¡GoConqr estrena diseño! Cuéntanos qué te parece.

Regístrate gratis

Test

Sin título

Jewella Macabenta 33 preguntas

Hacer el test

Preguntas del test (33)

  1. 1In standard SGD with Momentum (SGDM), which statement best describes the update?
    1. AMomentum changes the gradient computation during backprop by adding a “momentum gradient” term at each layer.
    2. BBackprop computes gradients normally; momentum accumulates a velocity from past gradients and then updates weights using that velocity.
    3. CMomentum replaces the local gradient δ at each layer.
    4. DMomentum performs bias-correction like Adam before applying weight decay.
  2. 2Which best matches the meaning of local gradient and weight correction in BP-based training?
    1. ALocal gradient is the optimizer’s momentum term; weight correction is the raw gradient.
    2. BLocal gradient refers to per-layer error signals; weight correction refers to parameter updates computed from gradients and optimizer rules.
    3. CLocal gradient is the loss value; weight correction is the learning rate schedule.
    4. DLocal gradient exists only in Adam; weight correction exists only in SGDM.
  3. 3Which statement best describes the difference between plain SGD and SGD with Momentum (SGDM) during training?
    1. ASGDM changes backpropagation by modifying the local gradients 𝛿 at each layer, while SGD does not.
    2. BBoth use the same gradients from backprop; SGDM adds a velocity (momentum) state that accumulates past gradients and uses it to compute the weight update.
    3. CSGDM updates weights using only the sign of the gradient, while SGD uses the full gradient magnitude.
    4. DSGDM removes the need for learning rate tuning because momentum fully determines the step size.
  4. 4A typical training iteration for a mini-batch is:
    1. AUpdate weights → forward pass → compute loss → backward pass
    2. BForward pass → compute loss → backward pass (compute gradients) → optimizer step (update weights)
    3. CBackward pass → forward pass → compute loss → optimizer step
    4. DCompute loss → optimizer step → forward pass → backward pass
  5. 5For single-label multi-class classification (exactly one correct class per sample), the most standard choice is:
    1. ASigmoid + Binary Cross-Entropy
    2. BSoftmax + Categorical Cross-Entropy
    3. CSoftmax + Mean Squared Error
    4. DLinear output + Hinge loss only
  6. 6For multi-label classification (multiple classes can be true simultaneously), which pairing is most appropriate?
    1. ASoftmax + Categorical Cross-Entropy
    2. BSigmoid (per class) + Binary Cross-Entropy (per class)
    3. CSoftmax + Binary Cross-Entropy
    4. DSigmoid + Categorical Cross-Entropy
  7. 7Many libraries offer "cross-entropy from logits" (i.e., internally applies log-softmax). Which is the best practice for numerical stability?
    1. AApply softmax in the model, then pass probabilities to a “from logits” loss.
    2. BDo not apply softmax in the model; pass raw logits to a cross-entropy loss that internally uses log-softmax.
    3. CAlways apply softmax twice (once in model, once in loss) to improve calibration.
    4. DReplace softmax with ReLU before the loss.
  8. 8Which statement is most correct?
    1. APadding changes the number of channels; stride changes only the number of filters.
    2. BIn most deep learning frameworks, “convolution” is implemented as cross-correlation (no kernel flip), but still works for learning.
    3. CBatchNorm must be placed after the final loss layer to normalize labels.
    4. DReLU is typically applied before the convolution to “linearize” the input.
  9. 9Why do many CNN architectures include max pooling (or downsampling)?
    1. ATo increase the number of channels while keeping spatial size fixed.
    2. BTo reduce spatial resolution, increasing receptive field and improving translation robustness, while lowering computation.
    3. CTo make the network strictly linear so it is easier to optimize.
    4. DTo ensure gradients are always non-zero.
  10. 10Why do CNNs often reduce spatial resolution (downsample) in deeper layers?
    1. ATo reduce computation/memory and increase the effective receptive field, while improving robustness to small translations.
    2. BBecause convolution cannot be applied to large feature maps.
    3. CBecause downsampling guarantees higher accuracy on any task.
    4. DBecause gradients vanish unless feature maps are small.

Información

Descripción
Student AISE 3010 Test sobre Sin título, creado por Jewella Macabenta el 08/02/2026.
Idioma
English
Visitas
6
Valoración
Aún sin valoraciones
Valora este test
Valora este test

Escríbenos

¿Sobre qué es tu mensaje?

¿Qué te parece el nuevo diseño?

¿Te gusta?