Neural Networks MCQ Questions And Answers

Neural Networks MCQ Questions And Answers

41. What is a Convolutional Neural Network (CNN) and how does it work?

  1. A type of RNN specifically designed to process convolutional (circular) data structures
  2. A deep learning architecture designed for processing grid-like data (images) that uses learnable convolutional filters to automatically detect spatial features — consisting of convolutional layers, pooling layers, and fully connected layers
  3. A neural network where neurons are arranged in a convex shape to optimize connectivity
  4. A network that convolves multiple training datasets to create composite feature maps

Answer : B
Explanation: A CNN processes images by automatically learning hierarchical features. Key components: Convolutional Layer — applies learnable filters (kernels) across the input, computing dot products at each position to produce feature maps that detect edges, textures, shapes. Activation Layer — typically ReLU applied element-wise after convolution. Pooling Layer — reduces spatial dimensions (Max Pooling takes the maximum value in each region, reducing size and computation while retaining dominant features). Fully Connected (Dense) Layer — flattens feature maps and makes the final classification decision. How CNNs learn: early layers detect simple features (edges, colors), middle layers combine these into shapes (eyes, wheels), deep layers recognize complex patterns (faces, cars). Parameter sharing (one filter applied across the entire image) dramatically reduces parameters compared to fully connected networks. Famous CNN architectures: LeNet, AlexNet, VGG, ResNet, InceptionNet, EfficientNet.

42. What is a pooling layer in a CNN and what are its types?

  1. A layer that pools (combines) multiple CNNs together into an ensemble model
  2. A layer that down-samples feature maps by summarizing regions — reducing spatial dimensions, computation, and parameters while retaining important features and providing translation invariance
  3. A layer that pools all training examples into a single batch for faster convolution
  4. A layer that collects (pools) activation values from multiple filters into one feature map

Answer : B
Explanation: Pooling reduces feature map dimensions after convolution. Types: Max Pooling — takes the maximum value in each pooling window (most common). Preserves the most prominent feature in each region. Provides translation invariance (slight shifts in the image don’t change the output). Average Pooling — takes the average value in each window. Global Average Pooling (GAP) — averages each entire feature map into a single value, often used before the classification layer in modern CNNs (replaces large fully connected layers). Global Max Pooling — takes the maximum across each entire feature map. Typical configuration: 2×2 max pooling with stride 2 halves both height and width (reducing spatial dimensions by 4×). Modern architectures (ResNet, EfficientNet) sometimes use strided convolutions instead of pooling for better gradient flow. Pooling provides no learnable parameters — it is a fixed operation.

43. What is a Recurrent Neural Network (RNN) and what makes it different from a feedforward network?

  1. An RNN is a type of CNN that processes images in a recursive circular pattern
  2. An RNN is a neural network designed for sequential data where the output from a previous time step is fed back as input to the current time step — allowing the network to maintain a hidden state (memory) across the sequence
  3. A neural network that recurrently retrains on the same dataset to improve accuracy
  4. A network where neurons recursively call each other to process structured data

Answer : B
Explanation: RNNs process sequences by maintaining a hidden state h that captures information from previous time steps. At each step t: hₜ = tanh(Wₕhₜ₋₁ + Wₓxₜ + b). This recurrent connection creates a “memory” of past inputs. Unlike feedforward networks (process one input independently), RNNs share weights across all time steps (parameter efficiency). Key applications: language modeling, machine translation, speech recognition, text generation, time series forecasting, sentiment analysis. Key limitations: Vanishing/exploding gradients over long sequences (backpropagation through time — BPTT — multiplies gradients across many steps). Difficulty capturing long-range dependencies (what was mentioned 50 words ago may not influence the hidden state much now). These limitations motivated the development of LSTM and GRU, which are dramatically more effective RNN variants.

44. What is LSTM (Long Short-Term Memory) and why was it developed?

  1. A type of memory chip designed for storing long-term neural network weights
  2. A specialized RNN architecture that uses three gates (input, forget, output) and a cell state to selectively remember and forget information over long sequences — solving the vanishing gradient problem of standard RNNs
  3. A network that limits short-term memory to improve long-term computational efficiency
  4. A long-term storage module attached to feedforward networks for persistent memory

Answer : B
Explanation: LSTM (Long Short-Term Memory), developed by Hochreiter and Schmidhuber (1997), solves the vanishing gradient problem in RNNs. Key components: Cell State (Cₜ) — the long-term memory highway. Gradients flow through it relatively unchanged. Forget Gate — σ(Wf·[hₜ₋₁, xₜ] + bf) — decides what to erase from cell state. Input Gate — σ(Wi·[hₜ₋₁, xₜ] + bi) — decides what new information to add to cell state. Output Gate — σ(Wo·[hₜ₋₁, xₜ] + bo) — decides what part of cell state to expose as hidden state. The gates use sigmoid (0 to 1 — open/close gate) and tanh (cell state values). Applications: machine translation (seq2seq with LSTM encoder-decoder), speech recognition, text generation, music generation, video captioning. LSTMs remain relevant but Transformers have largely replaced them for NLP tasks where parallel processing is advantageous.

45. What is GRU (Gated Recurrent Unit) and how does it differ from LSTM?

  1. A GPU-based processing unit used to accelerate RNN training on graphics hardware
  2. A simplified RNN variant with two gates (update and reset) instead of LSTM’s three gates — achieving comparable performance to LSTM with fewer parameters and faster training
  3. A recurrent unit that groups (gates) multiple RNN layers into a single processing block
  4. A gated architecture used exclusively in transformer attention mechanisms

Answer : B
Explanation: GRU (Gated Recurrent Unit), introduced by Cho et al. (2014), simplifies LSTM while maintaining comparable performance. Two gates: Update Gate — combines LSTM’s forget and input gates into one. Controls how much of the previous hidden state to keep vs. how much new information to incorporate. Reset Gate — controls how much of the previous hidden state to forget when computing the new candidate hidden state. Key differences from LSTM: No separate cell state — GRU uses only the hidden state. Fewer parameters (2 gates vs 3 gates + cell state) — faster training and inference. Less prone to overfitting on smaller datasets. When to use: GRU for shorter sequences, limited compute, or small datasets. LSTM for complex long-range dependencies, large datasets. In practice, performance is often similar — GRU is increasingly preferred for its simplicity. Both have been largely superseded by Transformers for most NLP tasks.

46. What is the difference between a feedforward neural network and a recurrent neural network?

  1. Feedforward networks use forward propagation; RNNs use only backward propagation
  2. In a feedforward network, information flows only forward (input → hidden → output) with no cycles — suitable for fixed-size inputs; in an RNN, connections form directed cycles allowing past outputs to influence current processing — suitable for sequential data
  3. Feedforward networks are more accurate than RNNs for all types of machine learning tasks
  4. RNNs process all inputs simultaneously while feedforward networks process one at a time

Answer : B
Explanation: Feedforward Neural Network (FNN/MLP): data flows in one direction only — from input layer through hidden layers to output. No loops, no memory of previous inputs. Each input is processed independently. Suitable for: tabular data, image classification (with CNN layers), regression. Cannot handle variable-length sequences or temporal dependencies. Recurrent Neural Network (RNN): has directed cycles — output from previous time steps feeds back as input to current time step. Maintains hidden state (memory). Suitable for: sequential data (text, speech, time series), variable-length inputs, tasks requiring temporal context. Trained with Backpropagation Through Time (BPTT) — unrolling the network across time steps and applying standard backpropagation. The choice between FNN and RNN depends on whether the task requires memory of past inputs — if yes, use RNN (or Transformer); if no, use FNN.

47. What is the multilayer perceptron (MLP) and how does it overcome the limitations of a single-layer perceptron?

  1. An MLP is a perceptron that processes data through multiple parallel streams simultaneously
  2. An MLP is a fully connected feedforward neural network with one or more hidden layers and non-linear activation functions — enabling it to learn non-linear decision boundaries and solve problems like XOR that are impossible for a single-layer perceptron
  3. A perceptron architecture that uses multiple datasets during training for better generalization
  4. A neural network where each layer contains multiple parallel perceptrons with different weights

Answer : B
Explanation: The Single-Layer Perceptron can only learn linearly separable problems — it cannot solve XOR. The MLP (Multilayer Perceptron) solves this by: Adding hidden layers — enabling the network to learn hierarchical feature representations. Using non-linear activation functions (sigmoid, tanh, ReLU) — without non-linearity, multiple linear layers are equivalent to one linear layer, providing no benefit from depth. Trained with backpropagation + gradient descent. Architecture: Input Layer → Hidden Layers (1 or more) → Output Layer. Each neuron in a layer is connected to every neuron in the next layer (fully connected / dense). MLPs are universal function approximators (Universal Approximation Theorem) — with enough hidden neurons, an MLP with one hidden layer can approximate any continuous function. MLPs are the foundation of all modern deep learning — CNNs and RNNs use MLP components for their fully connected classification layers.

48. What is dropout regularization in neural networks?

  1. A technique that removes underperforming neurons permanently from the neural network
  2. A regularization technique that randomly deactivates (sets to zero) a fraction of neurons during each training iteration — preventing co-adaptation and overfitting by forcing the network to learn redundant representations
  3. A method that drops the learning rate when training accuracy stops improving
  4. A technique that removes outlier training examples that cause overfitting

Answer : B
Explanation: Dropout, introduced by Srivastava et al. (2014), is one of the most effective regularization techniques for neural networks. During training: each neuron is independently set to zero with probability p (the dropout rate — typically 0.2 to 0.5). The remaining neurons’ outputs are scaled by 1/(1-p) to maintain expected values. During inference: all neurons are active (no dropout), but outputs are scaled appropriately. Why it works: Prevents co-adaptation (neurons can’t rely on specific other neurons always being present). Acts like training an ensemble of 2^n different network architectures simultaneously. Forces learning of more robust, distributed representations. Connection to bagging: dropout can be viewed as an extreme form of bagging (as noted in Q22 of your existing questions). Applied to: fully connected layers (most common), sometimes to convolutional layers (SpatialDropout). Not needed with batch normalization (which provides implicit regularization).

49. What is batch normalization in neural networks and why is it used?

  1. A technique that normalizes the size of training batches to a standard number of examples
  2. A technique that normalizes the activations of each layer within a mini-batch to have zero mean and unit variance — stabilizing and accelerating training, enabling higher learning rates, and acting as a mild regularizer
  3. A preprocessing step that normalizes all training data before feeding it to the network
  4. A method of normalizing weight updates to prevent the exploding gradient problem

Answer : B
Explanation: Batch Normalization (BN), introduced by Ioffe and Szegedy (2015), normalizes layer activations: for each mini-batch, compute mean and variance, normalize, then apply learnable scale (γ) and shift (β) parameters. Benefits: Reduces Internal Covariate Shift — prevents the distribution of layer inputs from changing during training (making each layer’s job easier). Enables higher learning rates — gradients are more stable. Reduces sensitivity to weight initialization — less need for careful initialization. Acts as regularizer — reduces the need for dropout in many architectures. Faster convergence — typically converges in fewer epochs. Where to place BN: typically after the linear transformation and before the activation function (or sometimes after). Limitation: performance degrades with very small batch sizes. Alternative: Layer Normalization (used in Transformers — normalizes across features for a single example, works with any batch size). BN is now standard in almost all deep learning architectures.

50. What is the difference between classification and regression in neural networks?

  1. Classification uses regression outputs to classify data; regression classifies numerical ranges
  2. Classification predicts discrete class labels (which category) using cross-entropy loss and sigmoid/softmax output; regression predicts continuous numerical values using MSE/MAE loss and linear output
  3. Regression neural networks are always deeper than classification networks
  4. Classification and regression use identical network architectures with different input data types

Answer : B
Explanation: Classification predicts which class/category an input belongs to. Binary Classification: output layer — 1 neuron with Sigmoid activation, loss function — Binary Cross-Entropy. Multi-class Classification: output layer — N neurons (one per class) with Softmax activation, loss function — Categorical Cross-Entropy. Regression predicts a continuous numerical value. Output layer — 1 neuron (or N for multi-output), Linear activation (no activation, or identity), loss function — Mean Squared Error (MSE), Mean Absolute Error (MAE). The rest of the architecture (hidden layers, activations, batch normalization, dropout) is often identical between classification and regression networks. Example: predicting house price → regression (output: $350,000). Predicting whether email is spam → binary classification (output: 0.92 = 92% probability of spam). Predicting handwritten digit → multi-class classification (output: probabilities for 0-9).

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