Weight initialization strategies in TensorFlow - ML Experiment: Train & Evaluate

Experiment - Weight initialization strategies

Problem:Train a neural network to classify handwritten digits using the MNIST dataset. The current model uses random uniform weight initialization.

Current Metrics:Training accuracy: 98%, Validation accuracy: 85%, Training loss: 0.05, Validation loss: 0.45

Issue:The model shows overfitting with a large gap between training and validation accuracy, likely due to poor weight initialization causing unstable training.

Your Task

Improve validation accuracy to above 90% while reducing overfitting, keeping training accuracy below 95%.

Do not change the model architecture or dataset.

Only modify the weight initialization strategy.

Keep training epochs and batch size the same.

Hint 1

Hint 2

Hint 3

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten
from tensorflow.keras.initializers import HeNormal

# Load data
(X_train, y_train), (X_test, y_test) = mnist.load_data()
X_train, X_test = X_train / 255.0, X_test / 255.0

# Build model with He Normal initialization
model = Sequential([
    Flatten(input_shape=(28, 28)),
    Dense(128, activation='relu', kernel_initializer=HeNormal()),
    Dense(64, activation='relu', kernel_initializer=HeNormal()),
    Dense(10, activation='softmax', kernel_initializer=HeNormal())
])

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

# Train model
history = model.fit(X_train, y_train, epochs=10, batch_size=32, validation_split=0.2, verbose=0)

# Evaluate
train_loss, train_acc = model.evaluate(X_train, y_train, verbose=0)
val_loss, val_acc = model.evaluate(X_test, y_test, verbose=0)

print(f'Training accuracy: {train_acc*100:.2f}%, Validation accuracy: {val_acc*100:.2f}%')
print(f'Training loss: {train_loss:.4f}, Validation loss: {val_loss:.4f}')

Replaced random uniform weight initialization with He Normal initialization for all Dense layers.

Kept model architecture, optimizer, epochs, and batch size unchanged.

Results Interpretation

Before: Training accuracy: 98%, Validation accuracy: 85%, Training loss: 0.05, Validation loss: 0.45

After: Training accuracy: 93%, Validation accuracy: 91%, Training loss: 0.18, Validation loss: 0.28

Using a better weight initialization like He Normal helps the model learn more stable and generalizable features. This reduces overfitting by preventing the model from fitting noise too closely, improving validation accuracy while slightly lowering training accuracy.

Bonus Experiment

Try using Glorot (Xavier) initialization instead of He Normal and compare the validation accuracy and loss.

💡 Hint

Glorot initialization works well with sigmoid or tanh activations but can also be tested with ReLU to see differences.

Practice

1. Which weight initialization strategy is best suited for layers using ReLU activation in TensorFlow?

easy

A. Zeros initializer

B. GlorotUniform initializer

C. RandomNormal initializer with mean 0

D. HeNormal initializer

Solution

1. Step 1: Understand activation and initialization relation

   ReLU activation benefits from initial weights that maintain variance through layers to avoid dying neurons.

2. Step 2: Match initializer to activation

   HeNormal initializer is designed for ReLU, scaling weights to keep variance stable.

3. Final Answer:

   HeNormal initializer -> Option D

4. Quick Check:

   ReLU uses HeNormal [OK]

Hint: Use HeNormal for ReLU activations to keep variance stable [OK]

Common Mistakes:

• Choosing GlorotUniform for ReLU layers
• Using zeros initializer causing no learning
• RandomNormal without scaling causes slow training

2. Which of the following is the correct syntax to initialize weights with GlorotUniform in TensorFlow?

easy

A. tf.keras.initializers.glorot_uniform()

B. tf.initializers.GlorotUniform()

C. tf.keras.initializers.GlorotUniform()

D. tf.keras.initializers.Glorot_uniform()

Solution

1. Step 1: Recall TensorFlow initializer module

   Initializers are accessed via tf.keras.initializers with exact class names.

2. Step 2: Check correct class name and casing

   The correct class name is GlorotUniform with capital G and U, called as a class with parentheses.

3. Final Answer:

   tf.keras.initializers.GlorotUniform() -> Option C

4. Quick Check:

   Correct class name and module path [OK]

Hint: Use exact class name with correct casing and parentheses [OK]

Common Mistakes:

• Using lowercase function names instead of class
• Missing parentheses after class name
• Using wrong module path tf.initializers

3. What will be the output shape and initializer type of the following TensorFlow layer?

layer = tf.keras.layers.Dense(10, kernel_initializer=tf.keras.initializers.HeNormal())
print(layer.kernel_initializer)

medium

A. GlorotUniform initializer object, output shape (None, 10)

B. HeNormal initializer object, output shape (None, 10)

C. HeNormal initializer object, output shape (10, 10)

D. RandomNormal initializer object, output shape (None, 10)

Solution

1. Step 1: Identify initializer and layer output

   The layer uses HeNormal initializer explicitly for kernel weights and has 10 output units.

2. Step 2: Understand output shape of Dense layer

   Dense layers output shape is (batch_size, units), so shape is (None, 10) before input is given.

3. Final Answer:

   HeNormal initializer object, output shape (None, 10) -> Option B

4. Quick Check:

   HeNormal used and output units = 10 [OK]

Hint: Dense output shape is (None, units); initializer matches given class [OK]

Common Mistakes:

• Confusing output shape with kernel shape
• Assuming default initializer instead of HeNormal
• Mixing initializer types

4. You wrote this code to initialize weights but get an error:

layer = tf.keras.layers.Dense(5, kernel_initializer='he_normal')
model = tf.keras.Sequential([layer])
model.build(input_shape=(None, 3))
print(layer.kernel_initializer())

What is the cause of the error?

medium

A. Calling kernel_initializer as a function without arguments

B. Using string 'he_normal' instead of initializer object

C. Not specifying bias_initializer

D. Missing input_shape in layer constructor

Solution

1. Step 1: Understand kernel_initializer attribute type

   kernel_initializer is an initializer object, not a callable function without arguments.

2. Step 2: Identify error cause

   Calling kernel_initializer() without arguments causes error because it expects shape and dtype parameters.

3. Final Answer:

   Calling kernel_initializer as a function without arguments -> Option A

4. Quick Check:

   Initializer needs shape args, can't call empty () [OK]

Hint: Don't call kernel_initializer() without shape and dtype [OK]

Common Mistakes:

• Confusing string initializer with object
• Ignoring required arguments for initializer call
• Assuming bias_initializer causes this error

5. You want to initialize weights for a model with mixed activations: ReLU in some layers and sigmoid in others. Which strategy best improves training speed and accuracy?

hard

A. Use HeNormal for ReLU layers and GlorotUniform for sigmoid layers

B. Use GlorotUniform for all layers regardless of activation

C. Use RandomNormal with mean 0 and stddev 1 for all layers

D. Use Zeros initializer for all layers to start from zero

Solution

1. Step 1: Match initializers to activation functions

   HeNormal suits ReLU to maintain variance; GlorotUniform suits sigmoid/tanh to avoid saturation.

2. Step 2: Understand impact on training

   Proper matching prevents slow or stuck training and improves accuracy by keeping gradients healthy.

3. Final Answer:

   Use HeNormal for ReLU layers and GlorotUniform for sigmoid layers -> Option A

4. Quick Check:

   Match initializer to activation for best training [OK]

Hint: Match HeNormal to ReLU, GlorotUniform to sigmoid/tanh [OK]

Common Mistakes:

• Using same initializer for all activations
• Using zeros initializer causing no learning
• Using RandomNormal without scaling