Deep Learning MCQ Questions And Answers

Deep Learning MCQ Questions And Answers

71. What is deep reinforcement learning?

  1. Deep reinforcement learning is a technique for deeply analyzing the rewards received by neural networks
  2. Deep reinforcement learning combines deep neural networks with reinforcement learning — using neural networks as function approximators to learn optimal policies or value functions from high-dimensional state spaces, enabling agents to learn complex behaviors from raw sensory input
  3. Deep reinforcement learning is identical to supervised learning but with reinforcement signals as labels
  4. Deep reinforcement learning only works in simulated environments and cannot be applied to real-world robotics

Answer : B
Explanation: Deep Reinforcement Learning (Deep RL) combines the representation power of deep learning with the decision-making framework of reinforcement learning. Key concepts: Agent: the decision-making entity (the neural network). Environment: where the agent acts. State: current observation (pixel images, sensor data). Action: what the agent does. Reward: numerical feedback from the environment. Policy: maps states to actions — what the neural network learns. Value function: estimates expected future reward from a state. Landmark achievements: DQN (DeepMind, 2013): first deep RL system to learn Atari games from raw pixels. Deep Q-Network: convolutional network + experience replay + target network. AlphaGo (DeepMind, 2016): beat world champion Go player — combined MCTS with deep RL. AlphaZero (2017): learns from self-play alone — mastered Chess, Go, Shogi. OpenAI Five (2019): beat world champions at Dota 2 5v5. ChatGPT RLHF: PPO algorithm used to align language models. Key algorithms: DQN, PPO (Proximal Policy Optimization), A3C, SAC, TD3. Applications: Robotics (manipulation, locomotion), autonomous vehicles, recommendation systems, chip design, drug discovery, trading strategies.

72. What is the exploding gradient problem in deep learning?

  1. Exploding gradient is when the GPU temperature exceeds safe limits during intensive training
  2. The exploding gradient problem occurs when gradients grow exponentially large during backpropagation through many layers or time steps — causing weight updates to become extremely large, making training unstable and the loss function diverge (often to NaN)
  3. Exploding gradients cause the model to learn too quickly and overfit the training data immediately
  4. Exploding gradients only occur in very small networks with fewer than 5 layers

Answer : B
Explanation: Exploding Gradients is the opposite of vanishing gradients — occurs when gradient magnitudes grow exponentially during backpropagation. Causes: chain rule multiplications: if weights are large (>1) and/or activation derivatives are large, each multiplication amplifies the gradient. Particularly problematic in deep RNNs (many time steps = many multiplications). Symptoms: loss suddenly becomes NaN or infinity. Weights become very large (blow up). Model accuracy collapses during training. Parameter values become NaN in debugging. Solutions: Gradient Clipping (most common and effective): clip the gradient norm to a maximum value. if ||∇||₂ > threshold: ∇ = ∇ × (threshold / ||∇||₂). PyTorch: torch.nn.utils.clip_grad_norm_(params, max_norm). Typically threshold = 1.0 or 5.0 for RNNs. Proper weight initialization: He, Xavier initialization prevents initial weights from being too large. Batch Normalization: normalizes activations — prevents very large activations feeding into gradient computations. LSTM/GRU gating: gates control information flow, preventing runaway gradients. Reduce learning rate: smaller steps prevent large weight updates. Weight regularization (L2): penalty discourages very large weights. Gradient clipping is the most direct solution and is standard practice for all RNN/LSTM training.

73. What is deep learning’s role in autonomous vehicles?

  1. Deep learning in autonomous vehicles is only used for infotainment systems and navigation maps
  2. Deep learning powers the perception stack of autonomous vehicles — enabling real-time object detection, lane detection, semantic segmentation, depth estimation, and behavior prediction from camera, LiDAR, and radar inputs to support safe navigation decisions
  3. Autonomous vehicles using deep learning only work on highways and cannot operate in urban environments
  4. Deep learning in autonomous vehicles replaces all traditional sensors with a single camera-based system

Answer : B
Explanation: Autonomous Vehicles (AVs) are one of the most demanding deep learning applications. Perception tasks: Object Detection: detect cars, pedestrians, cyclists, traffic signs, traffic lights. Models: YOLO variants, PointPillars (for LiDAR), 3D object detection. Semantic Segmentation: label every pixel — road, sidewalk, building, vegetation. Lane Detection: detect lane markings, road boundaries. Depth Estimation: monocular depth estimation, stereo depth. Sensor Fusion: combine camera + LiDAR + radar. BEV (Bird’s Eye View): represent scene from a top-down perspective. Prediction: Trajectory prediction: where will other agents move next? Occupancy prediction: which areas will be occupied in the future? Sensor modalities: Cameras: high-resolution, color, texture — but no depth. LiDAR: precise 3D point cloud — but expensive. Radar: works in poor weather — but low resolution. Ultrasonic: short-range parking sensors. Companies and approaches: Tesla (camera-only, FSD + neural networks), Waymo (camera+LiDAR+radar), Cruise, Aurora, Mobileye. End-to-end learning: directly map sensor input → steering/throttle/brake commands. Imitation learning from human drivers. Challenges: long-tail edge cases, sensor failures, adversarial conditions (rain, snow, night), regulatory approval.

74. What is the attention is all you need paper’s contribution to deep learning?

  1. The paper introduced the concept of paying more attention to difficult training examples during training
  2. The paper “Attention Is All You Need” (Vaswani et al., Google, 2017) introduced the Transformer architecture — showing that self-attention alone (without recurrence or convolutions) could achieve state-of-the-art machine translation results with significantly faster training, laying the foundation for BERT, GPT, and all modern large language models
  3. The paper introduced the attention mechanism as a replacement for the loss function in neural networks
  4. The paper showed that human attention patterns should be used to guide neural network training directly

Answer : B
Explanation: “Attention Is All You Need” is arguably the most impactful deep learning paper ever written. Before the paper: sequence-to-sequence models used LSTM/GRU encoders and decoders with attention. Sequential processing: RNNs process one token at a time → slow to train on modern parallel hardware. The breakthrough: proposed replacing all recurrence with self-attention. Self-attention: every token directly attends to every other token — O(1) steps vs O(n) for RNNs. Enables full parallelization — train 10× faster. Better long-range dependencies — no information bottleneck across many steps. Architecture components: Multi-head self-attention, Feed-forward layers, Residual connections, Layer normalization, Positional encoding. Results: Outperformed best LSTM-based models on WMT translation benchmarks. Trained in a fraction of the time. What it enabled: BERT (2018): bidirectional pre-training. GPT-1/2/3/4: scaling language generation. T5, BART: encoder-decoder transformers. ViT: apply Transformers to images. Modern LLMs: all built on Transformer. Citation count: one of the most cited papers in AI history (~90,000+ citations). The Transformer has become the fundamental building block of modern AI — not just NLP but also vision, audio, video, robotics, and scientific computing.

75. What is the difference between supervised, unsupervised, and self-supervised learning in deep learning?

  1. Supervised uses teachers; unsupervised uses no teachers; self-supervised teaches itself using mirrors
  2. Supervised learning trains on labeled (x, y) pairs; unsupervised learning discovers patterns in unlabeled data (x only); self-supervised learning creates its own labels from the structure of unlabeled data — enabling pre-training on massive unlabeled datasets like the internet
  3. Self-supervised learning is the same as semi-supervised learning with a different name
  4. Unsupervised deep learning is only applicable to clustering tasks and nothing else

Answer : B
Explanation: Learning paradigms in deep learning: Supervised Learning: training data: (input, label) pairs. Model learns mapping from input to label. Examples: image classification (image → class), sentiment analysis (text → positive/negative). Limitations: labeled data is expensive and time-consuming to collect. Unsupervised Learning: training data: inputs only (no labels). Discovers structure in data without guidance. Examples: clustering (k-means, deep clustering), dimensionality reduction (autoencoders, t-SNE), density estimation (VAEs, normalizing flows). Self-Supervised Learning: the most exciting modern paradigm. Creates supervisory signals automatically from the structure of data (no human labeling needed). Examples: BERT: predict masked tokens in text. GPT: predict the next token. SimCLR, MoCo: predict that two augmented views of the same image are similar. MAE (Masked Autoencoder): predict masked image patches. Why it’s transformative: enables learning from web-scale unlabeled data (billions of images, trillions of text tokens). Models (foundation models) learn rich representations transferable to many downstream tasks. BERT, GPT, CLIP, SAM are all pre-trained self-supervisedly. Semi-supervised: small labeled + large unlabeled dataset — combines both supervised and unsupervised signals.

Machine Learning MCQ Questions And Answers

76. What is the concept of a learning rate schedule in deep learning?

  1. A learning rate schedule is a timetable showing when each layer learns during the training process
  2. A learning rate schedule dynamically adjusts the learning rate during training — typically starting high (fast learning) and decreasing over time or in response to training dynamics, improving convergence stability and final model performance
  3. Learning rate schedules assign different learning rates to different layers for the entire training run
  4. Learning rate schedules are only needed for RNN training and are unnecessary for CNNs or Transformers

Answer : B
Explanation: Learning Rate (LR) Scheduling is crucial for training deep learning models effectively. Why schedules matter: too high LR: overshoots minimum, training diverges. Too low LR: slow convergence, may get stuck. Varying LR during training often achieves better results than any fixed LR. Common schedules: Step Decay: reduce LR by a factor every N epochs. LR × 0.1 every 30 epochs. Exponential Decay: LR × γ each epoch. Cosine Annealing: LR follows a cosine curve from max to min. Widely used — smooth reduction with potential warm restarts. ReduceLROnPlateau: reduce LR when validation loss stops improving. Warm Restarts (SGDR): periodically reset LR to its maximum value — allows escaping local minima. Warmup: start with very low LR, gradually increase to target LR. Essential for Transformers — prevents early instability. OneCycleLR: single cycle of LR from low → peak → low. Often achieves best results in fewer epochs. Cyclical Learning Rates (CLR): oscillate between min and max LR — helps escape saddle points. Implementation: PyTorch: torch.optim.lr_scheduler (StepLR, CosineAnnealingLR, OneCycleLR). TensorFlow/Keras: keras.optimizers.schedules. Transformers: linear warmup + cosine decay is the standard schedule.

77. What is the concept of weight initialization in deep learning?

  1. Weight initialization is the process of setting initial network weights by copying from a similar trained model
  2. Weight initialization sets the starting values of neural network weights before training — critically affecting training speed and convergence; poor initialization can cause vanishing or exploding gradients from the very first step
  3. Weight initialization sets the maximum values that weights can reach during training as an upper bound
  4. All deep learning networks must start with weights initialized to exactly zero for stable training

Answer : B
Explanation: Good weight initialization is crucial — it determines whether training starts in a stable state. Why zero initialization fails: all neurons compute the same output → same gradients → all neurons stay identical throughout training (symmetry problem). Why large random initialization fails: activations explode or vanish immediately. Common initialization methods: Xavier/Glorot Initialization (for sigmoid/tanh): variance = 2 / (fan_in + fan_out). Keeps activation variance constant across layers. Default in TensorFlow/Keras for dense layers. He/Kaiming Initialization (for ReLU): variance = 2 / fan_in. Accounts for ReLU zeroing negative values. Default in PyTorch for Conv layers. LeCun Initialization (for SELU): variance = 1 / fan_in. Used with self-normalizing networks. Orthogonal Initialization: initializes weight matrices to be orthogonal — useful for RNNs. Glorot Uniform/Normal: same principles, different distribution shapes. Modern practice: with batch normalization, sensitivity to initialization decreases significantly — BN stabilizes activations regardless of initial weights. Without BN: proper initialization is critical. Pre-trained weights: for transfer learning, use pre-trained weights — no initialization needed.

78. What is contrastive learning in deep learning?

  1. Contrastive learning is a technique that compares a student model’s output to the teacher model’s output
  2. Contrastive learning is a self-supervised learning approach that trains models to pull together representations of similar samples (positive pairs) and push apart representations of dissimilar samples (negative pairs) — without requiring labeled data
  3. Contrastive learning uses contrasting (opposite) loss functions to create adversarial training
  4. Contrastive learning is only applicable to text data and cannot be applied to image or audio inputs

Answer : B
Explanation: Contrastive Learning enables powerful self-supervised representation learning. Core idea: positive pairs (similar samples) → similar representations. Negative pairs (dissimilar samples) → different representations. No labels needed — similarity defined by data augmentation or structure. Key frameworks: SimCLR (Chen et al., Google, 2020): take one image, apply two random augmentations → positive pair. Large batch provides many negative pairs. NT-Xent (Normalized Temperature-scaled Cross Entropy) loss. MoCo (He et al., Facebook): momentum-updated memory bank of negative examples. More memory efficient than SimCLR. BYOL (Bootstrap Your Own Latent): no negative pairs — online and target network. Learns without explicitly pushing negatives apart. InfoNCE Loss: L = -log[exp(sim(z1,z2)/τ) / Σk exp(sim(z1,zk)/τ)]. CLIP (OpenAI, 2021): contrastive between image and text pairs. Text encoder and image encoder trained to align matching (image, text) pairs. Enables zero-shot image classification. Results: SimCLR pre-trained on ImageNet (no labels) + linear classifier ≈ supervised ResNet-50 performance. Foundation for modern vision-language models (CLIP, ALIGN). Active research area in NLP, graph learning, time series, and multi-modal learning.

79. What is knowledge distillation in deep learning?

  1. Knowledge distillation is a technique for extracting domain expert knowledge and encoding it into a model
  2. Knowledge distillation is a model compression technique where a small “student” network is trained to mimic the behavior of a large, accurate “teacher” network — using the teacher’s soft probability outputs as additional training targets, transferring knowledge more efficiently than training from scratch
  3. Knowledge distillation condenses multiple training datasets into a single representative dataset
  4. Knowledge distillation requires the teacher and student to have identical architectures for knowledge transfer

Answer : B
Explanation: Knowledge Distillation (Hinton et al., 2015) efficiently transfers knowledge from a large model to a small model. Why soft labels help: hard labels: [0, 0, 1, 0] (one-hot) — minimal information. Teacher’s soft labels: [0.02, 0.05, 0.85, 0.08] — contains rich information about similarity between classes (e.g., a cat image scores 0.05 for small dogs — capturing dog-cat similarity). This “dark knowledge” trains the student more effectively. Temperature scaling: T > 1 in softmax → softer probabilities → more information in soft labels. Process: Train a large, accurate teacher model. Generate teacher’s soft predictions on training data (with temperature T). Train student to minimize: αL_CE(hard labels) + (1-α)L_KD(soft labels from teacher). Student matches teacher’s behavior, not just the final answer. Results: DistilBERT: 40% smaller BERT, 97% performance, 60% faster. DistilGPT-2: 6× compressed GPT-2. MobileNet: distilled from large CNNs for mobile deployment. Applications: Deploy large models on edge devices (smartphones, IoT), Reduce inference latency, Save memory. Online distillation: multiple models teach each other simultaneously. Self-distillation: model distills into a smaller copy of itself across layers.

80. What is the concept of neural architecture search (NAS) in deep learning?

  1. NAS is a web scraping technique that searches the internet for neural network architecture papers
  2. Neural Architecture Search (NAS) is an automated machine learning technique that uses search algorithms to discover optimal neural network architectures for a given task — replacing manual architectural engineering with algorithmic search over a design space
  3. NAS searches for the best hyperparameters like learning rate and batch size for a fixed architecture
  4. Neural Architecture Search only works for image classification and cannot be applied to NLP tasks

Answer : B
Explanation: Neural Architecture Search (NAS) automates the design of neural network architectures — replacing years of human intuition with algorithmic search. NAS components: Search Space: defines possible architectures (layer types, connections, sizes, skip connections). Search Strategy: how to explore the search space. RL-based: controller RNN generates architectures, trained using validation accuracy as reward (Zoph & Le, 2016 — took 800 GPUs for 28 days). Evolutionary: evolve architectures through mutation and selection. Gradient-based (DARTS): make architecture decisions differentiable — backpropagate through architecture choices. One-shot/Weight sharing: train one supernetwork containing all possible architectures. Performance Estimation: predict performance without full training. Famous NAS results: NASNet: outperformed hand-designed architectures on ImageNet. EfficientNet (AutoML): compound scaling of depth/width/resolution — best accuracy/efficiency at time of release. EfficientDet: NAS-designed object detector. MobileNetV3: hardware-aware NAS for mobile deployment. AmoebaNet: evolutionary NAS. Modern trend: DARTS and its variants make NAS feasible on a single GPU. Hardware-aware NAS: optimize for specific hardware (mobile CPU, GPU, neural processing unit). NAS is a cornerstone of AutoML — making deep learning more accessible by automating architecture design.

Pages: 1 2 3 4 5 6 7 8 9 10