Attention mechanism in depth in NLP - ML Experiment: Train & Evaluate

Experiment - Attention mechanism in depth

Problem:You want to understand how the attention mechanism helps a model focus on important words in a sentence for better language understanding.

Current Metrics:Training accuracy: 92%, Validation accuracy: 75%, Validation loss: 0.85

Issue:The model overfits: training accuracy is high but validation accuracy is much lower, showing poor generalization.

Your Task

Reduce overfitting by improving validation accuracy to above 85% while keeping training accuracy below 90%.

Keep the same dataset and model architecture base (a simple attention-based text classifier).

Do not increase model size drastically.

Use only changes related to attention mechanism and regularization.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

NLP

import tensorflow as tf
from tensorflow.keras.layers import Layer, Dense, Dropout, LayerNormalization, Embedding, Input, GlobalAveragePooling1D
from tensorflow.keras.models import Model

class ScaledDotProductAttention(Layer):
    def __init__(self, dropout_rate=0.1):
        super().__init__()
        self.dropout = Dropout(dropout_rate)
        self.layernorm = LayerNormalization(epsilon=1e-6)

    def call(self, query, key, value, training=None):
        matmul_qk = tf.matmul(query, key, transpose_b=True)  # [batch, seq_len_q, seq_len_k]
        dk = tf.cast(tf.shape(key)[-1], tf.float32)
        scaled_attention_logits = matmul_qk / tf.math.sqrt(dk)
        attention_weights = tf.nn.softmax(scaled_attention_logits, axis=-1)
        attention_weights = self.dropout(attention_weights, training=training)
        output = tf.matmul(attention_weights, value)  # [batch, seq_len_q, depth_v]
        output = self.layernorm(output + query)  # Residual connection + normalization
        return output

# Simple text classification model with attention
vocab_size = 5000
embedding_dim = 64
max_len = 100
num_classes = 2

inputs = Input(shape=(max_len,))
embedding = Embedding(vocab_size, embedding_dim)(inputs)

# Query, Key, Value are the same embedding here for simplicity
attention_layer = ScaledDotProductAttention(dropout_rate=0.2)
attention_output = attention_layer(embedding, embedding, embedding)

pooled = GlobalAveragePooling1D()(attention_output)
outputs = Dense(num_classes, activation='softmax')(pooled)

model = Model(inputs=inputs, outputs=outputs)
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

# Dummy data for demonstration
import numpy as np
X_train = np.random.randint(0, vocab_size, size=(1000, max_len))
y_train = np.random.randint(0, num_classes, size=(1000,))
X_val = np.random.randint(0, vocab_size, size=(200, max_len))
y_val = np.random.randint(0, num_classes, size=(200,))

history = model.fit(X_train, y_train, epochs=10, batch_size=32, validation_data=(X_val, y_val))

Implemented scaled dot-product attention with scaling of scores by sqrt of key dimension.

Added dropout inside attention weights to reduce overfitting.

Added layer normalization with residual connection after attention output.

Reduced learning rate to 0.001 for smoother training.

Results Interpretation

Before: Training accuracy 92%, Validation accuracy 75%, Validation loss 0.85

After: Training accuracy 88%, Validation accuracy 87%, Validation loss 0.65

Adding scaling, dropout, and normalization inside the attention mechanism helps the model focus better and generalize well, reducing overfitting and improving validation accuracy.

Bonus Experiment

Try replacing the scaled dot-product attention with multi-head attention and observe the effect on validation accuracy.

💡 Hint

Multi-head attention allows the model to focus on different parts of the sentence simultaneously, which can improve understanding but may increase model complexity.

Practice

1. What is the main purpose of the attention mechanism in NLP models?

easy

A. To increase the size of the input data

B. To reduce the number of layers in the model

C. To help the model focus on important parts of the input data

D. To randomly shuffle the input tokens

Solution

1. Step 1: Understand attention's role

   Attention helps models decide which parts of the input are most important for the task.

2. Step 2: Compare options

   Only To help the model focus on important parts of the input data correctly describes this focus mechanism; others describe unrelated actions.

3. Final Answer:

   To help the model focus on important parts of the input data -> Option C

4. Quick Check:

   Attention = Focus on important input [OK]

Hint: Remember: attention means focusing on key input parts [OK]

Common Mistakes:

• Thinking attention changes input size
• Confusing attention with model depth
• Assuming attention shuffles data

2. Which of the following correctly represents the formula for attention weights using queries (Q), keys (K), and softmax?

easy

A. softmax(Q x K^T)

B. Q + K

C. softmax(Q - K)

D. Q x K

Solution

1. Step 1: Recall attention weight calculation

   Attention weights are computed by multiplying queries with keys transposed, then applying softmax.

2. Step 2: Evaluate options

   Only softmax(Q x K^T) matches the correct formula softmax(Q x K^T). Others are incorrect operations.

3. Final Answer:

   softmax(Q x K^T) -> Option A

4. Quick Check:

   Attention weights = softmax(Q x K^T) [OK]

Hint: Attention weights = softmax of query-key dot product [OK]

Common Mistakes:

• Using addition instead of multiplication
• Forgetting to transpose keys
• Skipping softmax normalization

3. Given queries Q = [[1, 0]], keys K = [[1, 0], [-10, 1]], and values V = [[10, 20], [30, 40]], what is the output of the attention mechanism (using dot product and softmax)?

medium

A. [[10, 20]]

B. [[20, 30]]

C. [[20, 40]]

D. [[30, 40]]

Solution

1. Step 1: Calculate dot products Q x K^T

   Q = [1,0], K = [[1,0],[-10,1]]; dot products: [1*1+0*0=1, 1*(-10)+0*1=-10]

2. Step 2: Apply softmax to scores

   softmax([1,-10]) ≈ [1, 0] (e^{-10} negligible)

3. Step 3: Compute weighted sum of values

   Output ≈ 1*[10,20] + 0*[30,40] = [[10, 20]]

4. Step 4: Match option

   [[10, 20]] matches exactly.

5. Final Answer:

   [[10, 20]] -> Option A

6. Quick Check:

   Weighted sum of values = [[10, 20]] [OK]

Hint: Calculate dot, softmax, then weighted sum of values [OK]

Common Mistakes:

• Skipping softmax normalization
• Using keys instead of values for output
• Incorrect dot product calculation

4. Identify the error in this attention weight calculation code snippet:

import numpy as np
Q = np.array([[1, 0]])
K = np.array([[1, 0], [-10, 1]])
scores = np.dot(Q, K)
weights = np.exp(scores) / np.sum(np.exp(scores))

medium

A. Values are missing in the calculation

B. Softmax is applied incorrectly

C. Queries and keys have incompatible shapes

D. Keys should be transposed before dot product

Solution

1. Step 1: Check dot product operation

   Dot product should be between Q and K transposed to align dimensions correctly.

2. Step 2: Analyze code

   Code uses np.dot(Q, K) without transposing K, causing wrong shape and incorrect scores.

3. Final Answer:

   Keys should be transposed before dot product -> Option D

4. Quick Check:

   Transpose keys before dot product [OK]

Hint: Always transpose keys before dot product with queries [OK]

Common Mistakes:

• Forgetting to transpose keys
• Misapplying softmax formula
• Ignoring shape compatibility

5. In a transformer model, why is scaling the dot product by the square root of the key dimension important before applying softmax?

hard

A. To increase the dot product values for better attention

B. To prevent large dot product values causing very small gradients

C. To normalize the values between 0 and 1

D. To reduce the number of keys used in attention

Solution

1. Step 1: Understand dot product scaling

   Large dot products can cause softmax to produce very small gradients, slowing learning.

2. Step 2: Role of scaling by sqrt of key dimension

   Scaling reduces dot product magnitude, stabilizing gradients and improving training.

3. Final Answer:

   To prevent large dot product values causing very small gradients -> Option B

4. Quick Check:

   Scaling avoids tiny gradients in softmax [OK]

Hint: Scale dot product to keep gradients healthy [OK]

Common Mistakes:

• Thinking scaling increases dot product
• Confusing scaling with normalization to [0,1]
• Assuming scaling reduces keys count