Introduction

There is no single global "default initializer" for every variable in TensorFlow. The correct answer depends on whether you are creating a raw tf.Variable yourself or letting a Keras layer create weights for you, because Keras layers define their own initializer defaults.

Raw tf.Variable Has No Hidden Default Value

When you create a tf.Variable directly, you provide the initial value explicitly. TensorFlow does not silently pick one for you.

1import tensorflow as tf
2
3weights = tf.Variable([[1.0, 2.0], [3.0, 4.0]])
4print(weights)

Here the initializer is effectively the value you passed in. If you want zeros, random values, or something else, you construct that value yourself:

1import tensorflow as tf
2
3zeros = tf.Variable(tf.zeros((2, 3)))
4random_values = tf.Variable(tf.random.normal((2, 3)))
5
6print(zeros)
7print(random_values)

So for raw tf.Variable, the question "what is the default initializer?" is the wrong framing. There is no global default initializer applied behind the scenes.

Keras Layers Do Have Default Initializers

Confusion usually comes from Keras layers, because they create trainable variables automatically. In that context, each weight type has a default initializer.

For a Dense layer, the default kernel initializer is usually GlorotUniform, while the default bias initializer is zeros.

1import tensorflow as tf
2
3layer = tf.keras.layers.Dense(8)
4layer.build((None, 4))
5
6print(layer.kernel_initializer)
7print(layer.bias_initializer)

That distinction matters:

• kernel weights often default to GlorotUniform,
• bias values often default to zeros,
• other layers may use different defaults.

So the most accurate answer is not "TensorFlow defaults to GlorotUniform." It is "many Keras layers, such as Dense, use GlorotUniform for kernels by default."

Why GlorotUniform Is Common

GlorotUniform, also called Xavier uniform initialization, is designed to keep activation and gradient scales reasonably balanced across layers. That makes it a strong general-purpose choice for many feed-forward networks.

You can inspect it directly:

1import tensorflow as tf
2
3initializer = tf.keras.initializers.GlorotUniform()
4values = initializer(shape=(4, 8))
5
6print(values.shape)
7print(values.dtype)

It is common because it works well across many ordinary architectures. That said, it is not universally optimal. For ReLU-heavy models, HeNormal or HeUniform can be a better fit. For embeddings or recurrent layers, the defaults may differ.

Setting the Initializer Explicitly

If you do not want to rely on layer defaults, declare the initializer directly.

1import tensorflow as tf
2
3layer = tf.keras.layers.Dense(
4    8,
5    activation="relu",
6    kernel_initializer=tf.keras.initializers.HeNormal(),
7    bias_initializer="zeros",
8)

Being explicit is often a good idea in experiments, especially when comparing training behavior across architectures.

You can also define variables yourself with a chosen initializer:

1import tensorflow as tf
2
3initializer = tf.keras.initializers.GlorotUniform()
4weights = tf.Variable(initializer(shape=(4, 8)), trainable=True)
5
6print(weights.shape)

This makes the initialization strategy visible instead of leaving it implicit.

Inspecting the built layer after build() or after the first forward pass is a practical habit. It lets you confirm which initializer the framework actually attached instead of relying on memory or assumptions.

Common Pitfalls

• Saying TensorFlow has one universal default initializer for all variables.
• Confusing raw tf.Variable creation with Keras layer-created weights.
• Assuming kernel and bias initializers use the same default.
• Believing GlorotUniform is always the best choice regardless of activation function or architecture.
• Forgetting to inspect layer defaults when debugging unstable training.

Summary

• Raw tf.Variable objects do not use a hidden default initializer; you pass the initial value yourself.
• Keras layers create weights automatically and often define initializer defaults.
• For many dense-style Keras layers, the kernel default is GlorotUniform and the bias default is zeros.
• Different layer types can use different defaults, so avoid overgeneralizing.
• If initialization matters for the experiment, set it explicitly rather than relying on assumptions.