Rehber bağlantılı sınav · Kolay Seviye

Derin Öğrenme Sınavı

Çok katmanlı ağları, parametre saymayı, doğrusal olmamayı, eğitime karşı çıkarımı, genellemeyi ve model puanlarının kalibre edilmiş yorumunu test edin.

İlgili kılavuz yollarıDerin Öğrenme
Soru 1 arasında 8

What does “deep” describe in a deep-learning model?

Her soruyu önizleyin8 sorular

Testi çözdüğünüzde cevaplar ve açıklamalar görünür.

  1. What does “deep” describe in a deep-learning model?
    • A neural-network architecture with multiple layers between input and output
    • A model that understands topics at a profound human level
    • A system trained only on unusually large datasets
    • A method that guarantees higher task accuracy
  2. A network has 2 inputs, hidden layers of 3 and 2 units, and 1 output. With a bias for each hidden/output unit, how many weights and biases are there?
    • 14
    • 20
    • 18
    • 23
  3. Why use nonlinear activations between hidden layers?
    • They convert every hidden unit into a labeled training example
    • They guarantee the network will generalize to unseen data
    • Without them, stacking linear transformations still produces a linear transformation
    • They remove the need for learned weights
  4. During inference, what normally happens to the trained model’s parameters?
    • They are retrained after every input
    • They are replaced by the labels in the prompt
    • They are reset before each output
    • The model uses them to process an input without necessarily updating them
  5. A model fits its training examples closely but performs poorly on comparable held-out data. What pattern does this show?
    • Aşırı uyum
    • Yetersiz uyum
    • Good generalization
    • A guarantee that the labels are correct
  6. Why compare a deeper network with a simpler baseline on held-out examples?
    • Because simpler models always outperform neural networks
    • To see whether added complexity improves the task enough to justify its practical costs
    • Because deeper models cannot be evaluated on new data
    • To avoid checking error patterns
  7. A classifier returns a score of 0.9. What does that alone establish?
    • The prediction is correct 90% of the time on every task
    • The model has been calibrated on deployment data
    • It is a model score, not proof of a calibrated 90% chance of correctness
    • The system understands the example with 90% certainty
  8. What does adding more layers alone prove about a model’s understanding or usefulness?
    • That it understands inputs like a person
    • That it will outperform a simpler model
    • That its errors are harmless
    • Nothing by itself; evaluate it on the intended task and relevant examples