loihub / AI KNOWLEDGE CHECK

How well do you understand AI?

Ten questions. Real reasoning. Test your understanding of machine learning, deep learning and evaluation, then explore the explanation behind every answer.

Free to try. No account needed. Complete the questions, then enter your name, email and country to see your results. This is a self-assessment, not a certification exam.

0 of 10 answered

01Transformers

For one attention head, the weights are [0.88, 0.12] and the corresponding value vectors are [10, 0] and [0, 10]. With no dropout, what is their weighted sum before any output projection?

[10, 10][8.8, 1.2][2, 0][0.88, 0.12]

02Transformers

Why does scaled dot-product attention divide query–key scores by √dₖ, where dₖ is the dimension of each key vector?

To reduce the number of tokensTo normalize scores before softmax so they sum to oneTo control score magnitude as the key dimension growsTo remove positional information

03Generative models

A VAE uses a diagonal Gaussian approximate posterior. Its encoder outputs vectors μ and log(σ²). Which expression samples from this distribution? Here log is the natural logarithm, exp acts element-wise, and ⊙ means element-wise multiplication.

z = μ + log(σ²)z = μ + exp(0.5 × log(σ²)) ⊙ ε, where ε ~ N(0, I)z = exp(μ) × σ²z = μ, with no random component

04ML foundations

Binary logistic regression uses logit a = wᵀx + b and p = sigmoid(a). For one label y ∈ {0, 1}, the loss is L = −y ln(p) − (1−y) ln(1−p), without class weights or regularization. What is ∂L/∂a?

p − yy − pp × (1 − p)(p − y)²

05ML foundations

At fixed model parameters, the objective is the mean of per-example losses on a fixed dataset. You sample examples uniformly with replacement and average their gradients. Assuming independent draws and finite gradient variance, what is true?

The estimate always equals the full-dataset gradientThe estimate is unbiased; increasing batch size generally reduces its varianceLarger batches automatically prevent overfittingBatch gradients cannot be unbiased

06ML foundations

For L(w) = ½w², gradient descent starts at w = 1 with learning rate η = 3. What happens over repeated steps?

It converges immediately to zeroIt approaches zero without changing signIt alternates sign and grows in magnitudeIt remains at one

07Deep learning

A convolution receives a 32 × 32 image, with a 5 × 5 kernel, stride 1, padding 0 and dilation 1. What is the output spatial size?

32 × 3230 × 3028 × 2816 × 16

08Deep learning

A unidirectional LSTM classifies complete variable-length sequences, with padding added after the real tokens. Which choice provides a sequence summary without processing the later padding steps?

The hidden state after the last real tokenThe hidden state after the last padding token, without maskingThe sequence length aloneThe first input token for every sequence

09Deep learning

For single-label classification with five mutually exclusive classes and hard class-index labels, what should be passed to PyTorch CrossEntropyLoss with its default settings?

Five raw logits per example and the target class indexThe predicted class index and target class indexFive sigmoid outputs treated as independent binary labelsOnly the highest predicted probability

10Evaluation

You standardize the entire dataset using its mean and standard deviation, then split it into training and test sets. What is the problem?

Standardization makes classification impossibleTest-set information influenced preprocessingThe test set is now too smallThere is no problem if labels were excluded

See your results

Enter your details to view your score and explanations. loihub stores your details and scores privately to administer this assessment. Your details are not shown publicly.

Full name *Email address *Country name *

I have read the privacy notice and understand that loihub will store my details and quiz scores to provide this assessment. *Email me about loihub courses, project reviews and consultancy. Optional; I can withdraw permission through Contact at any time.

Loading loihub…