Deep Learning MCQ Questions And Answers

Deep Learning MCQ Questions And Answers

51. What is BERT in deep learning and what makes it different from GPT?

  1. BERT and GPT are identical models trained by the same team using the same methodology
  2. BERT (Google) is an encoder-only Transformer pre-trained with masked language modeling — understanding context from both directions, best for classification and comprehension tasks; GPT (OpenAI) is a decoder-only Transformer pre-trained with next-token prediction — best for text generation tasks
  3. BERT is for image understanding; GPT is for text understanding — they handle different data types
  4. GPT is a smaller, less powerful version of BERT designed for edge device deployment

Answer : B
Explanation: Both BERT and GPT use the Transformer architecture but differ in design and purpose. BERT (Bidirectional Encoder Representations from Transformers, Google, 2018): Encoder-only Transformer. Pre-training: Masked Language Modeling (MLM) — predict randomly masked tokens using bidirectional context. + Next Sentence Prediction (NSP). Bidirectional — understands each word using both left and right context. Best for: text classification, NER, question answering, semantic similarity, text pair tasks. Fine-tuned for specific tasks. GPT (Generative Pre-trained Transformer, OpenAI): Decoder-only Transformer. Pre-training: autoregressive — predict the next token given previous tokens. Unidirectional (causal) — can only attend to previous tokens. Best for: text generation, completion, summarization, translation, code generation, conversational AI. ChatGPT = GPT-3.5/GPT-4 + RLHF. Modern models: Encoder-only: BERT, RoBERTa, ALBERT, DeBERTa. Decoder-only: GPT series, LLaMA, Gemini, Mistral, Claude. Encoder-Decoder: T5, BART, mT5 — best of both for seq2seq tasks.

52. What is transfer learning in deep learning?

  1. Transfer learning is a technique for transferring trained model files between different computers
  2. Transfer learning is the practice of reusing a model trained on one task as the starting point for a new related task — dramatically reducing the training time and data requirements for the new task by leveraging already learned features
  3. Transfer learning transfers gradients from one model to another to speed up convergence
  4. Transfer learning is only possible when both the source and target tasks use the same dataset

Answer : B
Explanation: Transfer Learning is one of the most important practical techniques in deep learning. How it works: Start with a pre-trained model (e.g., ResNet-50 trained on ImageNet with 1.4M images). Remove the final classification layer. Add new layers for the target task. Fine-tune either: frozen base (only train new layers — faster, less data needed) or full fine-tuning (update all weights with small learning rate — more powerful). Common in CV: ImageNet pre-trained models (ResNet, EfficientNet, ViT) transferred to medical imaging, satellite imagery, industrial defect detection. Common in NLP: BERT, GPT pre-trained on massive text → fine-tuned for sentiment analysis, QA, summarization. Why it works: early layers of neural networks learn general features (edges, textures in CV; syntax, semantics in NLP) that transfer across domains. Later layers learn task-specific features. Foundation Models: large models pre-trained on internet-scale data → fine-tuned for specific tasks. BERT, GPT-4, CLIP, SAM are examples. Few-shot learning: with foundation models, sometimes just providing a few examples in the prompt (no gradient updates) is sufficient.

53. What is the difference between a convolutional layer and a pooling layer in a CNN?

  1. Convolutional layers reduce spatial dimensions; pooling layers detect features using filters
  2. A convolutional layer applies learnable filters to detect features — producing feature maps with learned weights; a pooling layer performs a fixed downsampling operation (max or average) — reducing spatial dimensions without learnable parameters
  3. Pooling layers contain more parameters than convolutional layers due to their complex downsampling
  4. Convolutional layers are only used in the first layer of CNNs; pooling layers are used throughout

Answer : B
Explanation: Both are fundamental CNN components with distinct roles. Convolutional Layer: applies learnable filters (kernels) across the input. Filter slides over the input, computing dot products at each position. Parameters: filter_size × filter_size × in_channels × out_channels + bias. Learns to detect features (edges, textures, shapes, objects). Maintains spatial information. Output size: (input – filter + 2×padding)/stride + 1. Pooling Layer: fixed downsampling operation (no learnable parameters). Max Pooling: takes maximum value in each window → preserves dominant features, provides translation invariance. Average Pooling: takes mean → smoother summary. Global Average Pooling (GAP): averages entire feature map → one value per channel — replaces large FC layers in modern CNNs. Purpose: reduce spatial dimensions → reduces computation. Provides some translation invariance. Controls overfitting. Modern trend: strided convolutions (stride > 1) increasingly replace pooling — the convolution learns its own downsampling, potentially more powerful. EfficientNet, ResNet v2 use strided convolutions. Classic CNN pattern: Input → [Conv → ReLU → Pool] × N → Flatten/GAP → Dense → Softmax.

54. What is data augmentation in deep learning and why is it important?

  1. Data augmentation is a technique for purchasing and downloading more training data from the internet
  2. Data augmentation artificially expands the training dataset by applying realistic transformations to existing examples — improving model generalization and reducing overfitting without collecting new data
  3. Data augmentation is a post-training technique applied after the model has finished training
  4. Data augmentation is only applicable to text data and cannot be used for image training

Answer : B
Explanation: Data Augmentation is one of the most cost-effective ways to improve deep learning models. Image Augmentation: Geometric: Random horizontal/vertical flip, Random rotation, Random crop and resize, Shear, Perspective transform. Color: Color jitter (brightness, contrast, saturation, hue), Gaussian blur, Grayscale conversion. Cutout/Random erasing: mask random regions. Mixup: blend two images and interpolate their labels. CutMix: cut-paste region from another image. AutoAugment/RandAugment: automatically learn best augmentation policies. Text Augmentation: Synonym replacement, Random word insertion/deletion/swap, Back-translation (English → French → English). Audio Augmentation: Time stretching, Pitch shifting, Adding background noise, SpecAugment (mask time/frequency segments). Advanced: Test-time augmentation (TTA) — apply augmentation at inference time and average predictions. Why augmentation works: teaches the model invariances — a cat is still a cat when flipped, rotated, or color-shifted. Prevents memorizing specific visual patterns in training images. Libraries: torchvision.transforms (PyTorch), Albumentations (fast, comprehensive), imgaug.

55. What is the concept of epochs, batch size, and iterations in deep learning training?

  1. Epoch and iteration are the same thing; batch size is the number of training examples in total
  2. An epoch is one complete pass through the entire training dataset; batch size is the number of samples processed together in one forward/backward pass; iterations (steps) is the number of batches per epoch — equal to dataset size divided by batch size
  3. Batch size refers to the number of layers in the network; an epoch is one update to the weights
  4. More epochs always lead to better models regardless of batch size and learning rate settings

Answer : B
Explanation: Understanding these terms is essential for implementing deep learning training loops. Epoch: one complete pass through the entire training dataset. Multiple epochs needed — typically 10-200+ for deep learning tasks. More epochs → more exposure to data (risk of overfitting). Batch Size: number of training examples processed together before updating weights. Small batch (8-32): noisier gradient estimates, potentially better generalization, less GPU memory needed. Large batch (256-4096): stable gradients, parallelization efficient, faster per epoch but may converge to sharp minima. Mini-batch SGD (most common): balance between speed and gradient quality. Iterations (steps): iterations per epoch = dataset size / batch size. Example: 50,000 training images, batch size 100 → 500 iterations per epoch. If training for 100 epochs → 50,000 total weight updates. Important hyperparameters: Learning rate: most critical. Batch size: affects gradient quality. Number of epochs: combined with early stopping. Learning rate schedule: adjust LR over training. Rule of thumb: linear scaling rule — when increasing batch size k×, multiply learning rate by k. Used in distributed training.

56. What is the difference between a loss function and an optimizer in deep learning?

  1. A loss function updates the weights; an optimizer measures how wrong predictions are
  2. A loss function measures the difference between predictions and true values (what to minimize); an optimizer uses the gradients from the loss function to update the model’s weights (how to minimize the loss)
  3. Loss functions are used only during testing; optimizers are used only during training
  4. The loss function and optimizer are the same component with different names in different frameworks

Answer : B
Explanation: Loss Function: quantifies how wrong the model is — the objective to minimize. Common loss functions: Mean Squared Error (MSE): for regression. L = (1/n)Σ(y – ŷ)². Binary Cross-Entropy: for binary classification. L = -(y·log(ŷ) + (1-y)·log(1-ŷ)). Categorical Cross-Entropy: for multi-class classification with softmax output. Huber Loss: robust regression (combines MSE and MAE). Contrastive Loss, Triplet Loss: for metric learning. Focal Loss: for class imbalance (used in RetinaNet). Optimizer: algorithm that uses gradients to update weights toward the minimum loss. SGD (Stochastic Gradient Descent): w = w – α×∂L/∂w. Basic, noisy. Momentum: accelerates in consistent directions. Adam (Adaptive Moment Estimation): adaptive learning rates per parameter. Default for most deep learning — α=0.001, β₁=0.9, β₂=0.999. AdamW: Adam + weight decay decoupled — better generalization for Transformers. RMSprop: good for RNNs. Adagrad, Adadelta. Choosing: Classification → Cross-Entropy loss + Adam optimizer. Regression → MSE/Huber loss + Adam or SGD+Momentum.

57. What is deep learning’s role in computer vision?

  1. Deep learning in computer vision only classifies pre-labeled images and cannot detect new objects
  2. Deep learning has revolutionized computer vision — enabling machines to surpass human-level performance on tasks including image classification, object detection, semantic segmentation, face recognition, medical image analysis, and visual question answering through learned hierarchical representations
  3. Deep learning’s role in computer vision is limited to grayscale image processing only
  4. Computer vision using deep learning requires the objects of interest to be manually highlighted first

Answer : B
Explanation: Deep learning transformed computer vision starting with AlexNet’s 2012 ImageNet victory. Key CV tasks and deep learning approaches: Image Classification: ResNet, EfficientNet, Vision Transformer (ViT) — single label per image. AlexNet (2012) → VGG → GoogLeNet → ResNet → DenseNet → EfficientNet → ViT → ConvNeXt. Object Detection: YOLO (real-time), Faster R-CNN (high accuracy), DETR (transformer-based). Semantic Segmentation: U-Net (medical), DeepLab, SegFormer. Instance Segmentation: Mask R-CNN. Face Recognition: FaceNet, ArcFace — used in iPhone Face ID, Aadhaar, law enforcement. Optical Character Recognition (OCR): TrOCR, Tesseract with deep learning. Image Generation: GANs (StyleGAN), Diffusion Models (Stable Diffusion, DALL-E, Midjourney). Visual Question Answering: CLIP, Flamingo, LLaVA, GPT-4V. Pose Estimation: MediaPipe, OpenPose. Medical Imaging: pathology slide analysis, radiology (chest X-ray, CT, MRI), dermatology, ophthalmology. Key milestone: AlexNet (2012) achieved 15.3% top-5 error on ImageNet vs 26.2% for the runner-up — demonstrating deep learning’s superiority over traditional CV methods.

58. What is the convolution operation in deep learning?

  1. Convolution in deep learning is the same mathematical operation as the cross-correlation in signal processing
  2. Convolution is a mathematical operation where a learnable filter (kernel) slides over the input, computing element-wise products and summing them — detecting specific patterns (edges, textures) at every spatial location through weight sharing
  3. Convolution in deep learning refers to the mixing of multiple training datasets before processing
  4. Convolution is only applied to grayscale images and requires separate processing for color images

Answer : B
Explanation: Convolution is the fundamental operation in CNNs. Mathematical operation: (f * g)[n] = Σ f[m] × g[n-m]. In deep learning, technically cross-correlation is used (no flipping), but called convolution. How it works: a filter (e.g., 3×3×3 for RGB image) slides across the input image. At each position, compute element-wise product between filter and the input patch, then sum → one value in output feature map. For a 6×6 image with 3×3 filter, stride 1: output is 4×4 feature map. With padding=1: output is 6×6 (same size). Multiple filters: each filter produces one feature map — a layer with 64 filters produces a 64-channel output. Parameter sharing: one set of weights shared across all positions — exploits spatial stationarity. Feature map value = filter’s response at that location. What different filters detect: Early layers: edges (horizontal, vertical, diagonal), corners. Mid layers: textures, patterns, parts of objects. Deep layers: complex objects, semantic concepts. Dilated convolution: gaps between filter elements → larger receptive field with same parameters. Depthwise separable convolution: factorize into depthwise + pointwise → fewer parameters (used in MobileNet).

59. What is the difference between image classification and image recognition?

  1. Image classification assigns multiple labels per image; image recognition assigns only one label
  2. Image classification assigns one or more predefined class labels to an entire image; image recognition is a broader term encompassing classification plus also identifying specific entities like faces, text, or landmarks within images
  3. Image recognition is the older term and has been completely replaced by image classification in DL
  4. Image classification works on grayscale images only; image recognition requires color images

Answer : B
Explanation: Image Classification: input → single image. Output → one class label (or probabilities for multiple classes). “Is this a cat, dog, or bird?” Binary: is this a tumor or not? Multi-class: which of 1000 ImageNet categories? Multi-label: does this image contain a cat AND a dog? Models: ResNet, VGG, EfficientNet, Vision Transformer. Benchmark: ImageNet (1.4M images, 1000 classes) — classification error reduced from 26% (2011) to 1.6% (2021). Image Recognition (broader): includes classification + identification of specific instances. Face recognition: which specific person is this? Optical Character Recognition (OCR): read text in images. Logo recognition, landmark recognition. Image understanding tasks beyond labeling. Applications: Image classification: medical diagnosis (disease type), product categorization, content moderation. Image recognition: access control (face ID), document digitization, visual search (Google Lens, Pinterest Lens). Key models: Classification: ResNet-50 (25M params, ~76% top-1 accuracy), EfficientNet-B7 (66M params, ~84%), ViT-L (307M params, ~88%). Classification is the foundation — once learned, it transfers to detection, segmentation, and generation tasks through fine-tuning.

60. What is deep learning’s role in speech recognition?

  1. Deep learning in speech recognition only converts speech to text in English with no other languages
  2. Deep learning has transformed speech recognition — enabling end-to-end models that directly convert audio waveforms to text, achieving human-level accuracy on many benchmarks through architectures like Deep Speech, Wav2Vec, and Whisper
  3. Speech recognition using deep learning requires manually transcribed data for every 10 seconds of audio
  4. Deep learning only improved speech recognition speed, not its accuracy over traditional methods

Answer : B
Explanation: Speech Recognition (Automatic Speech Recognition / ASR) evolution with deep learning: Pre-deep learning: GMM-HMM (Gaussian Mixture Model — Hidden Markov Model) systems required manual feature engineering (MFCC features). Deep learning era: Deep Speech (Baidu, 2014): RNN/CNN-based, CTC loss — end-to-end training without alignment. Wav2Letter (Facebook): CNN-based end-to-end. Listen, Attend and Spell (Google): sequence-to-sequence with attention. Transformer-based: Conformer (CNN + Transformer) — state-of-the-art on many benchmarks. Wav2Vec 2.0 (Facebook, 2020): self-supervised pre-training on unlabeled audio → fine-tune on small labeled dataset. Whisper (OpenAI, 2022): trained on 680K hours of multilingual audio — robust, multilingual, multi-task (transcription + translation). CTC (Connectionist Temporal Classification) loss: allows training without explicit alignment between audio frames and text tokens. Applications: Virtual assistants (Siri, Alexa, Google Assistant), Live captioning, Medical transcription, Call center automation, Voice search. Remaining challenges: accented speech, noisy environments, code-switching (mixing languages), rare words, proper nouns.

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