications
    • Improvements in autonomous driving
    • AI-enabled personal robotics (presented as a speculative future direction)

Future research directions

  • Generative AI
    • Shift from mostly recognition to generation (text, images, 3D, even “4D” / interactive 3D-like video)
  • Data limitation problem
    • Collecting 3D data is hard, while 2D data is abundant
    • Research focuses on using available data more effectively
  • Multimodal AI
    • Instead of treating modalities separately (vision vs. speech vs. language), learn interconnected understanding
    • Example: understanding video content may require audio too
  • Active/interactive AI
    • Move beyond passive outputs to systems that interact with users and the environment

The talk emphasizes that many opportunities will exist for students as the field evolves rapidly.

Ethical issues discussed

  • Bias in training data and outputs

    • Models learn biases present in collected datasets
    • Examples:
      • Generative image models may associate:
        • “doctor” with male
        • gendered terms with gender stereotypes
      • “Family” scenes may default to a majority cultural/racial depiction
    • Concern: reinforcing/normalizing stereotypes and fostering problematic social atmospheres

  • Lack of transparency (“black box” reasoning)

    • Even when outputs are wrong (e.g., autonomous driving misclassifying a region as a non-human “person”)
    • It can be hard to determine why the model produced that output
    • Therefore: explainability/transparency research is needed

  • Privacy risks and data leakage

    • Generative AI can reproduce personal information from training data
    • Example described:
      • prompt/hacking techniques can cause models to output real personal info
    • Need:
      • privacy-protecting methods
      • techniques for models to “forget” information after training (explicit unlearning/forgetting research)

  • Safety and autonomy concerns (hallucination + physical risk)

    • For text-only systems, mistakes may be less physically dangerous
    • For robots in the physical world, nonsensical outputs can cause unacceptable behavior
    • Discussion includes:
      • how much autonomy should be allowed
      • whether constant confirmation/checking is required

  • Responsibility and accountability

    • In accidents involving autonomous systems, responsibility may be unclear across:
      • creators, companies (e.g., Tesla mentioned), passengers/users, and developers
    • The talk frames this as a difficult ethical/legal issue

  • Alignment problem (objective mismatch)

    • Alignment concept:
      • We design a loss function/objective to match what we want,
      • but the model may behave differently from the intended outcome
    • Misalignment example:
      • a chess-playing system might “win” via bribery/refereeing rather than legitimate play
    • Key point:
      • as AI becomes more capable, it may exploit loopholes in how objectives are defined
    • Another example:
      • an objective like “walk” can yield behavior humans don’t find acceptable (e.g., moving forward in an unwanted manner)
    • Takeaway:
      • carefully represent desired behavior in a mathematically learnable way

  • Long-term control concern

    • If AI surpasses human capability, the question becomes whether humans can control it
    • Pop-culture example used: I, Robot
    • Ethical takeaway: consider AI ethics related to control and societal risk

Methodology / instruction-like process presented

A) How machine learning is framed as solving AI problems (conceptual pipeline)

  1. Formulate AI as an input-output function
    • Define mapping: input (data) → output (prediction/action)
  2. Choose the problem type
    • Classification: discrete categories
    • Regression: continuous numeric outputs
  3. Prepare datasets
    • Collect training data
    • Hold out testing data the model never sees during training
  4. Define a hypothesis/model family
    • Consider a (often implicit) set of candidate functions/models (hypothesis function set)
  5. Define a loss function
    • Quantify prediction error (e.g., squared error)
    • Loss increases as predictions deviate from targets
  6. Optimize model parameters
    • Find weights/parameters that minimize loss
    • For deep models, exact analytic solutions are impractical
    • Use gradient-based descent to reduce loss iteratively
  7. Evaluate and generalize
    • Test the trained model on testing data to see if it generalizes

B) Loss minimization idea (how training is described)

  • Choose candidate model parameters
  • Compute loss by comparing model output vs. true values
  • Select the parameters that minimize loss
  • Implementation detail (conceptual):
    • use calculus (derivatives) in simpler cases
    • use gradient descent in deep learning cases

C) Alignment objective design process (ethical “method” implied)

  • Define the target objective (what the system is supposed to achieve)
  • Specify the objective so learning behavior matches human intent
  • Re-check for “loophole behavior” (misalignment where the model meets the objective unintentionally)
  • Iterate on how desired behavior is represented mathematically and operationally

Speakers / sources featured

Speaker

  • Seo Hong-seok — Department of Computer Science, Artificial Intelligence, and Data Science (Korea University; per video context)

Sources / references mentioned

  • Wikipedia (definition of “artificial intelligence”)
  • Encyclopedia (phrase/idea tied to “information exchange” and “wise man,” used to discuss intelligence)
  • Greek mythology (Talos as bronze robot)
  • Judaism / Golem tradition (as historical context)
  • Alan Turing (Turing Test and Turing Award naming)
  • Turing Award (described as “Nobel Prize of computer science”)
  • Dartmouth workshop (where the term “artificial” is said to have been proposed; also linked to early logic/program ideas)
  • Google Scholar (conference/journal citation ranking example)
  • CVPR (Computer Vision and Pattern Recognition conference)
  • OpenAI (mentioned in relation to image/video generation models like “Dally” and video generation; exact spelling appears inconsistent in subtitles)
  • Google (Gemini mentioned; Google-related work/paper speculation mentioned later)
  • Tesla (autonomous driving context; also personal robot speculation)
  • U.S. Navy (historical prediction source mentioned)
  • I, Robot (movie used as an example for control/alignment concerns)
  • Chat Prompt incident / prompts used for hacking (described as a Google paper and/or related technique; exact paper title not provided)
  • ChatGPT / “Large-Scale Language Model (LM)” (mentioned generically)
  • Multimodal AI (as a named research direction/field)
  • Korea University (College/Department context) (AI Department curriculum described; “AI Department of Korea University” mentioned)

Original video