Machine Learning and Deep Learning practice questions

Accuracy

Why wrong: Accuracy is not suitable for imbalanced datasets as it can be high even if the model fails to detect fraud.

Recall

Why wrong: Recall only considers false negatives, not false positives, so it is insufficient alone.

Precision

Why wrong: Precision only considers false positives, not false negatives, so it does not fully capture model performance on fraud detection.

Increase learning rate

Why wrong: Increasing the learning rate can lead to divergence and instability during training.

Feature scaling

Why wrong: Feature scaling normalizes input ranges but does not directly address sensitivity to small changes.

Cross-validation

Why wrong: Cross-validation is used to assess model performance, not to improve prediction stability.

Batch normalization

Why wrong: Batch normalization normalizes layer inputs, speeding up training but not primarily used to reduce overfitting.

Increase number of epochs

Why wrong: Increasing epochs would likely cause more overfitting.

Gradient clipping

Why wrong: Gradient clipping prevents exploding gradients, not overfitting.

Transfer learning

Why wrong: Transfer learning leverages pre-trained models for new tasks but does not inherently support continuous updates.

Batch learning

Why wrong: Batch learning requires retraining on the entire dataset periodically, which is inefficient for continuous data streams.

Unsupervised learning

Why wrong: Unsupervised learning finds patterns without labels, not designed for continuous supervised adaptation.

Normal equation

Why wrong: Normal equation requires computing (X^T X)^{-1}, which is computationally expensive for large datasets.

Batch gradient descent

Why wrong: Batch gradient descent uses the whole dataset for each update, which is slow for large datasets.

Principal component analysis

Why wrong: PCA reduces dimensionality but does not train a model.

Increase the learning rate

Why wrong: Increasing the learning rate may cause the loss to oscillate or diverge.

Increase the number of epochs

Why wrong: More epochs would likely increase overfitting further.

Add more convolutional layers

Why wrong: Adding more layers increases complexity and would likely worsen overfitting.

Association rule mining

Why wrong: Association rule mining finds frequent itemsets, not suited for sparse user-item matrices.

k-nearest neighbors

Why wrong: k-NN relies on similarity measures that may be inaccurate with sparse data.

Content-based filtering

Why wrong: Content-based filtering uses item attributes, not user-item interactions, and does not directly address matrix sparsity.

Normalization

Why wrong: Normalization scales features to a range, not for missing data.

One-hot encoding

Why wrong: One-hot encoding converts categorical variables to binary, not a missing data technique.

Dimensionality reduction

Why wrong: Dimensionality reduction reduces number of features, not for missing data.

Softmax

Why wrong: Softmax is used for multi-class output, but the question asks for common activation functions; it's less common in hidden layers.

Linear

Why wrong: Linear activation is rarely used in hidden layers because it makes the network linear.

Increase batch size

Why wrong: Increasing batch size can sometimes lead to sharper minima and may not reduce overfitting.

Increase learning rate

Why wrong: Increasing learning rate can cause instability, not reduce overfitting.

Gradient clipping

Why wrong: Gradient clipping prevents exploding gradients, not overfitting.

Batch normalization is disabled

Why wrong: Disabling batch normalization may slow training but would not cause constant loss.

The dataset is imbalanced

Why wrong: Imbalanced data would cause the model to predict the majority class, not constant loss at 2.3026.

The model is overfitting

Why wrong: Overfitting would show decreasing training loss and increasing validation loss, not constant values.

100,224

Why wrong: This would be 784*128 without biases.

109,258

Why wrong: This is the total number of parameters across all layers.

8,256

Why wrong: This is the parameter count for the second dense layer.

Remove one or more hidden layers from the network

Why wrong: Removing layers may cause underfitting.

Increase the number of training epochs

Why wrong: Increasing epochs typically increases overfitting.

Add more features to the input data

Why wrong: Adding features is not a primary remedy for overfitting in image classification.

Accuracy

Why wrong: Accuracy is misleading when classes are imbalanced.

Log loss

Why wrong: Log loss does not directly reflect performance on the minority class.

AUC-ROC

Why wrong: AUC-ROC is useful but less direct for the minority class.

Increase the reward for reaching the destination to +10

Why wrong: Increasing destination reward may still not overcome the reward for circling if the agent can accumulate many steps.

Use a discount factor closer to 0

Why wrong: A low discount factor makes the agent short-sighted, which could worsen the problem.

Add a penalty for each turn the vehicle makes

Why wrong: Penalizing turns may prevent necessary navigation and does not address the core issue.

Apply Principal Component Analysis (PCA) to reduce dimensionality

Why wrong: PCA may discard useful information and is not necessary for Random Forest.

Normalize the numerical features to have zero mean and unit variance

Why wrong: Random Forest is not sensitive to feature scaling.

Encode all features using one-hot encoding

Why wrong: One-hot encoding is not universally needed and may increase dimensionality unnecessarily.

Using a larger learning rate

Why wrong: Larger learning rates can cause divergence, not prevent overfitting.

L1 regularization

Why wrong: L1 regularization is used but not as common as dropout and early stopping for overfitting.

Increasing the number of layers

Why wrong: Adding layers often increases overfitting.

Underfitting

Why wrong: Underfitting would show high loss on both training and validation sets.

Data leakage

Why wrong: Data leakage would cause overly optimistic performance, not divergence.

Vanishing gradient

Why wrong: No evidence of gradients vanishing; loss is decreasing on training.

Undersample the majority class to balance the dataset

Why wrong: Undersampling discards valuable data and may reduce model performance.

Oversample the minority class using synthetic image generation

Why wrong: Oversampling can cause overfitting, especially with limited positive samples.

Replace the CNN with a transformer-based architecture

Why wrong: Changing architecture does not directly address class imbalance.

Add the customer support text as a feature using TF-IDF vectors

Why wrong: Adding text features requires significant preprocessing and may increase training time.

Use an ensemble of gradient boosting and random forest models

Why wrong: Ensemble methods increase training time and may not be justified.

Perform hyperparameter tuning using grid search

Why wrong: Hyperparameter tuning can help but may not provide the largest improvement.