One of the most confusing topics for beginners in Deep Learning is deciding which output layer, activation function, loss function, and dataset label mode should be used for a particular machine learning problem.

Fortunately, there is a simple way to think about it.

Rule #1: Don’t start by thinking about the model. Start by asking:

“What kind of prediction am I trying to make?”

Once you answer that question, the rest becomes almost automatic.

---

1. Binary Classification

Definition

Binary classification means there are only two possible classes.

Examples:

• Cat vs Dog
• Horse vs Human
• Happy vs Sad
• Spam vs Not Spam
• Cancer vs Healthy

Only one answer can be correct.

Image
   │
   ▼
Cat   OR   Dog

Model Architecture

Since there are only two possible outcomes, the network only needs one output neuron.

model = tf.keras.Sequential([
    ...
    tf.keras.layers.Dense(1, activation="sigmoid")
])

Why Sigmoid?

Sigmoid converts the output into a probability between 0 and 1.

Example:

0.02  → Cat

0.95  → Dog

Compile

model.compile(
    optimizer="adam",
    loss="binary_crossentropy",
    metrics=["accuracy"]
)

Dataset Labels

0
1

or

False
True

When using image_dataset_from_directory():

label_mode="binary"

---

Memory Trick

2 Classes

↓

1 Output Neuron

↓

Sigmoid

↓

Binary Crossentropy

---

2. Multi-Class Classification (One Correct Answer)

Definition

There are more than two classes, but only one class is correct.

Examples:

• MNIST Digits (10 classes)
• Fashion MNIST (10 classes)
• Flowers (5 classes)
• Fruits (24 classes)

Example:

Image

↓

Apple
Banana
Orange
Grape

Only ONE is correct.

Model Architecture

If there are 10 classes:

tf.keras.layers.Dense(10, activation="softmax")

If there are 24 classes:

tf.keras.layers.Dense(24, activation="softmax")

Why Softmax?

Softmax converts outputs into probabilities whose total equals 1.0.

Example:

Dog      0.03

Cat      0.02

Horse    0.91

Cow      0.04

The model predicts Horse because it has the highest probability.

---

Case A — One-Hot Encoded Labels

Labels look like this:

Cat

[1 0 0]

Dog

[0 1 0]

Horse

[0 0 1]

Compile using:

model.compile(
    optimizer="adam",
    loss="categorical_crossentropy",
    metrics=["accuracy"]
)

Dataset:

label_mode="categorical"

---

Case B — Integer Labels

Sometimes labels are stored simply as numbers.

Example:

0

2

5

9

instead of

[1 0 0 0 0 ...]

[0 0 1 0 0 ...]

...

The output layer is exactly the same:

Dense(10, activation="softmax")

The only thing that changes is the loss function.

loss="sparse_categorical_crossentropy"

Dataset:

label_mode="int"

---

Memory Trick

One Correct Class

↓

Softmax

↓

One-Hot Labels
↓

Categorical Crossentropy

OR

Integer Labels
↓

Sparse Categorical Crossentropy

---

3. Multi-Label Classification

Definition

An image can belong to multiple classes at the same time.

Example:

Image

✓ Dog

✓ Car

✓ Tree

✗ Cat

✗ Person

Unlike multi-class classification, multiple answers can all be correct.

---

Model

If there are five possible labels:

Dense(5, activation="sigmoid")

Notice that we use Sigmoid, not Softmax.

Each neuron independently answers:

Yes or No?

---

Compile

model.compile(
    optimizer="adam",
    loss="binary_crossentropy",
    metrics=["accuracy"]
)

---

Memory Trick

Many Answers

↓

Sigmoid

↓

Binary Crossentropy

---

4. Regression

Regression predicts a continuous numerical value, not a class.

Examples:

• House price
• Temperature
• Salary
• Stock price
• Age estimation

Example output:

$350,000

---

Model

Dense(1)

Notice that there is no activation function.

---

Compile

model.compile(
    optimizer="adam",
    loss="mse",
    metrics=["mae"]
)

Other common loss functions include:

loss="mae"

or

loss="huber"

---

Memory Trick

Predict Number

↓

No Activation

↓

MSE / MAE

---

Understanding the Relationship

The easiest way to memorize TensorFlow is to start with the labels.

Binary Labels

0

1

↓

Dense(1)

Sigmoid

Binary Crossentropy

---

Integer Labels

0

3

9

15

↓

Dense(N)

Softmax

Sparse Crossentropy

---

One-Hot Labels

[0 0 1 0]

[1 0 0 0]

[0 1 0 0]

↓

Dense(N)

Softmax

Categorical Crossentropy

---

Multi-Label

[1 0 1 0 1]

↓

Dense(N)

Sigmoid

Binary Crossentropy

---

Continuous Numbers

23.4

125000

19.8

↓

Dense(1)

No Activation

MSE / MAE

---

Complete Cheat Sheet

Problem	Example	Last Layer	Activation	Loss Function	Label Mode
Binary Classification	Cat vs Dog	Dense(1)	Sigmoid	Binary Crossentropy	binary
Multi-Class (One-Hot)	Fashion MNIST	Dense(N)	Softmax	Categorical Crossentropy	categorical
Multi-Class (Integer)	MNIST	Dense(N)	Softmax	Sparse Categorical Crossentropy	int
Multi-Label	Dog + Car + Tree	Dense(N)	Sigmoid	Binary Crossentropy	Multi-hot
Regression	House Price	Dense(1)	None	MSE / MAE	Numeric

---

A Simple Way to Remember Everything

Whenever you’re building a neural network, ask these questions in order:

Question 1

Am I predicting a class or a number?

• Number → Regression
• Class → Classification

Question 2

How many correct answers can exist?

• One → Softmax
• Multiple → Sigmoid

Question 3

How are my labels stored?

• Integers → Sparse Categorical Crossentropy
• One-hot vectors → Categorical Crossentropy
• Binary labels → Binary Crossentropy

---

Final Mnemonic

2 Classes
→ 1 Neuron
→ Sigmoid
→ Binary Crossentropy

One Correct Class
→ N Neurons
→ Softmax
→ Categorical (or Sparse Categorical) Crossentropy

Many Correct Classes
→ N Neurons
→ Sigmoid
→ Binary Crossentropy

Predict a Number
→ 1 Neuron
→ No Activation
→ MSE / MAE

If you memorize these four patterns, you’ll be able to build the correct output layer and choose the appropriate loss function for almost every TensorFlow/Keras model you encounter.