First neural network in TensorFlow - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of a simple neural network prediction

What is the output of the following code snippet that creates a simple neural network and predicts on a single input?

TensorFlow

import tensorflow as tf
import numpy as np

model = tf.keras.Sequential([
    tf.keras.layers.Dense(1, input_shape=(2,), activation='linear')
])

model.layers[0].set_weights([np.array([[1.0],[2.0]]), np.array([0.5])])

input_data = np.array([[3.0, 4.0]])
prediction = model.predict(input_data)
print(prediction)

A [[14.0]]

B [[11.5]]

C [[10.5]]

D [[7.5]]

💡 Hint

Remember the formula for a dense layer output: output = input * weights + bias.

The weights matrix is [[1.0],[2.0]] and bias is [0.5]. For input [3.0, 4.0], output = 3*1 + 4*2 + 0.5 = 3 + 8 + 0.5 = 11.5.

❓ Model Choice

intermediate

Choosing the correct activation function for regression

You want to build a neural network to predict house prices (a continuous value). Which activation function should you use in the output layer?

A Linear

B Sigmoid

C ReLU

D Softmax

💡 Hint

For continuous output without limits, choose an activation that does not restrict output range.

Linear activation allows output to be any real number, suitable for regression tasks like predicting prices.

❓ Hyperparameter

advanced

Effect of batch size on training

Which statement best describes the effect of increasing batch size during neural network training?

A Larger batch sizes always improve model generalization.

B Batch size does not affect training speed or memory usage.

C Increasing batch size reduces the number of training epochs needed but increases memory usage.

D Smaller batch sizes provide noisier gradient estimates but can improve generalization.

💡 Hint

Think about how batch size affects gradient noise and memory.

Smaller batches give noisier gradients which can help escape local minima and improve generalization, but training is slower. Larger batches use more memory and can speed up training but may reduce generalization.

❓ Metrics

advanced

Choosing the right metric for classification

You train a neural network to classify images into 3 categories. Which metric is best to evaluate overall model accuracy?

A Mean Absolute Error

B Mean Squared Error

C Accuracy

D Root Mean Squared Error

💡 Hint

Think about metrics that measure correct class predictions.

Accuracy measures the proportion of correct predictions and is suitable for classification tasks.

🔧 Debug

expert

Identifying the cause of training failure

You train a neural network but the loss stays constant and does not decrease. What is the most likely cause?

TensorFlow

import tensorflow as tf

model = tf.keras.Sequential([
    tf.keras.layers.Dense(10, activation='relu', input_shape=(5,)),
    tf.keras.layers.Dense(1, activation='sigmoid')
])

model.compile(optimizer='adam', loss='binary_crossentropy')

import numpy as np
X = np.random.rand(100, 5)
y = np.ones((100, 1))

model.fit(X, y, epochs=10)

A The labels y are all ones, causing no gradient to update weights.

B The optimizer 'adam' is not suitable for binary classification.

C The input shape is incorrect for the first layer.

D The model uses sigmoid activation but binary_crossentropy loss, which is incompatible.

💡 Hint

Check if the labels provide enough variation for learning.

If all labels are the same, the model cannot learn meaningful differences, so loss stays constant.