Deep Learning – IIT Ropar Week 6 Assignment Answers

1 day ago
Sanket Kumar
3 minutes
Deep Learning - IIT Ropar

Deep Learning – IIT Ropar Week 6 Assignment Answers (Jan-Apr 2026)

1. The team designs an autoencoder with a hidden layer of size 40. What is the main effect of this design choice?

  • It converts the task into supervised learning
  • It enforces compression of input information
  • It guarantees exact reconstruction under all conditions
  • It removes the need for explicit regularization during training
Answer : b

2. Why could choosing a hidden layer larger than the input dimension be problematic in this scenario?

  • It significantly slows down inference during deployment
  • It restricts the diversity of learned features
  • It may allow the model to learn an identity mapping
  • It prevents anomaly detection by eliminating reconstruction error
Answer : c

3. Which decoder activation function is most appropriate for reconstructing the original sensor values?

  • Sigmoid
  • tanh
  • Linear
  • Softmax
Answer : c

4. The team notices that reconstruction error is consistently higher for certain late-night data samples. What is the most reasonable explanation?

  • The optimizer is failing to converge properly
  • These samples differ from the training distribution
  • The decoder weights are accidentally frozen
  • The loss function is incorrectly scaled across batches
Answer : b

5. Why is mean squared error a suitable loss function for this autoencoder?

  • It promotes sparsity in hidden layers automatically
  • It models continuous outputs with Gaussian noise assumptions
  • It requires sigmoid activations in the decoder
  • It is unaffected by feature scaling or normalization
Answer : b

6. Even after normalizing the sensor inputs, which statement remains valid?

  • Binary cross-entropy becomes preferable for reconstruction
  • Linear decoders are still appropriate
  • Outputs must be strictly binary
  • PCA equivalence no longer holds in practice
Answer : b

7. Under which conditions does an autoencoder behave similarly to PCA?

  • Nonlinear encoder with dropout regularization
  • Linear encoder, linear decoder, squared error loss
  • Sparse encoder with sigmoid activation
  • Overcomplete hidden representation with constraints
Answer : b

8. Despite this similarity, why might PCA still be preferred in some cases?

  • PCA supports nonlinear mappings
  • PCA has a closed-form solution and stable optimization
  • PCA requires labeled data for training
  • PCA produces lower-dimensional outputs by default
Answer : b

9. Which autoencoder variant is most suitable for learning robust representations from corrupted inputs?

  • Linear autoencoder
  • Overcomplete autoencoder
  • Denoising autoencoder
  • PCA-based encoder
Answer : c

10. What is the defining characteristic of this training strategy?

  • Introducing labels during reconstruction
  • Randomly corrupting inputs during training
  • Increasing decoder depth significantly
  • Sharing encoder weights across layers
Answer : b

11. Why might a standard autoencoder perform poorly when tested on noisy medical data?

  • It requires supervised labels for learning
  • It assumes binary inputs by design
  • It may overfit to clean training data
  • It enforces strong regularization by default
Answer : c

12. The team introduces a sparsity constraint on the hidden layer. What is the intended outcome?

  • Faster gradient updates during backpropagation
  • Learning selective and interpretable features
  • Guaranteed dimensionality reduction
  • Removal of all noise from inputs
Answer : b

13. Which regularization technique explicitly penalizes sensitivity of hidden representations to small input changes?

  • Weight decay
  • Dropout
  • Contractive regularization
  • Early stopping
Answer : c

14. Why is learning a trivial identity mapping particularly harmful in medical signal analysis?

  • It increases inference time on edge devices
  • It hides abnormal or clinically relevant patterns
  • It causes unstable gradients during training
  • It reduces the depth of the network
Answer : b

15. Which combination best helps prevent identity mapping in high-capacity autoencoders?

  • Larger hidden layers without constraints
  • Noise injection, sparsity, and contractive penalties
  • Linear encoders with no regularization
  • Removing the decoder entirely
Answer : b