ورود به دنیای هوش مصنوعی | جلسه‌ی ۳ | یادگیری ماشین

Main Ideas & Lessons

1) Data is the true “fuel” of AI (and must be high-quality)

• AI value depends on data—not random text/images.
• Examples of valuable datasets:
  • Medical imaging data (MRI/CT scans, etc.)
  • Accounting/business records
  • Web/app activity data (purchases, visited pages)
  • Video/photo/conversation content (e.g., media posted online)
• Businesses often don’t realize the value of their collected data; that data can be used to train AI models.

2) Poor data quality (and missing metadata) ruins AI usefulness

• Even if data is plentiful, it must be stored correctly.
• Medical example: Students provided cases but used low-quality mobile photos of MRI/CT instead of high-resolution images → reducing AI usefulness.
• Infertility example: Images of embryos are valuable, but metadata is equally crucial:
  • which embryo/image belongs to which patient
  • identifiers (name/national ID), date, technique, drugs
  • whether the embryo was transplanted and outcome
• Key point: Train AI with both image data and metadata.
• Practical suggestion: Start storing data today in a format usable for AI.

3) Features (X) and labels (Y) define supervised learning

Every data item has:

• Features (X): measurable inputs
  • tabular data: columns
  • images: pixels
  • sound: time segments and signal intensity
• Labels (Y): the target to predict

Regression labels (continuous)

• Examples: house price, person height, stock price, Bitcoin price, etc.
• Goal: predict numeric values.

Classification labels (discrete)

• Examples: cat/dog, healthy/sick, disease type, etc.

Important data-storage instruction

• Ensure discrete labels are stored as proper numeric/categorical codes, not text.
• Example coding principle (illustrative):
  • 1 = man, 2 = woman, etc. (or 1/2/3 for disease categories)
• Avoid typos or inconsistent label numbering.

4) Generalization vs. memorization (the purpose of ML)

• ML aims to learn patterns that generalize, not merely memorize training examples.
• If you memorize last year’s exam questions, you’ll fail on new variants.
• Generalization is like high-level math contest performance—new questions require new ideas, not rote repetition.
• Example referenced: Maryam Mirzakhani’s ability to solve problems beyond seen patterns.

5) The “most important split test”: train/validation/test separation

To estimate whether a model truly generalizes, the dataset must be split.

Dataset splitting (typical approach described)

• Shuffle the dataset (like shuffling cards).
• Divide into:
  • Training set: ~70%
  • Validation set: ~10%
  • Test set: ~20%

Purpose of each split

• Training set
  • Learn model parameters (and possibly fit the model).
• Validation set
  • Choose good hyperparameters (model complexity choices).
  • Used for model selection by measuring prediction error on validation data.
• Test set
  • Final unbiased evaluation after hyperparameters are fixed.
  • Must be used only once; using it to tune models is “cheating.”

What hyperparameters mean

• Hyperparameters determine model structure/complexity, such as:
  • polynomial degree (e.g., linear vs quadratic vs 10th-order)
  • neural network structure:
    • number of hidden layers
    • number of neurons per layer
    • activation function choices
• Parameters are learned from data (e.g., weights).

Overfitting concept

• Higher complexity models (e.g., very high-degree polynomial) can fit training points too well:
  • training/fit error becomes tiny,
  • but validation/test performance worsens.
• Overfitting = “too many clothes,” memorization rather than learning the underlying pattern.
• Therefore: choose the model with lowest validation error to promote generalization.

6) Prevent data leakage (multiple kinds)

Two major leakage problems were emphasized:

A) Duplicate/near-duplicate individuals split across sets

• If the same patient appears in both training and test (different photos/days of the same person):
  • the model may memorize that person’s label
  • test accuracy becomes meaningless.
• Rule given: splitting should be based on the entity (e.g., person-level), so all of one person’s images stay together in either training/validation/test.

B) Time-series leakage (shuffling destroys temporal meaning)

• For prediction tasks over time (e.g., Bitcoin price in the next 5 minutes):
  • labels depend on the future relative to features.
• If you randomly shuffle rows, you may allow the model to “see the future relationship,” making test performance falsely excellent.
• Rule given: for time-series splits:
  • training should only include earlier timestamps
  • test should include later timestamps that were never seen
  • conceptually: “train up to time T, then test on T+5min/day,” etc.

7) Loss functions / cost functions (how error is measured)

• Instead of “total error” canceling out conceptually, AI uses a loss function to quantify error.

Regression loss / cost functions described

• Mean Absolute Error (MAE)
  • avoids cancellation by using absolute values:
  • sum of |y_pred - y_true| over all samples, then take the mean.
• Mean Squared Error (MSE) (referred to as mean score error)
  • squares differences so negatives become positive and large errors are penalized more:
  • sum of (y_pred - y_true)^2 over samples, then take the mean.

Lesson’s main point

• Choose hyperparameters/models by minimizing a validation loss/cost, and report final results using test loss.

8) Course progression: code comes next session

• The speaker states this lesson focused on concepts and definitions.
• Next session will shift to coding in practice, solving problems more hands-on (e.g., Kaggle-style tasks mentioned).

---

Speakers / Sources Featured

• Unnamed primary speaker: the course instructor (speaks throughout; references prior sessions and teaches ML concepts).
• Mr. Leinster Walzakh: cited as a source of a quote about talking not being valuable (the name appears distorted in subtitles).
• Maryam Mirzakhani: referenced as an example for generalization in mathematics.
• YouTube commenter(s): anonymous comment(s) said “My eyes hurt” (quoted via the instructor).
• Sarina Esmailzadeh: dedicated to (mentioned in closing).
• Taha Hoshyar: dedicated to (mentioned in closing).