Skip to main content

Command Palette

Search for a command to run...

Master Machine Learning

Updated
43 min readView as Markdown
S
I am an AI Research Engineer with a combined motivation of building AI models as well as developing AI integrated apps. I am currently exploring Robot Learning and groundbreaking DL, RL and Robotics papers and trying to understand how this is shaping the future.

The mental model that makes Machine Learning click

Machine Learning can look like a bag of disconnected techniques:

  • regression

  • classification

  • clustering

  • dimensionality reduction

  • cross-validation

  • regularization

  • metrics

  • calibration

  • feature engineering

But almost everything revolves around one idea:

Instead of explicitly programming every rule, give a machine examples and let it learn a useful function from data.

Traditional programming often looks like:

Rules + Data
     ↓
Program
     ↓
Answer

Machine learning looks more like:

Data + Answers/Structure
        ↓
Learning Algorithm
        ↓
Learned Model
        ↓
Predictions on new data

For example, instead of programming:

IF ears pointed
AND whiskers present
AND face shape ...
THEN cat

we provide thousands of labeled cat/dog images.

The model learns a function:

image → probability of cat

The word learn here has a precise engineering meaning:

Adjust internal parameters so that predictions become better according to some objective.

That is Machine Learning.


1. The Core ML Problem

Suppose we observe an input:

x

and want to predict something:

y

Machine learning tries to learn a function:

f(x) ≈ y

Examples:

house information → house price

email → spam / not spam

image → object class

speech → transcript

customer behavior → churn probability

We usually do not know the perfect function.

Instead, we select a family of functions:

fθ(x)

where:

θ = model parameters

Training means finding parameters that make the model perform well on data.

Conceptually:

Choose θ
   ↓
Make predictions
   ↓
Measure errors
   ↓
Adjust θ
   ↓
Repeat

2. The Most Important Idea in Machine Learning

The real goal is not:

Perform perfectly on the examples you already have.

The real goal is:

Learn patterns that still work on new examples you have never seen.

This is called generalization.

A student who memorizes every answer from a practice exam may score perfectly on those exact questions.

But if they cannot solve slightly different questions, they did not actually learn the subject.

Machine learning has exactly the same problem.

Therefore:

Training performance ≠ real success

What matters is performance on unseen data.

Almost the entire discipline of practical ML exists because generalization is difficult.


3. Supervised Learning

Supervised learning means the training data contains both:

input x

and:

correct target y

Example dataset:

House size     Price

800 sqft       $120k
1200 sqft      $180k
1500 sqft      $230k
2000 sqft      $310k

The model sees examples of:

x → y

and learns the relationship.

Another example:

Email text                  Label

"Claim your prize!"         spam
"Meeting tomorrow at 2"     not spam

The word supervised means the desired answer is available during training.

Think:

Teacher provides questions + answers.

4. The Two Major Supervised Problems

Most introductory supervised learning problems fall into:

REGRESSION

or:

CLASSIFICATION

The distinction is simple.


5. Regression

Regression predicts a continuous numerical value.

Examples:

house → $243,000

temperature sensors → 23.7°C

vehicle features → fuel consumption

patient features → estimated blood pressure

Output:

number

A simple regression model might learn:

house price
=
base price
+
size contribution
+
location contribution
+
age contribution

6. Linear Regression

Suppose house price approximately increases with house size.

We could model:

ŷ = b₀ + b₁x

where:

x  = house size
ŷ  = predicted price

b₀ = intercept
b₁ = slope

The model might learn:

Price = 50,000 + 150 × square_feet

For:

1000 sqft

prediction:

50,000 + 150 × 1000

= $200,000

Training determines the best values of:

b₀
b₁

from the data.

Regression is essentially:

Learn a numerical relationship between inputs and outputs.


7. Residuals

Suppose:

true house price = $220,000
prediction       = $200,000

The difference:

true - prediction

is a residual.

residual = 20,000

Training often tries to make these residuals small.

This naturally leads to loss functions later.


8. Classification

Classification predicts a category.

Examples:

image → cat / dog

email → spam / not spam

transaction → fraud / legitimate

medical scan → disease / no disease

Classification models often output probabilities.

Example:

Cat: 0.92
Dog: 0.08

Then a decision rule may convert probabilities to labels.

For binary classification:

P(cat) ≥ 0.5 → predict cat
P(cat) < 0.5 → predict dog

But 0.5 is not sacred.

Changing the threshold changes system behavior.

This becomes extremely important when false positives and false negatives have different costs.


9. Classification Is Usually Probability Estimation + Decision Rule

A useful engineering view is:

Input
  ↓
Model
  ↓
Probability / score
  ↓
Decision threshold
  ↓
Final class

Example:

Transaction
      ↓
P(fraud) = 0.71
      ↓
threshold = 0.60
      ↓
BLOCK

Another company might use:

threshold = 0.90

and allow the same transaction.

Same model.

Different operational decision.

This distinction is extremely important:

Prediction and decision are not always the same thing.


10. Unsupervised Learning

What if you have data but no labels?

Example:

Customer 1: behavior vector
Customer 2: behavior vector
Customer 3: behavior vector
...

but nobody has provided:

"premium customer"
"price-sensitive customer"
"occasional customer"

Then we enter unsupervised learning.

The system tries to discover useful structure without explicit target labels.

Typical tasks:

clustering
dimensionality reduction
density estimation
anomaly detection
representation learning

Think:

Supervised:
"Here are examples and their answers."

Unsupervised:
"Here is the data. Discover structure."

11. Clustering

Clustering tries to group similar observations.

Imagine points:

       • •
      • • •

                         × ×
                       × × ×

         ▲ ▲
       ▲ ▲ ▲

A clustering algorithm might discover three groups.

No teacher said:

these are class A
these are class B
these are class C

The structure came from the data.

Applications include:

customer segmentation

grouping documents by topic

biological population analysis

image organization

behavior discovery

anomaly investigation

12. K-Means Intuition

One classic clustering algorithm is k-means.

Suppose:

k = 3

We want three clusters.

Conceptually:

1. Place 3 cluster centers.

2. Assign every point to its nearest center.

3. Move each center to the average position of its assigned points.

4. Repeat.

Eventually:

points close together

tend to share a cluster.

K-means is beautifully simple.

But it implicitly prefers roughly spherical clusters and depends heavily on distance.

That means feature scaling matters enormously.


13. What Does "Similar" Mean?

This question is deeper than it looks.

Suppose two people:

Person A:
age = 25
income = $100,000

Person B:
age = 30
income = $105,000

If we measure Euclidean distance directly:

income difference dominates age difference

because dollars have much larger numerical magnitude.

So before clustering, features may need:

normalization
standardization

Machine learning algorithms do not automatically understand human units.

The representation determines what "similar" means.


14. Dimensionality Reduction

Real datasets may have thousands or millions of features.

Example image:

224 × 224 × 3

has:

150,528 pixel values

That means one image exists in a space with more than 150,000 dimensions.

Yet the meaningful structure may depend on far fewer latent factors:

object shape
orientation
lighting
texture
background

Dimensionality reduction tries to represent high-dimensional data using fewer dimensions while preserving useful structure.

10,000 dimensions
       ↓
      100
       ↓
       10

15. Why Reduce Dimensions?

Reasons include:

faster computation

less memory

visualization

noise reduction

removing redundancy

helping downstream models

discovering hidden structure

A common example is PCA — Principal Component Analysis.

PCA finds directions in the data that capture large amounts of variance.

Very roughly:

Original dimensions
x₁, x₂, x₃, x₄, ...

        ↓

New directions ordered by importance

PC1
PC2
PC3
...

Keep the most informative components.

Discard some lower-information directions.


16. Semi-Supervised Learning

Labeling data can be expensive.

Imagine:

1,000,000 medical images

but only:

10,000

were labeled by specialists.

Semi-supervised learning uses:

small labeled dataset
+
large unlabeled dataset

to learn better than using the labeled data alone.

Conceptually:

Labeled examples:
"This is pneumonia."
"This is normal."

+

Unlabeled images:
?????????

↓

Learn useful structure from everything.

This is useful whenever:

raw data is cheap

labels are expensive

which describes many real-world ML problems.


17. Self-Supervised Learning

Self-supervised learning is one of the most important ideas in modern AI.

The trick is:

Generate the supervision from the data itself.

No human needs to manually label every example.


Language example

Given:

"The robot moved toward the ___"

the model predicts:

door

The correct answer already existed in the original text.

The data supervises itself.


Modern language model example

Given:

The capital of Nepal is

predict the next token.

Training target:

Kathmandu

Again:

input and target are extracted from raw text

rather than manually annotated.


Vision example

Take an image.

Hide part of it.

Ask the model to reconstruct what was hidden.

Or generate two modified views of the same image and teach the model that their underlying representations should be similar.


18. Why Self-Supervised Learning Changed AI

Human-labeled datasets are limited.

The internet contains enormous amounts of:

text
images
video
audio
code

without manually created labels.

Self-supervised learning converts this raw data into training signals.

That enabled massive representation learning.

This is one reason modern foundation models became possible.


19. Supervised vs Semi-Supervised vs Self-Supervised

Remember:

SUPERVISED
input + human/provided target
UNSUPERVISED
data without explicit target
discover structure
SEMI-SUPERVISED
a little labeled data
+
lots of unlabeled data
SELF-SUPERVISED
targets automatically created
from the data itself

A model can also move between paradigms.

For example:

self-supervised pretraining
        ↓
supervised fine-tuning

This is extremely common today.


20. Feature Engineering

A feature is information provided to a model.

Suppose we want to predict house price.

Raw information:

construction date
bedrooms
area
latitude
longitude

Feature engineering might create:

house_age

price_per_neighborhood_average

distance_to_city_center

bedrooms_per_square_meter

The goal is to represent useful patterns in a form the model can exploit.


21. Representation Matters

Suppose we want to predict whether a date is a weekend.

Feature:

day_number = 1...365

This representation does not make weekends obvious.

Better feature:

day_of_week

Now the pattern is trivial.

This demonstrates:

Sometimes the difference between a difficult learning problem and an easy one is the representation.

Deep learning became powerful partly because neural networks can learn useful internal features automatically.

Older ML workflows often depended heavily on manual feature engineering.


22. Raw Features vs Learned Features

Traditional ML:

Human designs features
        ↓
ML algorithm
        ↓
prediction

Deep learning:

Raw-ish data
    ↓
Neural network learns representations
    ↓
prediction

Example image recognition.

Traditional:

edges
corners
texture descriptors
shape descriptors
         ↓
classifier

Deep learning:

pixels
  ↓
learned low-level features
  ↓
learned shapes
  ↓
learned object concepts
  ↓
classifier

This does not mean feature engineering disappeared.

It changed form.

Data representation, preprocessing, architecture, tokenization, context design, augmentation, and label construction remain forms of feature engineering.


23. Training Data

Training data is what the model learns from.

Example:

80,000 examples

During training, model parameters are adjusted using these examples.

A common beginner mistake is evaluating the model on the same data.

That tells you:

how well the model remembers/fits training examples

not necessarily how well it generalizes.

We therefore divide data.


24. Train / Validation / Test

The classic split:

TRAIN
VALIDATION
TEST

Each has a different purpose.


25. Training Set

Used to learn model parameters.

training data
    ↓
gradient updates
    ↓
model weights

The model is allowed to repeatedly see this data.


26. Validation Set

Used during development to make choices.

Examples:

Which model architecture?

Which regularization strength?

How many trees?

What learning rate?

What threshold?

When should training stop?

You compare alternatives using validation performance.

So:

training set
→ learn parameters

validation set
→ choose design/hyperparameters

27. Test Set

The test set answers:

After all development decisions are finished, how well does the final system perform on genuinely unseen data?

The test set should ideally remain untouched until the end.

Think of school:

Training set
=
practice exercises

Validation set
=
mock exams

Test set
=
final examination

If you repeatedly inspect the final exam while studying, it stops being a genuine final exam.

Exactly the same thing happens in ML.


28. Validation Leakage

Suppose you try:

Model A
Model B
Model C
...
Model Z

and repeatedly tune based on the test set.

Eventually, you indirectly optimize against the test set.

It has effectively become a validation set.

Your reported performance becomes overly optimistic.

Therefore:

Data can influence the model even without being directly used in gradient descent.

This broader idea is essential to understanding leakage.


29. Cross-Validation

Sometimes the dataset is too small for a single validation split.

Suppose:

1000 examples

Using 20% for validation leaves only:

800

for training.

Cross-validation lets us reuse the dataset more efficiently for evaluation.


30. K-Fold Cross-Validation

Example:

k = 5

Split data:

Fold 1
Fold 2
Fold 3
Fold 4
Fold 5

Run:

Train: 2,3,4,5
Validate: 1

Train: 1,3,4,5
Validate: 2

Train: 1,2,4,5
Validate: 3

...

Then average the validation results.

This gives a more stable estimate of model performance.


31. When Cross-Validation Can Be Wrong

Naive random folds are not appropriate for every dataset.

For time series:

future data must not predict the past

For multiple records from the same patient:

same patient should not appear
in both train and validation

For multiple samples from one user:

group-based splitting may be required

Data splitting is part of modeling.

It is not merely administrative bookkeeping.


32. Underfitting

Suppose the real relationship is curved:

       •
     •
   •
 •

but the model can only represent a crude straight line.

The model fails even on training data.

That is underfitting.

Symptoms:

training error high
validation error high

Causes may include:

model too simple

features inadequate

training insufficient

too much regularization

The model has not captured enough structure.


33. Overfitting

Now imagine a model so flexible that it perfectly memorizes every training point.

training accuracy = 100%

validation accuracy = 72%

That is overfitting.

The model learned:

real signal
+
training-specific noise

instead of the general pattern.

Think:

memorization ≠ generalization

34. The Goldilocks Problem

We want a model that is:

not too simple

not too flexible

but appropriate for the data.

Underfit          Good Fit           Overfit

too simple        captures           memorizes
                  general structure  noise

This is connected to the bias–variance trade-off.


35. Bias and Variance

These words are frequently misunderstood because ML uses them technically.


High Bias

The model has overly restrictive assumptions.

It consistently misses important patterns.

Example:

trying to fit a straight line
to a strongly curved relationship

Typical result:

underfitting

High Variance

The model reacts too strongly to details of the training dataset.

Train it on slightly different data and it may produce very different behavior.

Typical result:

overfitting

36. Bias–Variance Intuition

Imagine repeatedly training models on slightly different samples.

High-bias model:

always makes approximately
the same wrong prediction

High-variance model:

predictions change dramatically
depending on training sample

The desired model:

captures the real signal
without reacting excessively to noise

A simplified mental relationship:

Simple model
→ more bias
→ less variance

Complex model
→ less bias
→ more variance

Not an absolute law, but a very useful intuition.


37. Regularization

Regularization discourages the model from becoming unnecessarily complex.

Think:

Fit the data, but do not contort yourself to explain every tiny irregularity.

For linear models, we may minimize:

Loss
+
penalty for large parameters

Instead of:

"Make training error as small as mathematically possible."

we say:

"Fit the data well while preferring simpler solutions."

38. L2 Regularization

A common form penalizes squared parameter magnitudes.

Conceptually:

Total objective
=
prediction loss
+
λ × parameter size penalty

Large:

λ

means stronger regularization.

Effect:

weights encouraged toward smaller values

This is also called:

weight decay

in many neural-network contexts.


39. L1 Regularization

L1 penalizes absolute parameter magnitude.

It often encourages some coefficients to become exactly zero.

That can create sparse models.

Useful intuition:

L2:
"Keep weights small."

L1:
"Use fewer important weights."

40. Other Forms of Regularization

Regularization is much broader than L1/L2.

Examples:

dropout

early stopping

data augmentation

smaller models

weight decay

noise injection

label smoothing

constraints

ensembling effects

The common purpose is:

Improve generalization rather than merely optimize the training data.


41. Loss Functions

A loss function tells the training algorithm:

How wrong was this prediction?

Training attempts to minimize loss.

Example:

Prediction: $200k
True price: $250k

Loss converts that error into a numerical training signal.


42. Mean Squared Error — MSE

Common regression loss:

prediction error = ŷ - y

Square the error:

(ŷ - y)²

Average across examples.

Squaring means large errors are punished strongly.

Example:

error = 2
squared error = 4

error = 10
squared error = 100

So large mistakes dominate.


43. Mean Absolute Error — MAE

Another regression loss uses:

|ŷ - y|

Large errors are not amplified quadratically.

MAE is often more robust to outliers than MSE.

Example:

MSE:
large errors hurt dramatically

MAE:
errors grow linearly

The choice of loss encodes what kinds of mistakes matter.


44. Classification Loss

Suppose true label:

cat

Model outputs:

P(cat) = 0.99

Good.

Now:

P(cat) = 0.01

The model is confidently wrong.

A common classification loss is cross-entropy / log loss.

Its important behavior is:

correct and confident
→ low loss

wrong but uncertain
→ moderate loss

wrong and extremely confident
→ huge loss

This is exactly what we want when training probabilistic classifiers.

A model should be punished more for saying:

"I'm 99.999% sure"

and being wrong than for saying:

"I'm 55% sure"

and being wrong.


45. Loss Is Not the Same as Metric

This distinction matters.

Loss is usually used for optimization.

Metric is usually used for evaluation.

Example:

Train using:
cross-entropy loss

Evaluate using:
accuracy
precision
recall
F1
AUROC

Sometimes the loss and evaluation metric can be the same or related.

But they serve different roles.


46. Metrics

A model can look excellent under one metric and terrible under another.

There is no universal "best metric."

The correct metric depends on:

business objective

error costs

class distribution

decision threshold

deployment conditions

Metric selection is part of problem definition.


47. Accuracy

Accuracy:

correct predictions
-------------------
total predictions

Suppose:

950 correct
50 wrong

Accuracy:

95%

Simple.

But sometimes dangerously misleading.


48. The Class Imbalance Trap

Suppose:

10,000 transactions

9,900 legitimate
100 fraudulent

A useless model predicts:

LEGITIMATE

every time.

Accuracy:

99%

Looks excellent.

Fraud detection ability:

0%

This is why accuracy alone can be meaningless under class imbalance.


49. Confusion Matrix

Binary classification can be summarized:

                    TRUE

                Positive   Negative

Pred Positive      TP         FP

Pred Negative      FN         TN

Where:

TP = true positive
FP = false positive
FN = false negative
TN = true negative

Everything else becomes easier once this table is clear.


50. Precision

Precision asks:

Of everything I predicted positive, how many were actually positive?

Precision = TP / (TP + FP)

Example:

Model flags 100 emails as spam.

90 actually spam.
10 legitimate.

Precision:

90%

High precision means:

When I say positive,
I am usually right.

51. Recall

Recall asks:

Of all actual positives, how many did I successfully find?

Recall = TP / (TP + FN)

Suppose:

100 fraudulent transactions exist

model catches 80
misses 20

Recall:

80%

High recall means:

I miss few true positives.

52. Precision vs Recall

This trade-off appears constantly.

Spam filter:

High precision:
avoid marking legitimate email as spam.

High recall:
catch almost all spam.

Cancer screening:

High recall may be especially important
because missing a true case is costly.

Precision and recall should not be interpreted abstractly.

Ask:

What does a false positive cost?

and:

What does a false negative cost?


53. F1 Score

F1 combines precision and recall using their harmonic mean.

Useful when both matter and class imbalance exists.

Intuition:

Precision high
Recall low
→ F1 penalized

Precision low
Recall high
→ F1 penalized

Both high
→ F1 high

But F1 does not automatically represent every real-world cost structure.

Use it intentionally.


54. ROC and AUROC

A classifier often produces scores.

By moving the decision threshold, you get different:

true-positive rates
false-positive rates

The ROC curve summarizes this behavior.

AUROC asks roughly:

How well does the model rank positive examples above negative examples across thresholds?

Useful, but it can sometimes look optimistic with severe class imbalance.

Precision-recall curves may then be more informative.


55. Regression Metrics

Common regression metrics include:

MAE

MSE

RMSE

R²

Each emphasizes errors differently.


MAE

Average absolute error.

Easy interpretation.

"The prediction is wrong by about $8,000 on average."

RMSE

Square root of average squared error.

It punishes large mistakes more strongly.

Useful when big misses matter disproportionately.


Roughly indicates how much variance in the target the model explains relative to a baseline.

But it should not be treated as a universal measure of model quality.

Again:

Metric interpretation depends on the problem.


56. Calibration

This is one of the most important concepts that many ML engineers initially overlook.

Suppose a model makes 100 predictions with:

P(success) ≈ 0.8

A well-calibrated model should be correct roughly:

80% of the time

among those predictions.

Calibration asks:

Do predicted probabilities correspond to real-world frequencies?


57. Confidence vs Correctness

Imagine Model A says:

P(cat) = 0.99

for 100 images.

If only:

70

are cats, the model is overconfident.

Model B says:

P(cat) = 0.70

and around 70 out of 100 are cats.

Model B is better calibrated.

This matters enormously for:

medicine

robotics

risk systems

fraud

autonomous driving

decision-making agents

because downstream systems may rely on the probability itself.


58. Accuracy and Calibration Are Different

A model can be:

accurate but poorly calibrated

or:

less accurate but well calibrated

For example, two classifiers could make identical label predictions but output:

Model A:
0.51, 0.52, 0.53...

Model B:
0.99, 0.999, 0.9999...

They have the same accuracy.

But their confidence behavior is very different.

Therefore:

classification quality
≠
probability quality

59. Class Imbalance

Class imbalance occurs when some classes are much more common than others.

Example:

Normal transactions: 99.9%
Fraud:               0.1%

Problems include:

accuracy becomes misleading

minority class may be ignored

training gradients dominated by majority examples

decision threshold may become inappropriate

Possible techniques include:

class weighting

oversampling

undersampling

focal loss

threshold adjustment

better metrics

collecting more minority examples

But blindly "balancing" every dataset is not automatically correct.

Deployment prevalence matters.


60. Data Leakage

Data leakage is one of the easiest ways to accidentally create a fake "great" model.

Leakage occurs when training receives information that would not legitimately be available when making real predictions.

Example:

We want to predict:

Will this patient be diagnosed with disease?

Feature:

medication_given_after_diagnosis

That feature reveals the future answer.

The model may score:

99.9%

but it learned a shortcut unavailable in real deployment.


61. Simple Leakage Example

Predict whether a student passed an exam.

Features:

study hours
attendance
homework completion
final_certificate_status

But final_certificate_status is generated after the exam.

The model sees:

future information

and appears brilliant.

Deployment:

feature does not exist yet.

Failure.


62. Leakage Through Preprocessing

Leakage can be subtle.

Suppose you normalize features using:

mean of entire dataset

before splitting train and test.

Then test-set information influenced training preprocessing.

Better:

split first

fit preprocessing using training set only

apply learned transformation to validation/test

Same principle for:

feature selection

imputation

PCA

target encoding

normalization

data augmentation decisions

The test set must behave like unseen future information.


63. Leakage Through Duplicate or Related Samples

Suppose:

same patient

has 20 scans.

Random splitting may put:

15 scans in training
5 scans in validation

The model can partly recognize patient-specific patterns.

Validation performance becomes artificially high.

Similarly:

same speaker

same document

same video

same customer

same machine

same location

may require grouped splitting.

A correct split should simulate the real deployment boundary.


64. Distribution Shift

Machine learning assumes training data tells us something useful about future data.

But the world changes.

That creates distribution shift.

Formally, training examples come from something like:

P_train(X, Y)

Deployment examples come from:

P_test(X, Y)

If these differ substantially:

model performance can collapse

even if training and validation looked excellent.


65. Distribution Shift Example

Imagine training an autonomous-driving vision model mostly on:

sunny California roads

Then deploying on:

snowy Nepal mountain roads

Differences:

lighting

weather

road geometry

vehicles

signs

pedestrians

camera conditions

The learned data distribution changed.

This is not necessarily "overfitting" in the ordinary sense.

The environment itself moved.


66. Types of Shift

Useful categories include:

Covariate shift

P(X) changes

Input distribution changes.

Example:

new camera sensor

Label shift

P(Y) changes

Class prevalence changes.

Example:

fraud becomes much more common during an attack

Concept drift

The relationship itself changes.

P(Y | X) changes

Example:

Spam attackers discover how your filter works.

Old patterns stop being reliable.

This is common in adversarial or evolving environments.


67. Why Production ML Is Harder Than Kaggle ML

A benchmark dataset sits still.

The real world does not.

Production systems face:

distribution shift

broken sensors

new user behavior

schema changes

missing features

feedback loops

adversarial behavior

seasonality

software changes

policy changes

So production ML is partly:

modeling

and partly:

monitoring the assumptions under which the model was valid.

68. Uncertainty in Machine Learning

A prediction should not always be treated as certain.

Suppose a model says:

cat

We should ask:

How confident is it?

Has it seen similar data before?

Are multiple plausible predictions possible?

Is the model itself uncertain?

Uncertainty becomes especially important in autonomous systems.


69. Aleatoric Uncertainty

Aleatoric uncertainty comes from inherent randomness/noise in the observations or world.

Example:

extremely blurry image

Even the perfect model may not know what object is present.

Or:

future traffic arrival

contains intrinsic unpredictability.

Think:

Data itself is ambiguous.

70. Epistemic Uncertainty

Epistemic uncertainty comes from limited model knowledge.

Example:

A robot vision model trained only on:

cars
dogs
people

sees:

an elephant

The problem is not merely image noise.

The model lacks knowledge about this region of the world.

Think:

The model doesn't know enough.

With better or more diverse training data, epistemic uncertainty can potentially decrease.


71. Aleatoric vs Epistemic

Memorize the intuition:

ALEATORIC

"The world/data is uncertain."
EPISTEMIC

"My model's knowledge is uncertain."

Example:

Fog obscures pedestrian:

aleatoric

Never trained on snow-covered roads:

epistemic

Both matter enormously in robotics and safety-critical AI.


72. Out-of-Distribution Data

Suppose a model trained on handwritten digits:

0...9

receives a photo of:

a giraffe

A bad model might confidently say:

"7 — 99.3%"

Softmax does not inherently mean:

"I recognize this."

It may simply mean:

"Among the classes I know, 7 received the largest score."

This is why:

confidence

must not automatically be interpreted as:

knowledge.

Out-of-distribution detection remains an important ML problem.


73. Interpretability

Interpretability asks:

Can humans understand why the model made a prediction?

Consider a loan model:

Rejected.

A user might reasonably ask:

Why?

Potential explanations:

income too low

debt ratio too high

short credit history

For some systems, explanation is merely helpful.

For others, it is operationally, scientifically, ethically, or legally important.


74. Intrinsically Interpretable Models

Some models are naturally easier to inspect.

Examples:

linear regression

small decision trees

simple rule systems

Linear model:

price
=
100 × area
- 3000 × age
+ ...

You can inspect coefficients.

Decision tree:

IF income > X
    IF debt < Y
       approve

Human-readable structure.


75. Post-Hoc Interpretability

Complex models may need tools that approximate why they behaved a certain way.

Examples include:

feature importance

SHAP

LIME

saliency maps

counterfactual explanations

But be careful.

An explanation method is itself another model or approximation.

It may not perfectly reveal the true internal reasoning process.

Interpretability should not be confused with storytelling.


76. Feature Importance

Suppose a model predicts house price.

Feature importance might indicate:

Location      45%
Area          30%
Age           15%
Bedrooms      10%

This can help understand global model behavior.

But importance does not automatically imply:

causation

For example:

ice cream sales

may correlate with:

sunburn

because both increase in hot weather.

Ice cream does not cause sunburn.

Machine learning usually learns statistical relationships, not necessarily causal relationships.


77. Correlation vs Causation

This deserves permanent attention.

A model may learn:

X predicts Y

That does not mean:

changing X causes Y to change

Prediction:

Umbrellas → rain

Umbrellas strongly predict rain.

But deploying more umbrellas does not cause rainfall.

Predictive ML and causal reasoning are different problems.

This becomes crucial when moving from:

prediction

to:

intervention and decision making.

78. The Full Training Pipeline

A robust ML workflow often resembles:

Define problem
     ↓
Collect data
     ↓
Understand data-generation process
     ↓
Define train/validation/test split
     ↓
Clean/preprocess
     ↓
Engineer/learn features
     ↓
Choose baseline
     ↓
Train
     ↓
Evaluate validation performance
     ↓
Analyze errors
     ↓
Tune / iterate
     ↓
Lock decisions
     ↓
Evaluate test set
     ↓
Deploy
     ↓
Monitor real-world performance
     ↓
Detect drift / failures
     ↓
Retrain or redesign

Notice:

training the model

is only one stage.


79. Start With a Baseline

A powerful engineering habit:

Before building a sophisticated model, build a simple baseline.

Examples:

Classification:

always predict majority class

Regression:

always predict training mean

Then:

logistic regression

small tree

simple heuristic

Why?

Suppose your giant neural network gets:

82%

but logistic regression gets:

81.8%

Your complexity may not be justified.

Baselines prevent self-deception.


80. Error Analysis

Suppose accuracy is:

88%

That number alone tells you very little.

Look at failures.

Maybe:

40% are low-light images

25% are partially occluded

20% are mislabeled

10% are unusual object classes

5% are random

Now engineering becomes actionable.

Perhaps the solution is not:

bigger model

but:

better low-light data

Error analysis often creates more improvement than blind hyperparameter tuning.


81. Data Quality vs Model Quality

Beginners often assume:

better algorithm = better system

In practice:

better data

can matter more.

Problems include:

incorrect labels

missing values

duplicates

sampling bias

unrepresentative users

bad sensors

ambiguous targets

leakage

inconsistent preprocessing

A powerful model trained on bad data becomes a powerful reproducer of bad patterns.


82. Garbage In, Garbage Out — But More Precisely

The common phrase:

Garbage in → garbage out

is useful but incomplete.

Modern models can sometimes tolerate noisy data.

The deeper problem is:

The model learns the statistical structure that the dataset actually contains, not the structure you intended it to contain.

If your dataset contains a shortcut:

all wolf photos contain snow

the model may learn:

snow → wolf

instead of:

animal appearance → wolf

The model did not "cheat."

It optimized the task you accidentally gave it.


83. Spurious Correlations

A spurious correlation is a relationship that predicts well in training data but does not represent the intended underlying relationship.

Example:

Hospital A uses one scanner type mostly for sick patients.

Hospital B uses another scanner mostly for healthy patients.

The model may learn:

scanner artifacts → disease

rather than:

medical pathology → disease

Validation may look excellent if the same hospitals appear in both sets.

Deployment at Hospital C:

performance collapses.

This connects:

data leakage

distribution shift

shortcut learning

generalization

84. Hyperparameters vs Parameters

Do not mix these.

Parameters

Learned by the model.

Examples:

neural-network weights

linear-regression coefficients

Hyperparameters

Chosen by the engineer/training procedure.

Examples:

learning rate

tree depth

regularization strength

batch size

number of layers

k in k-means

Validation data helps choose hyperparameters.

Training data learns parameters.


85. Model Capacity

Capacity means roughly:

How complex a function can this model represent?

Low-capacity model:

simple straight line

High-capacity model:

huge neural network

Higher capacity can capture richer relationships.

But it can also capture:

noise

shortcuts

memorization

Therefore:

capacity
+
data amount
+
regularization
+
optimization

must work together.


86. More Data Can Act Like Regularization

Suppose a giant model sees only:

100 examples.

Memorization is easy.

If it sees:

100 million diverse examples

then memorizing useless quirks becomes much less helpful.

More representative data can greatly improve generalization.

This is one reason large-scale ML has become so successful.

But:

more biased data

does not necessarily solve bias.

Quantity does not automatically fix data quality.


87. Training vs Inference

Another essential distinction.

Training

data
 ↓
compute gradients / optimize
 ↓
update model parameters

Potentially expensive.

Inference

new input
 ↓
fixed trained model
 ↓
prediction

Example:

Training:
learn cat detector using 10 million images.

Inference:
classify one new photo.

For deployed systems, inference latency, memory, throughput, and cost matter enormously.


88. Offline Metrics vs Online Reality

A model may have:

95% offline accuracy

and still be a bad product.

Why?

Maybe:

too slow

too expensive

poorly calibrated

fails on important minority users

performance degrades over time

wrong errors are catastrophic

requires unavailable features

causes bad human behavior

Therefore ML evaluation has layers:

model metric
    ↓
system metric
    ↓
product/business outcome
    ↓
real-world impact

Optimizing only the first can mislead you.


89. Machine Learning Is Statistical, Not Magical

A model learns from examples.

If the data does not contain enough information to predict the target, no algorithm can magically recover it.

Suppose:

Input:
random ID number

Target:
tomorrow's lottery result

No meaningful predictive relationship exists.

A sophisticated model may still find accidental patterns in training data.

That is overfitting, not intelligence.

Always ask:

Is the information needed for prediction actually present in the inputs?


90. The Three Fundamental Sources of Failure

Most ML failures can be understood through three questions.

1. Data problem?

wrong labels
bad sampling
shift
leakage
missing information

2. Representation/model problem?

features inadequate
capacity too low/high
wrong inductive bias

3. Evaluation/objective problem?

wrong loss
wrong metric
bad threshold
bad split
wrong definition of success

Do not automatically assume every problem is solved by:

train a larger model

91. A Complete Example — Fraud Detection

Let's connect everything.

Goal:

Detect fraudulent card transactions.

Data

Each transaction has:

amount

merchant

location

time

device

historical behavior

Target:

fraud
legitimate

Therefore:

supervised classification

Feature Engineering

Create:

distance from usual location

transactions in last 10 minutes

amount relative to personal average

new device indicator

Split

Do not randomly mix all historical transactions.

A more realistic split might be:

train = older months
validation = later month
test = newest month

because deployment predicts future transactions.


Class Imbalance

Maybe:

fraud = 0.2%

Accuracy is nearly useless.

Use:

precision

recall

PR-AUC

cost-based metrics

Loss

Perhaps use:

weighted cross-entropy

because fraud examples matter strongly.


Calibration

Suppose:

P(fraud)=0.90

should actually mean something operational.

Risk engines may depend on that probability.


Threshold

Maybe:

P(fraud) > 0.95
→ block

0.70–0.95
→ request verification

<0.70
→ allow

Same model.

Different actions.


Distribution Shift

Fraudsters change tactics.

P(Y|X)

changes.

Performance degrades.

Monitoring detects drift.

Retraining becomes necessary.


Interpretability

Analyst asks:

Why was transaction blocked?

System reports:

new country

unusual amount

new device

multiple recent attempts

Now the complete ML system makes sense.


92. Another Complete Example — Robot Object Detection

Input:

camera frame

Output:

pedestrian
vehicle
box
unknown

This is supervised classification/detection.

Training data:

images + bounding-box labels

Self-supervised pretraining may first learn visual representations.

Augmentation:

brightness

rotation

crop

blur

acts partly as regularization.

Train/validation/test must represent realistic environments.

Metrics:

precision

recall

mAP

Class imbalance:

many cars
few wheelchairs

needs careful attention.

Calibration matters because downstream planning may use:

P(pedestrian)

Distribution shift appears when:

weather changes

camera changes

country changes

night arrives

Uncertainty matters because:

"unknown object"

may be safer than confidently misclassifying something unseen.

And interpretability/error analysis helps identify:

fails mostly under occlusion

This connects ML directly to autonomous systems.


93. The Deep Relationship Between the Learning Paradigms

Think of supervision as a spectrum.

LOTS OF EXPLICIT HUMAN TARGETS

        supervised
            ↓
      semi-supervised
            ↓
      self-supervised
            ↓
       unsupervised

LITTLE / NO EXPLICIT HUMAN TARGETS

Modern systems often combine them:

Raw internet data
        ↓
self-supervised pretraining
        ↓
small human-labeled dataset
        ↓
supervised fine-tuning
        ↓
deployment feedback
        ↓
further adaptation

These categories are not mutually exclusive stages of a system.


94. The Deep Relationship Between Fit, Bias, Variance, and Regularization

Here is the mental map:

MODEL TOO SIMPLE
      ↓
high bias
      ↓
underfitting

MODEL TOO FLEXIBLE
      ↓
high variance
      ↓
overfitting

Regularization pushes against excessive flexibility.

More representative data often reduces variance.

Better features can reduce bias.

Better model architecture may reduce both.

So when debugging:

training bad
validation bad
→ likely underfitting / data problem

training excellent
validation poor
→ likely overfitting / distribution mismatch

This is not a perfect diagnostic law, but it is an excellent starting point.


95. The Deep Relationship Between Loss and Metrics

Think:

LOSS
"What signal should training optimize?"
METRIC
"How should humans judge performance?"
BUSINESS / SYSTEM COST
"What mistakes actually matter in reality?"

Ideally these align.

But often they do not perfectly.

Example:

Train:

cross-entropy

Validate:

PR-AUC

Deploy:

cost of fraud
+
cost of falsely blocking customers

The final system objective may be more complex than the training loss.


96. The Deep Relationship Between Calibration and Decision Making

Suppose a model outputs:

P(failure)=0.3

If that probability is calibrated, a decision system can combine it with cost:

Expected cost
=
P(failure) × failure_cost

Poor calibration corrupts this calculation.

Therefore:

Probability estimation
      ↓
Calibration
      ↓
Decision theory
      ↓
Action

This directly connects Machine Learning to the AI Foundations chapter.


97. ML vs Classical AI

Classical AI often asks:

How should knowledge and rules be represented?

Machine learning asks:

Can useful behavior be learned from data?

Example:

Classical rule:

IF object has wheels
AND is large
THEN maybe vehicle

Machine learning:

millions of images
        ↓
learn representation
        ↓
vehicle probability

Modern systems combine them.

For a robot:

ML perception
      ↓
probabilistic state estimation
      ↓
planning
      ↓
decision making
      ↓
control

ML often provides predictions.

It does not automatically solve the complete intelligence problem.


98. Prediction vs Decision — Again

This distinction is critical enough to repeat.

Machine learning model:

P(pedestrian)=0.37

Decision system:

Should I brake?

You cannot answer from probability alone.

You need:

consequences
risk
utility
safety constraints

Thus:

ML
often answers:
"What is likely?"

AI agent
must answer:
"What should I do?"

99. Machine Learning vs Deep Learning

Machine Learning is the larger field.

Machine Learning
│
├── linear models
├── decision trees
├── random forests
├── SVMs
├── nearest neighbors
├── probabilistic models
├── clustering
└── neural networks
      ↓
   Deep Learning

Deep learning is machine learning using multilayer neural networks that learn powerful representations.

So:

Deep Learning ⊂ Machine Learning

not the other way around.


100. Machine Learning vs Reinforcement Learning

Supervised learning:

input
+
correct answer
→ learn mapping

Reinforcement learning:

state
→ action
→ reward
→ next state

The agent learns from consequences rather than a correct label for every decision.

RL connects directly to:

MDPs

policies

value functions

Q-values

Machine learning therefore extends naturally from prediction into sequential decision making.


101. A Senior Engineer's View of an ML System

Do not think:

Model = system

Think:

                DATA PIPELINE
                     │
                     ▼
              FEATURE PIPELINE
                     │
                     ▼
                 MODEL
                     │
                     ▼
               CALIBRATION
                     │
                     ▼
            DECISION LOGIC
                     │
                     ▼
              APPLICATION
                     │
                     ▼
                FEEDBACK
                     │
                     ▼
                MONITORING
                     │
                     ▼
                RETRAINING

A production ML engineer must understand the entire loop.


102. The Questions You Should Ask Before Training Anything

Before choosing an algorithm, ask:

What exactly is the target?

How was the data generated?

What information exists at prediction time?

What does one sample represent?

What is the deployment distribution?

What mistakes matter most?

What metric reflects those mistakes?

How will train/test separation simulate deployment?

Can labels be trusted?

What happens when the model is uncertain?

How will drift be detected?

What is the baseline?

These questions often matter more than the model architecture.


103. Permanent Cheat Sheet

Machine Learning

Learn useful patterns/functions from data
that generalize to unseen examples.

Supervised Learning

inputs + targets
→ learn mapping

Unsupervised Learning

unlabeled data
→ discover structure

Semi-Supervised Learning

small labeled dataset
+
large unlabeled dataset

Self-Supervised Learning

create training targets from raw data itself

Regression

predict a number

Examples:

price
temperature
speed

Classification

predict a class / class probability

Examples:

cat/dog

spam/not spam

fraud/not fraud

Clustering

group similar examples
without known labels

Dimensionality Reduction

many dimensions
→ fewer informative dimensions

Feature Engineering

Represent raw information
in a form that exposes useful patterns.

Train Set

learn parameters

Validation Set

make development/hyperparameter choices

Test Set

final unbiased evaluation

Cross-Validation

rotate validation folds
to estimate performance more reliably

Underfitting

model too simple / insufficiently learned

train bad
validation bad

Overfitting

memorizes training-specific structure

train great
validation poor

Bias

systematic error from overly restrictive assumptions

Variance

excessive sensitivity to the particular training sample

Regularization

Discourage unnecessary complexity
to improve generalization.

Loss

How wrong is the prediction?

Used to train.

Metric

How should performance be judged?

Used to evaluate.

Accuracy

correct / total

Dangerous under strong class imbalance.


Precision

Of predicted positives,
how many were truly positive?
TP / (TP + FP)

Recall

Of actual positives,
how many did we find?
TP / (TP + FN)

Calibration

When the model says 80%,
does the event really occur about 80% of the time?

Class Imbalance

some classes are much rarer than others

Therefore accuracy may mislead.


Data Leakage

training indirectly receives information
that should only exist in the future/test/deployment world

Leakage creates fake performance.


Distribution Shift

training world
≠
deployment world

A good historical model can become a bad current model.


Aleatoric Uncertainty

uncertainty inherent in data/world

Epistemic Uncertainty

uncertainty caused by limited model knowledge

Interpretability

Can we understand
why the model behaved as it did?

104. The Entire Field in One Diagram

                         DATA
                           │
                           ▼
                    REPRESENTATION
                           │
                  feature engineering
                  learned features
                           │
                           ▼
                     ML PARADIGM
                           │
       ┌───────────────────┼──────────────────┐
       │                   │                  │
 supervised          unsupervised       self-supervised
       │                   │                  │
       ▼                   ▼                  ▼
 regression           clustering        representation
 classification       dimensionality      learning
                         reduction
       │
       ▼
                    TRAIN MODEL
                           │
                      minimize loss
                           │
                           ▼
                       VALIDATE
                           │
                 tune hyperparameters
                           │
                           ▼
                       TEST
                           │
                 estimate generalization
                           │
                           ▼
                       DEPLOY
                           │
                           ▼
                 NEW REAL-WORLD DATA
                           │
              ┌────────────┴─────────────┐
              │                          │
       predictions                  uncertainty
              │                          │
              ▼                          ▼
          decisions                 confidence /
                                    calibration
              │
              ▼
         real outcomes
              │
              ▼
          MONITORING
              │
      distribution shift?
      class drift?
      performance drop?
              │
              ▼
          RETRAIN / UPDATE

105. The Mental Ladder

If you forget everything else, reconstruct Machine Learning using these questions:

1. WHAT AM I TRYING TO PREDICT?

   number?
   class?
   structure?
2. WHAT DATA DO I HAVE?

   labeled?
   unlabeled?
   partially labeled?
3. WHAT INFORMATION SHOULD THE MODEL SEE?

   features
   representation
4. HOW DO I SEPARATE LEARNING FROM EVALUATION?

   train
   validation
   test
5. HOW DO I DEFINE "WRONG"?

   loss
6. HOW DO I DEFINE "GOOD"?

   metrics
7. IS THE MODEL LEARNING OR MEMORIZING?

   bias
   variance
   underfitting
   overfitting
8. HOW DO I IMPROVE GENERALIZATION?

   data
   features
   regularization
   model choice
9. CAN I TRUST ITS PROBABILITIES?

   calibration
   uncertainty
10. ARE IMPORTANT CASES RARE?

    class imbalance
11. DID INFORMATION LEAK?

    leakage
12. WILL THE FUTURE LOOK LIKE THE TRAINING DATA?

    distribution shift
13. CAN I UNDERSTAND ITS FAILURES?

    interpretability
    error analysis
14. DOES IT STILL WORK AFTER DEPLOYMENT?

    monitoring
    drift detection
    retraining

Final Mental Picture

Machine Learning is not fundamentally:

"find the fanciest algorithm."

It is:

Observe examples
      ↓
Represent them correctly
      ↓
Define what should be learned
      ↓
Choose an appropriate model
      ↓
Define a loss
      ↓
Learn from training data
      ↓
Measure generalization
      ↓
Understand the errors
      ↓
Estimate uncertainty
      ↓
Deploy into a changing world
      ↓
Monitor whether assumptions still hold

The central engineering challenge is generalization.

You have finite historical observations:

training data

but you care about:

future unseen reality.

Everything else—

validation
regularization
bias/variance
cross-validation
calibration
leakage prevention
distribution-shift monitoring

—exists because those two things are not automatically the same.

So if there is one sentence worth retaining permanently, make it this:

Machine Learning is the engineering of systems that extract useful statistical structure from past data and continue making reliable predictions on future, unseen data.

And for a senior engineer, add one more sentence:

A model is only trustworthy to the extent that its data, objective, evaluation procedure, uncertainty estimates, and deployment assumptions are trustworthy.

Once those two ideas are solid, most of Machine Learning has a natural place in your head.