NCA-GENL Core Machine Learning and AI Knowledge Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  To squash output values to a range of -1 to 1.

  Why it's wrong here

  The Tanh function squashes values into the range (-1, 1). Softmax, by definition, maps outputs to the (0, 1) range and ensures they sum to 1. This probability interpretation is specific to Softmax and is the standard requirement for multi-class classification outputs in deep learning models.

• ✗

  To linearize the output for regression tasks.

  Why it's wrong here

  Regression tasks typically use a linear output layer without an activation function, or sometimes ReLU if values must be positive. Softmax is strictly for classification, as its normalization properties enforce a probability distribution across discrete classes, which would be completely inappropriate for predicting continuous values in a regression problem.

• ✓

  To convert logits into a probability distribution.

  Why this is correct

  Softmax normalizes the raw output scores (logits) by exponentiating them and dividing by the sum of all exponentiated values. This ensures that every output represents the probability of that class, and the sum of all probabilities is 1.0, which is necessary for multi-class classification tasks.

• ✗

  To enforce sparsity on the output vector.

  Why it's wrong here

  Softmax does not enforce sparsity; in fact, it ensures that every class has a non-zero probability. Sparsity is typically achieved through techniques like L1 regularization or ReLU, which can set values to exactly zero. Softmax forces all classes to contribute to the final probability distribution, preventing zero-value outputs.