AI Concepts and Techniques practice questions

The learning rate is too high; decrease it to 0.0001

Why wrong: A high learning rate would cause the loss to oscillate or increase, not slowly decrease.

The model is overfitting; add L2 regularisation

Why wrong: Overfitting does not cause slow decrease of loss; regularisation may reduce overfitting but not address the learning rate.

The batch size is too large; reduce it

Why wrong: Batch size affects noise but not the rate of decrease as directly as learning rate.

Accuracy

Why wrong: Accuracy is high even if the model predicts negative for all cases, hiding poor performance on positives.

Perplexity

Why wrong: Perplexity is used for language model evaluation, not binary classification.

Variational autoencoder (VAE)

Why wrong: VAEs generate images but often produce blurry outputs compared to diffusion models.

Generative adversarial network (GAN)

Why wrong: GANs can be difficult to train and may not match the quality and diversity of modern diffusion models.

Recurrent neural network (RNN)

Why wrong: RNNs are designed for sequential data, not image generation.

Use a smaller batch size

Why wrong: Smaller batch sizes can increase variance, potentially making fluctuations worse.

Increase the learning rate

Why wrong: Increasing learning rate would likely worsen fluctuations.

Increase the number of epochs

Why wrong: More epochs don't address stability; fluctuations may persist.

Increase the number of epochs

Why wrong: More epochs can lead to overfitting without other regularization.

Use L2 regularization

Why wrong: L2 regularization helps but is generally less effective for CNNs compared to dropout and augmentation.

Add more convolutional layers

Why wrong: Adding more layers increases model capacity, which can worsen overfitting.

Train a custom model from scratch on the policy documents each month

Why wrong: Training from scratch requires massive compute resources and weeks of time — entirely disproportionate for monthly document updates.

Use a larger foundation model with a longer context window and paste all documents into each prompt

Why wrong: Pasting all documents into every prompt is expensive, hits context limits for large document sets, and does not scale as the document library grows.

Fine-tune a base LLM on the policy documents monthly

Why wrong: Fine-tuning is expensive and time-consuming; monthly cycles are impractical and fine-tuned knowledge becomes stale immediately after cutoff.

Reinforcement learning

Why wrong: Reinforcement learning involves an agent interacting with an environment, learning from rewards.

Regression

Why wrong: Regression predicts continuous numeric values, not binary outcomes.

Clustering

Why wrong: Clustering is unsupervised and does not use labeled outcomes.

Vanishing gradients

Why wrong: Vanishing gradients would prevent loss from decreasing at all, not cause a divergence.

Incorrect learning rate scheduling

Why wrong: While learning rate issues can cause problems, the pattern described is classic overfitting.

Underfitting

Why wrong: Underfitting would show high loss on both sets, not an increasing validation loss.

Decrease the top-p value

Why wrong: Decreasing top-p reduces the set of tokens considered, often making outputs less diverse.

Increase the frequency penalty

Why wrong: Frequency penalty reduces repetition, which could help with creativity, but temperature directly impacts creativity more effectively.

Decrease the max tokens limit

Why wrong: Decreasing max tokens makes the output shorter, which is opposite of what is desired.

Use chain-of-thought prompting

Why wrong: Chain-of-thought encourages step-by-step reasoning, often resulting in longer outputs.

Decrease the top-k value

Why wrong: Top-k limits token choices but does not specifically encourage conciseness.

Increase the temperature

Why wrong: Higher temperature increases randomness, which may produce more verbose or irrelevant content.

It can perform any intellectual task that a human can

Why wrong: That describes General AI.

It requires no training data

Why wrong: Most narrow AI systems require training data.

It surpasses human intelligence in all domains

Why wrong: That describes Super AI.

Set the temperature to 0

Why wrong: Temperature affects output randomness, not injection vulnerability.

Use a system prompt that instructs the model to ignore injection attempts

Why wrong: System prompts can be overridden by user input; not reliable.

Use a larger, more powerful LLM

Why wrong: Larger models may still be vulnerable to injection.

Generative Adversarial Network (GAN)

Why wrong: Used for generating images, not classification.

Recurrent Neural Network (RNN)

Why wrong: Designed for sequences, not images.

Multi-layer Perceptron (MLP)

Why wrong: Less effective for image classification due to lack of spatial hierarchy.

BLEU score

Why wrong: BLEU measures text generation quality, not classification performance.

Accuracy

Why wrong: Accuracy can be misleading in imbalanced datasets; a model predicting all legitimate achieves 99.9% accuracy but fails to catch fraud.

Perplexity

Why wrong: Perplexity is used for language models, not for binary classification.

Unsupervised learning

Why wrong: Unsupervised learning finds patterns in unlabeled data without rewards.

Supervised learning

Why wrong: Supervised learning uses labeled data, not a reward-based environment.

Self-supervised learning

Why wrong: Self-supervised learning generates labels from the data itself, not from environmental rewards.

Transformer

Why wrong: Transformers process sequences in parallel using attention, not a recurrent hidden state.

Convolutional Neural Network (CNN)

Why wrong: CNNs are designed for spatial data like images, not sequences.

Multi-layer Perceptron (MLP)

Why wrong: MLPs are feedforward networks without sequential memory.

Supervised classification with logistic regression

Why wrong: Requires labeled data, which is not available here.

Semi-supervised learning with a small labeled set

Why wrong: Semi-supervised still requires some labels, which are not available.

Reinforcement learning with a reward function

Why wrong: RL is for sequential decision-making, not static grouping.

Increase learning rate

Why wrong: Higher learning rate may cause divergence, not reduce overfitting.

Increase the number of layers

Why wrong: More layers increase model capacity, likely worsening overfitting.

Reduce training data size

Why wrong: Less data typically increases overfitting, not reduces it.

Remove dropout layers to allow more gradient flow

Why wrong: Removing dropout may increase overfitting but not necessarily help convergence.

Use a smaller batch size to reduce memory

Why wrong: Smaller batch size introduces noise, which may not help convergence.

Increase the learning rate by 10x

Why wrong: A high learning rate may cause divergence, not convergence.

Remove the minority class samples to make the dataset balanced

Why wrong: Removing samples discards valuable data and reduces model's ability to learn the minority class.

Report only accuracy as the evaluation metric

Why wrong: Accuracy is misleading for imbalanced data.