Textbook of AI

Self-assessment

A short multi-choice quiz. Click an option to commit; the correct answer and an explanation appear. Your answers are remembered in this browser.

1. Question 1. A standard feed-forward neural network with no hidden layers, sigmoid output, and cross-entropy loss is mathematically equivalent to:

   1. Logistic regression
   2. A support vector machine with a linear kernel
   3. Principal-component regression
   4. A k-nearest-neighbours classifier with $k = 1$

2. Question 2. The vanishing-gradient problem in deep networks is essentially:

   1. Gradients shrinking exponentially when backpropagated through many layers with squashing activations
   2. The optimiser oscillating around the minimum
   3. The loss going to infinity
   4. The network failing to converge because the learning rate is too high

3. Question 3. A convolutional layer is preferred over a fully-connected layer for images mainly because:

   1. Weight sharing exploits translation equivariance and dramatically reduces the parameter count
   2. Convolutions are mathematically more powerful than dense layers
   3. Convolutions cannot overfit
   4. Convolutions are required for the gradient to exist

4. Question 4. A variational autoencoder differs from a plain autoencoder in that it:

   1. Imposes a distribution over the latent code and trains by maximising an evidence lower bound (ELBO)
   2. Uses convolutional layers instead of dense layers
   3. Replaces the reconstruction loss with adversarial loss
   4. Removes the decoder

5. Question 5. In Q-learning, the Bellman update for $Q(s, a)$ after observing transition $(s, a, r, s')$ is:

   1. $Q(s,a) \leftarrow Q(s,a) + \alpha \big[ r + \gamma \max_{a'} Q(s', a') - Q(s,a) \big]$
   2. $Q(s,a) \leftarrow r + \gamma Q(s', a)$
   3. $Q(s,a) \leftarrow Q(s,a) \cdot (1 + \alpha r)$
   4. $Q(s,a) \leftarrow \arg\max_{a'} Q(s,a')$

6. Question 6. In-context learning — a large language model solving a new task from a prompt with a few examples — is best described as:

   1. An emergent behaviour: the model's weights do not change; the few-shot examples merely condition its forward pass
   2. An automatic fine-tuning step applied to the model on each prompt
   3. A retrieval step that looks up the answer in a database
   4. A separately-trained mechanism that overrides the main model