Video summary

Kecerdasan Buatan - Week 4

Main summary

Key takeaways

Educational

Main ideas & lessons

Machine Learning (ML) in AI

  • ML is programming computers to optimize performance by using data examples or past experience.
  • ML is a subset of AI that learns patterns from data.
    • In the lecturer’s framing, it does not truly “learn its own data” creativity/innovations the way humans do.
  • The course positions this as Week 4 content, continuing from earlier weeks on AI and learning paradigms.

Three learning paradigms (learning approaches)

The lecturer emphasizes feedback type as the main distinction:

  • Supervised learning: uses labels
  • Unsupervised learning: uses no labels
  • Reinforcement learning: uses rewards/punishments

Later, semi-supervised learning is mentioned as a 4th category:

  • Semi-supervised: uses a mix of labeled + unlabeled data to improve supervised learning.

Importance of data

  • “Garbage in, garbage out”: if the dataset is bad/rotten, even the best model can’t produce good results.
  • Data quality and representativeness are repeatedly stressed as more decisive than choosing a sophisticated algorithm.
  • Dataset validation includes checking that training and test sets have consistent “nature” (i.e., match the real-world population/distribution).

Generalization vs memorization

  • ML success is measured by handling new/unseen data (test), not just high training performance.
  • Overfitting (conceptually) is discussed critically:
    • Often defined as higher training performance than test performance.
    • The lecturer argues it may not be “ignorance of knowledge,” but rather distribution/nature differences between train and test.
  • Overfitting doesn’t always imply the model is inherently bad—especially if test performance remains competitive.

Definitions: AI vs ML vs Deep Learning (quick distinctions)

  • AI: the broad field of artificial intelligence.
  • Machine Learning: AI that can learn from data.
  • Deep Learning: ML that typically avoids manual feature extraction (often described as “end-to-end” learning), enabling more autonomy.

Conventional programming vs ML workflow

  • Conventional programming:
    • uses data + explicit rules/code to produce outputs.
  • ML workflow:
    • involves training (learn model/rules from data),
    • then inference (produce outputs on new inputs).

ML pipeline / workflow (stages for research/thesis)

The lecturer outlines an ML workflow that supports both research writing and practical execution:

  1. Problem formulation
  2. Data collection
    • Prefer existing datasets if available.
    • Collecting new data is difficult and must be valid/representative.
  3. Pre-processing (especially for non-deep-learning ML)
  4. Feature extraction / features
  5. Model selection
  6. Error function / loss
  7. Evaluation (depends on the task)
  8. Implementation & monitoring (inference)
    • may extend to real-world deployment and further research

Task types (“bags”) and ML outputs

  • Tasks are framed as separate “buckets” (focused problem types).
  • The lecturer argues specialized models for a single bucket often outperform general ones.

Classification

  • Predicts a class label (often binary, e.g., diseased vs not diseased).
  • Classes should be specific (example logic: “kidney disease” vs other diseases—one bucket = one task that can’t be further broken down).

Regression

  • Predicts numeric/continuous outcomes from historical patterns.
  • Related study areas mentioned: ARIMA/SARIMA and linear regression.

Segmentation

  • Predicts regions (e.g., tumor areas in MRI).
  • Data dimensionality notes:
    • 3D: spatial dimensions (length/width/height)
    • 4D: includes time (e.g., video/temporal data)
  • Common in health/medical research and may be combined with classification.

Methodologies & instruction-like lists

A) Types of learning (paradigms)

  • Supervised learning

    • Inputs: X
    • Targets/labels: Y
    • Goal: learn a mapping f(X) → Y

  • Unsupervised learning

    • Inputs: X only
    • No target labels
    • Goal: discover structure/patterns (e.g., clustering, dimensionality reduction, association rules)

  • Semi-supervised learning

    • Uses unlabeled data to improve performance of supervised learning
    • Typically treated as a supplement to labeled training

  • Reinforcement learning

    • An agent interacts with an environment
    • Receives rewards / punishments (including negative rewards)
    • Goal: maximize reward via repeated trial-and-error
    • Example mentioned: Q-learning

B) ML research/thesis stages (workflow)

  1. Problem formulation

    • Define the research objective/task precisely.

  2. Data collection

    • If a valid existing dataset exists, use it.
    • Otherwise, consider data crawling/retrieval, but ensure validity and representativeness.
    • Validate whether the dataset reflects the real-world distribution.

  3. Pre-processing (especially for non-deep-learning ML)

    • Lecturer claims preprocessing may reduce performance in some cases for deep learning (end-to-end reduces need).
    • Examples mentioned:
      • cleaning/removing “empty values”
      • normalization (e.g., scale 0–10 → 0–1)
      • encoding (assigning values to abstract items like words/graphs/user personas)

  4. Features

    • Determine what representation X uses.
    • Emphasis: feature quality affects model quality.

  5. Model selection

    • Choose model type based on the task:
      • regression (e.g., linear regression; ARIMA/SARIMA references)
      • classification (see classification section below)
      • segmentation (image-region prediction)

  6. Error function / loss

    • Defines how far predictions deviate from the target.

  7. Evaluation

    • Use test data not seen during training to assess generalization.
    • Metrics depend on task:
      • Classification: accuracy, precision, recall, F1-score (ev-score mentioned)
      • Segmentation: task-specific segmentation metrics (described as “various evaluation techniques”)
      • Regression: RMSE, MSE (and other regression metrics mentioned)

  8. Implementation & monitoring (inference)

    • Turn the model into a usable method/system.
    • Example:
      • build a brain tumor segmentation model, then apply it in a hospital setting
      • can lead to additional real-world research beyond evaluation

C) Examples of unsupervised methods (as described)

  • Clustering

    • Group data by similarity/structure
    • Examples: K-medoids, K

  • Dimensionality reduction

    • PCA (principal component analysis)
    • Lecturer contrasts PCA with an alternative “variance/encoder” approach that may sometimes outperform PCA (based on lecturer experience).

  • Association rules

    • Find patterns like A → B
    • Example scenario: product placement (e.g., toothbrush + toothpaste at Indomaret)
    • Mentioned: Apriori algorithm
    • Distinguished from:
      • clustering (grouping)
      • dimensionality reduction (compressing/projection)

  • Reinforcement learning method

    • Mentioned: Q-learning
    • Agent selects actions; environment returns reward/punishment; policy updates iteratively.

D) Classification algorithms discussed + how they work (conceptually)

  • Naive Bayes

    • Uses probability to infer the most likely class.

  • k-Nearest Neighbors (KNN)

    • Choose class based on k closest neighbors
    • Examples:
      • K = 1: nearest neighbor decides
      • K = 2, 3, …: majority voting among closest points
    • Lecturer observation: K = 1 can sometimes outperform larger K.

  • Support Vector Machine (SVM)

    • Finds a decision boundary (hyperplane) to separate classes
    • Notes extension from binary to multi-class via modifications.

  • Classification “bags” (task scoping)

    • Binary examples: spam/not spam, kidney failure/not kidney failure, diseased/not diseased.
    • Multi-label is possible (separate labels per condition), but the key idea is keeping tasks appropriately scoped.

E) Regression / segmentation examples discussed

  • Regression

    • Predict numeric/historical patterns
    • Mentions linear regression and time-series tools (ARIMA/SARIMA)

  • Segmentation

    • Predict affected regions in images (MRI example)
    • Output includes spatial region labels (e.g., box/circle/areas rather than one class)

Speakers / sources featured (identified)

Primary speaker

  • Course lecturer / instructor (unnamed)
    • Referred to as “Sir” by the lecturer and “I” by the lecturer.

Sources mentioned (non-speaker references)

  • YouTube (example comparison: supervised vs reinforcement learning for car racing)
  • Indomaret (association-rule intuition example)
  • Q-learning
  • Apriori algorithm
  • PCA
  • KNN, Naive Bayes, SVM
  • ARIMA/SARIMA
  • LLM (mentioned in the context of classification comparison and knowledge-capacity boundaries)
  • UAT (user acceptance testing) mentioned in deployment context

Original video