Neural Networks Fundamentals Assessment — Questions and Answers

Question 1: What is the Leaky ReLU and how does it differ from standard ReLU?

It adds a learned bias to the ReLU output
It doubles the gradient for positive inputs
It applies a sigmoid function for negative inputs
It allows a small non-zero gradient for negative inputs instead of outputting zero (Correct answer)

Correct answer: It allows a small non-zero gradient for negative inputs instead of outputting zero

Leaky ReLU outputs a small slope (e.g., 0.01x) for negative inputs rather than zero, preventing neurons from dying and ensuring some gradient flow for negative activations.

Question 2: What is second-order optimization and why is it rarely used in deep learning?

Training with two optimizers in parallel
Optimization using two loss functions simultaneously
Using second derivatives only for the output layer
Using curvature information (Hessian) for updates, but too computationally expensive for large networks (Correct answer)

Correct answer: Using curvature information (Hessian) for updates, but too computationally expensive for large networks

Second-order methods use the Hessian matrix of second derivatives for more accurate updates, but computing and inverting the Hessian is computationally prohibitive for large networks.

Question 3: What is depthwise separable convolution and what is its main advantage?

A 3D convolution for volumetric data
A convolution that separates positive and negative activations
A factored convolution applied depthwise then pointwise, dramatically reducing parameters and computation (Correct answer)
A convolution that processes depth information only

Correct answer: A factored convolution applied depthwise then pointwise, dramatically reducing parameters and computation

Depthwise separable convolution applies a single filter per input channel (depthwise) then combines channels with 1×1 convolution (pointwise), drastically reducing computation vs standard convolution.

Question 4: What is the main weakness of Adagrad that RMSProp was designed to fix?

Adagrad requires computing the full Hessian matrix
Adagrad doesn't use per-parameter learning rates
Adagrad's learning rate decreases monotonically and can become too small, stopping learning prematurely (Correct answer)
Adagrad doesn't work for convex optimization

Correct answer: Adagrad's learning rate decreases monotonically and can become too small, stopping learning prematurely

Adagrad's accumulation of all past squared gradients causes the learning rate to decrease monotonically to near-zero; RMSProp uses an exponential moving average to keep the rate from shrinking too much.

Question 5: What is the peephole connection in LSTM variants?

A shortcut connection from input to output bypassing the gates
An additional output gate for multi-task learning
A skip connection from early to late layers
Connections allowing the gates to observe the cell state directly when computing gate values (Correct answer)

Correct answer: Connections allowing the gates to observe the cell state directly when computing gate values

Peephole connections allow the forget, input, and output gates to look at the current cell state (not just the hidden state), giving gates access to more information for their decisions.

Question 6: What does the strength of association rule that is indicated by support and

Binding strength factor
Deviation factor
Depth factor
Confidence factor (Correct answer)

Correct answer: Confidence factor

Explanation: The support and confidence in an association rule can be used to determine its strength. The frequency with which a rule applies to a specific situation is determined by support.

Question 7: What does the 'dying ReLU' problem refer to?

ReLU activations that oscillate between 0 and 1
Neurons that permanently output zero due to large negative biases (Correct answer)
ReLU units that saturate at very high values
ReLU gradients that become too large

Correct answer: Neurons that permanently output zero due to large negative biases

Dying ReLU occurs when neurons receive consistently negative inputs, causing them to always output zero and receive no gradient, effectively becoming permanently inactive.

Question 8: What is the purpose of the temperature parameter in the softmax function during knowledge distillation?

To control the mixing ratio of hard and soft targets
To soften the probability distribution so the teacher's output provides more information about inter-class similarities (Correct answer)
To control the learning rate during distillation
To scale the student network's logits before computing loss

Correct answer: To soften the probability distribution so the teacher's output provides more information about inter-class similarities

Higher temperature in softmax produces softer, more spread-out probability distributions from the teacher, revealing relative similarities between classes that hard labels hide.

Question 9: What is a recommendation system using neural networks and how does it differ from traditional collaborative filtering?

It recommends neural network architectures for tasks
It uses deep embeddings to learn complex user-item interactions beyond linear patterns in traditional matrix factorization (Correct answer)
It uses the network's loss as a recommendation confidence score
It filters out low-quality training data automatically

Correct answer: It uses deep embeddings to learn complex user-item interactions beyond linear patterns in traditional matrix factorization

Neural collaborative filtering learns non-linear user-item interactions through deep embeddings and MLP layers, capturing complex patterns that traditional dot-product-based methods miss.

Question 10: What is a bidirectional RNN?

An RNN that uses two separate hidden states for long and short term
An RNN where gradients flow in two directions during backpropagation
An RNN that processes data both forwards and backwards in time simultaneously (Correct answer)
An RNN that can handle both text and image inputs

Correct answer: An RNN that processes data both forwards and backwards in time simultaneously

A bidirectional RNN processes sequences in both forward and backward temporal directions, concatenating or summing the two hidden states to capture context from both past and future.

Question 11: What is Backpropagation Through Time (BPTT)?

A method for transferring RNN weights across time periods
A regularization technique specific to sequential models
The algorithm for computing gradients in RNNs by unrolling the network across time steps (Correct answer)
A technique to speed up RNN training using time-based learning rates

Correct answer: The algorithm for computing gradients in RNNs by unrolling the network across time steps

BPTT unrolls the RNN across all time steps and applies standard backpropagation, computing gradients with respect to each time step's weights.

Question 12: What is non-maximum suppression (NMS) used for in object detection CNNs?

Normalizing confidence scores across all detections
Removing neurons with near-zero activations
Eliminating redundant overlapping bounding box predictions to retain the most confident detections (Correct answer)
Suppressing gradient magnitudes during backpropagation

Correct answer: Eliminating redundant overlapping bounding box predictions to retain the most confident detections

NMS removes duplicate overlapping bounding boxes by retaining only the box with the highest confidence score among boxes with high IoU overlap.

Question 13: What is feature map visualization and why is it used in CNN analysis?

Visualizing the training loss over time
Showing the confusion matrix of predictions
Displaying the intermediate activations to understand what features each layer detects (Correct answer)
Plotting the distribution of filter weights

Correct answer: Displaying the intermediate activations to understand what features each layer detects

Feature map visualization displays the activations of intermediate layers, revealing what patterns (edges, textures, shapes) the network has learned to detect at each stage.

Question 14: Early stopping is a regularization technique that halts training when:

Validation loss stops improving or begins to increase (Correct answer)
The learning rate drops below a threshold
Training loss reaches zero
All weight values converge to the same value

Correct answer: Validation loss stops improving or begins to increase

Early stopping monitors validation loss and stops training when it ceases to improve, preventing the model from continuing to overfit the training data.

Question 15: What is curriculum learning in neural network training?

Training the network on easier examples first, then gradually increasing difficulty (Correct answer)
Applying different learning rates to different layers
Using a fixed curriculum of hyperparameter schedules
Presenting training samples in random order

Correct answer: Training the network on easier examples first, then gradually increasing difficulty

Curriculum learning organizes training data from easy to hard examples, mimicking how humans learn and often leading to better generalization and faster convergence.

Question 16: What is the primary challenge of applying neural networks to tabular data compared to images?

Neural networks cannot handle categorical features
Tabular data lacks the spatial structure that makes CNNs effective, and gradient boosting often outperforms deep learning (Correct answer)
Tabular data requires different activation functions
Tabular data has too many features for neural networks

Correct answer: Tabular data lacks the spatial structure that makes CNNs effective, and gradient boosting often outperforms deep learning

Unlike images, tabular data lacks spatial/sequential structure, and tree-based methods like XGBoost often outperform neural networks; recent architectures like TabNet and FT-Transformer aim to close this gap.

Question 17: What is a Siamese network and what tasks is it used for?

A network with two different architectures trained jointly
Two identical networks sharing weights that learn similarity between input pairs, used for verification tasks (Correct answer)
A network with dual output heads for multi-task learning
A network trained on data from two different domains simultaneously

Correct answer: Two identical networks sharing weights that learn similarity between input pairs, used for verification tasks

Siamese networks use two weight-sharing branches to encode two inputs into embeddings, measuring their similarity or distance for tasks like face verification, signature matching, or one-shot learning.

Question 18: What is the purpose of hooks in PyTorch and when are they used?

To load model checkpoints during training
To sync model weights across distributed training nodes
To register functions that execute on forward/backward passes for tasks like feature extraction or gradient analysis (Correct answer)
To attach optimizers to model parameters

Correct answer: To register functions that execute on forward/backward passes for tasks like feature extraction or gradient analysis

PyTorch hooks (forward hooks and backward hooks) let you register callback functions on modules or tensors to inspect or modify activations and gradients during computation, useful for debugging and feature visualization.

Question 19: In TensorFlow/Keras, what does model.compile() configure?

The model's layers and architecture
The optimizer, loss function, and evaluation metrics for training (Correct answer)
The input and output tensor shapes
The hardware device (CPU/GPU) for training

Correct answer: The optimizer, loss function, and evaluation metrics for training

model.compile() in Keras sets up the optimizer, loss function, and metrics that will be used during model.fit() training, preparing the model for the training loop.

Question 20: What is the primary use case for TensorFlow Lite?

Distributed training across multiple machines
Training large models on cloud TPUs
Visualizing TensorFlow model architectures
Running pretrained neural network models on mobile and embedded devices with low latency (Correct answer)

Correct answer: Running pretrained neural network models on mobile and embedded devices with low latency

TensorFlow Lite is a lightweight runtime for deploying trained TensorFlow models on mobile (Android/iOS) and embedded (microcontrollers) devices with optimized performance.

Question 21: What does the stride parameter control in a convolutional layer?

The depth of the feature maps produced
The number of filters in the layer
The learning rate for that layer's weights
The step size with which the filter slides across the input (Correct answer)

Correct answer: The step size with which the filter slides across the input

Stride controls how many pixels the convolutional filter moves at each step; a stride of 2 halves the spatial dimensions of the output compared to stride 1.

Question 22: What does max pooling accomplish in a CNN?

It applies a learned transformation to reduce dimensions
It learns to detect edges in the input
It normalizes feature map values between 0 and 1
It downsamples feature maps by taking the maximum value in each pooling window (Correct answer)

Correct answer: It downsamples feature maps by taking the maximum value in each pooling window

Max pooling reduces the spatial dimensions of feature maps by selecting the maximum activation within each pooling window, providing spatial invariance.

Question 23: In CNN-based image classification, what does the global average pooling layer do before the final classifier?

Applies max pooling across all feature maps simultaneously
Concatenates all feature maps into a single vector
Averages each entire feature map to a single value, reducing spatial dimensions to 1×1 (Correct answer)
Pools gradients globally for stability

Correct answer: Averages each entire feature map to a single value, reducing spatial dimensions to 1×1

Global average pooling computes the mean of each feature map, converting spatial feature maps of any size into a fixed-length vector suitable for classification.

Question 24: What distinguishes ReLU from sigmoid and tanh activation functions?

ReLU is a second-order polynomial function
ReLU outputs probabilities between 0 and 1
ReLU does not saturate for positive values, avoiding vanishing gradients in that region (Correct answer)
ReLU is differentiable everywhere unlike sigmoid and tanh

Correct answer: ReLU does not saturate for positive values, avoiding vanishing gradients in that region

ReLU (f(x) = max(0, x)) has a constant gradient of 1 for positive inputs, unlike sigmoid and tanh which saturate (near-zero gradients) at extreme values, alleviating vanishing gradients.

Question 25: What does 'label smoothing' do to the target distribution in classification?

It replaces all labels with uniform probabilities
It replaces hard one-hot labels with a soft distribution that assigns small probability to other classes (Correct answer)
It normalizes labels to have zero mean
It randomly flips labels during training

Correct answer: It replaces hard one-hot labels with a soft distribution that assigns small probability to other classes

Label smoothing replaces one-hot targets with softened distributions (e.g., 0.9 for correct class, 0.1/K distributed among others), improving calibration and reducing overconfidence.

Question 26: An auto-associative network is defined as:

a single layer feed-forward neural network with pre-processing
has only single loop
Has a feedback mechanism (Correct answer)
does not have loops

Correct answer: Has a feedback mechanism

Explanation: A neural network with feedback is the same as an auto-associative network. It is not necessary to have only one feedback path (loop)

Question 27: What is the difference between model inference and model training in the context of neural networks?

Training updates weights via backpropagation; inference runs a forward pass only to produce predictions (Correct answer)
Inference requires labeled data; training does not
Training uses smaller batch sizes than inference
Inference uses different hardware than training always

Correct answer: Training updates weights via backpropagation; inference runs a forward pass only to produce predictions

Training involves both forward passes and backward passes to update weights; inference only performs the forward pass to generate predictions from learned weights, requiring no gradient computation.

Question 28: What is a 1×1 convolution and why is it useful?

A convolution that mixes channel information without affecting spatial dimensions (Correct answer)
A convolution that applies no transformation
A convolution used only for the final classification layer
A convolution that doubles the spatial dimensions

Correct answer: A convolution that mixes channel information without affecting spatial dimensions

A 1×1 convolution performs a linear combination across channels at each spatial position, enabling channel dimensionality reduction or expansion without affecting spatial size.

Question 29: What is the primary effect of dropout regularization during training?

It randomly deactivates a fraction of neurons, preventing co-adaptation (Correct answer)
It clips gradient values to prevent explosion
It reduces the learning rate over time
It normalizes the activations at each layer

Correct answer: It randomly deactivates a fraction of neurons, preventing co-adaptation

Dropout randomly sets a fraction of neuron activations to zero during each training step, forcing the network to learn redundant representations and reducing co-adaptation.

Question 30: What is the typical dropout rate recommended for hidden layers in a deep neural network?

0.5 to 0.8
0.2 to 0.5 (Correct answer)
0.9 to 1.0
0.1 to 0.2

Correct answer: 0.2 to 0.5

A dropout rate of 0.2 to 0.5 (20–50% of neurons dropped) is commonly recommended for hidden layers, balancing regularization without excessively impeding learning.

Question 31: What is dropout regularization and how does it prevent overfitting?

It reduces the learning rate whenever validation loss increases
It randomly deactivates a fraction of neurons during each training step, preventing co-adaptation (Correct answer)
It drops training samples with high loss values from each mini-batch
It removes neurons with small weights from the architecture permanently

Correct answer: It randomly deactivates a fraction of neurons during each training step, preventing co-adaptation

Dropout randomly sets a fraction of neuron activations to zero during each training forward pass, forcing the network to learn redundant representations and preventing co-adaptation.

Question 32: What is the GELU activation function and where is it commonly used?

Gaussian Error Linear Unit; commonly used in Transformers and BERT (Correct answer)
Gated Exponential Learning Unit; used in LSTMs
Grouped Efficient Linear Unit; used in lightweight CNNs
Generalized Exponential Linear Unit; used in ResNets

Correct answer: Gaussian Error Linear Unit; commonly used in Transformers and BERT

GELU (Gaussian Error Linear Unit) weights inputs by their Gaussian CDF, providing a smooth, probabilistic gating effect; it's the default activation in BERT, GPT, and most modern Transformers.

Question 33: What is the ELU (Exponential Linear Unit) activation and its advantage over ReLU?

ELU applies a learnable parameter to all inputs
ELU outputs larger values for positive inputs, speeding up training
ELU clips large activations to prevent gradient explosion
ELU uses an exponential for negative inputs, producing negative mean activations that help self-normalize (Correct answer)

Correct answer: ELU uses an exponential for negative inputs, producing negative mean activations that help self-normalize

ELU uses an exponential function for negative inputs, producing negative mean activations that push the mean closer to zero, reducing bias shift and often outperforming ReLU.

Question 34: What distinguishes L1 regularization from L2 regularization in terms of the weight solutions they produce?

L1 and L2 produce identical weight distributions
L1 produces dense small weights; L2 produces sparse weights
L1 is used only for output layers; L2 is used only for hidden layers
L1 tends to produce sparse weights with some exactly zero; L2 produces small but non-zero weights (Correct answer)

Correct answer: L1 tends to produce sparse weights with some exactly zero; L2 produces small but non-zero weights

L1 regularization's absolute value penalty creates a sparse solution by driving some weights to exactly zero, while L2's squared penalty shrinks all weights toward zero but rarely makes them exactly zero.

Question 35: What neural network has no hidden layers?

Recurrent Neural Networks
Neural Networks
Single-layer Perceptron (Correct answer)
Multi-layer perceptron (MLP)

Correct answer: Single-layer Perceptron

Explanation: A single-layer neural network is the simplest type of neural network, with only one layer of input nodes sending weighted inputs to a later layer of receiving nodes, or in certain situations, to only one receiving node.

Question 36: What does SELU (Scaled ELU) achieve that makes it 'self-normalizing'?

It normalizes gradients to unit variance
It applies batch normalization at each step
It scales weights to lie within a fixed range
With proper initialization, it drives activations to zero mean and unit variance automatically across layers (Correct answer)

Correct answer: With proper initialization, it drives activations to zero mean and unit variance automatically across layers

SELU uses carefully chosen scale and alpha parameters that, with Lecun normal initialization, maintain zero mean and unit variance of activations throughout the network without explicit normalization layers.

Question 37: Which of the following is a symptom of underfitting in a neural network?

High training accuracy and low test accuracy
Low training accuracy and high test accuracy
High training accuracy and high test accuracy
Low training accuracy and low test accuracy (Correct answer)

Correct answer: Low training accuracy and low test accuracy

Underfitting occurs when the model is too simple to capture the underlying data patterns, resulting in poor performance on both the training set and the test set.

Question 38: What is ONNX (Open Neural Network Exchange) used for?

A standard format for representing deep learning models enabling interoperability between frameworks (Correct answer)
A tool for visualizing neural network architectures
An open-source framework for training neural networks
A dataset format for neural network benchmarking

Correct answer: A standard format for representing deep learning models enabling interoperability between frameworks

ONNX provides an open format for neural network models, allowing models trained in PyTorch, TensorFlow, or other frameworks to be exported and run in different inference runtimes.

Question 39: In elastic net regularization, which combination of penalties is applied to the loss function?

A multiplicative combination of L1 and L2 penalties
A linear combination of both L1 and L2 penalties (Correct answer)
L2 penalty only, applied to output weights
L1 penalty only, applied to all layers

Correct answer: A linear combination of both L1 and L2 penalties

Elastic net regularization combines both L1 and L2 penalties as a weighted sum, achieving both sparsity (from L1) and small weight magnitudes (from L2).

Question 40: What does L2 regularization (weight decay) do to the weights during training?

It normalizes weights to have unit norm after each update
It increases large weights to improve learning speed
It encourages weights to remain small by penalizing their squared magnitude (Correct answer)
It sets small weights to exactly zero

Correct answer: It encourages weights to remain small by penalizing their squared magnitude

L2 regularization adds the sum of squared weights to the loss, creating a penalty that shrinks weights toward zero proportionally, preventing any single weight from dominating.

Question 41: What does LSTM stand for and what problem was it designed to solve?

Layered Sequential Training Module; overfitting in time-series models
Large Scale Temporal Model; slow training in deep networks
Long Short-Term Memory; vanishing/exploding gradients in standard RNNs (Correct answer)
Learned State Transition Memory; sparse gradient updates

Correct answer: Long Short-Term Memory; vanishing/exploding gradients in standard RNNs

LSTM (Long Short-Term Memory) was designed by Hochreiter and Schmidhuber to address vanishing and exploding gradients in standard RNNs by using gated cell states.

Question 42: What is AutoML and how does it relate to neural network development?

Automatic model deployment to cloud infrastructure
A method for automatically labeling training data
Automatic parallelization of neural network training across GPUs
Automated machine learning that automates architecture search, hyperparameter tuning, and pipeline optimization (Correct answer)

Correct answer: Automated machine learning that automates architecture search, hyperparameter tuning, and pipeline optimization

AutoML automates key steps in the ML pipeline — feature engineering, model selection, hyperparameter optimization, and NAS — reducing the expertise needed to build high-performing neural networks.

Question 43: What is transfer learning in the context of CNNs?

Converting a CNN model to run on mobile devices
Reusing pretrained CNN weights as a starting point for a new task (Correct answer)
Transferring a model from one GPU to another
Sharing weights between encoder and decoder networks

Correct answer: Reusing pretrained CNN weights as a starting point for a new task

Transfer learning uses a CNN pretrained on a large dataset (like ImageNet) as the starting point for a new task, leveraging learned feature representations to reduce training time and data requirements.

Question 44: What is federated learning in the context of neural network training?

Distributing training across multiple GPUs in a data center
A method for federating pretrained models from multiple sources
Training models across decentralized devices while keeping data local and only sharing model updates (Correct answer)
Distributed hyperparameter tuning across cloud providers

Correct answer: Training models across decentralized devices while keeping data local and only sharing model updates

Federated learning trains a shared model across many devices (e.g., smartphones) without centralizing data — each device trains locally and only shares weight updates with a central server.

Question 45: What metric is used to measure the overlap between predicted and ground-truth bounding boxes in object detection?

Intersection over Union (IoU) (Correct answer)
Dice coefficient
Mean Average Precision only
F1 score

Correct answer: Intersection over Union (IoU)

Intersection over Union (IoU) measures the ratio of the overlapping area to the combined area of predicted and ground-truth boxes, ranging from 0 (no overlap) to 1 (perfect match).

Question 46: What is learning rate warmup in neural network training?

Using different learning rates for different layers
Applying cyclic learning rate schedules
Decreasing the learning rate at the end of training
Gradually increasing the learning rate at the start of training (Correct answer)

Correct answer: Gradually increasing the learning rate at the start of training

Learning rate warmup starts with a very small learning rate and gradually increases it during the early training steps to stabilize optimization.

Question 47: Which neural network has an entry point into complicated neural nets, where input data goes via several layers of artificial neurons and each node is connected to all neurons in the following layer, resulting in a fully connected neural network?

Convolutional Neural Network
Recurrent Neural Networks
Multi-Layer Perceptron (Correct answer)
Feedforward Neural Networks

Correct answer: Multi-Layer Perceptron

Explanation: A multilayer perceptron (MLP) is a feedforward artificial neural network that creates outputs from inputs. Multiple layers of input nodes are connected as a directed graph between the input and output layers of an MLP. Multilayer perceptron uses backpropogation to train the network.

Question 48: What is the bias-variance tradeoff in the context of neural networks?

Balancing training speed against model accuracy
Balancing underfitting (high bias) against overfitting (high variance) (Correct answer)
Balancing the number of layers against the number of neurons per layer
Balancing weight initialization against learning rate

Correct answer: Balancing underfitting (high bias) against overfitting (high variance)

The bias-variance tradeoff describes the tension between a model that is too simple (high bias, underfits) and one that is too complex (high variance, overfits).

Question 49: What is the swish activation function and why might it outperform ReLU?

A piecewise linear function approximating sigmoid
A clipped version of tanh for better gradient flow
A learned activation function with trainable parameters
f(x) = x · sigmoid(x), which is smooth and non-monotonic, often outperforming ReLU in deep networks (Correct answer)

Correct answer: f(x) = x · sigmoid(x), which is smooth and non-monotonic, often outperforming ReLU in deep networks

Swish (x·σ(x)) is smooth, non-monotonic, and unbounded above, properties that often lead to improved performance over ReLU in deeper architectures according to Google Brain research.

Question 50: Which of the following regularization strategies is most effective when the training dataset is very small?

Using dropout with a high drop rate of 0.9
Removing all regularization to preserve capacity
Data augmentation and transfer learning (Correct answer)
Increasing model depth

Correct answer: Data augmentation and transfer learning

When training data is scarce, data augmentation increases effective dataset size and transfer learning leverages pre-trained features, both significantly reducing overfitting.

Question 51: Which component of a Variational Autoencoder (VAE) enables backpropagation through the sampling step?

The reparameterization trick (Correct answer)
The encoder bottleneck
The reconstruction loss
The KL divergence term

Correct answer: The reparameterization trick

The reparameterization trick expresses the random sample as a deterministic function of the latent mean and variance, enabling gradients to flow through.

Question 52: What is the purpose of an activation function in a neural network?

To normalize the inputs to each layer
To introduce non-linearity so the network can learn complex patterns (Correct answer)
To reduce the number of parameters in the model
To compute the gradient during backpropagation

Correct answer: To introduce non-linearity so the network can learn complex patterns

Without activation functions, a neural network would be equivalent to a single linear transformation regardless of depth; activation functions introduce non-linearity enabling complex function approximation.

Question 53: What is the purpose of padding in convolutional layers?

To preserve spatial dimensions and allow filters to be applied at border regions (Correct answer)
To increase the number of learned filters
To add extra training examples to the dataset
To prevent overfitting by adding noise

Correct answer: To preserve spatial dimensions and allow filters to be applied at border regions

Padding (typically zero-padding) adds border pixels to the input so that the output feature map maintains the same spatial dimensions and edge information is not lost.

Question 54: What is TensorBoard primarily used for?

Deploying models to production servers
Visualizing training metrics, model graphs, embeddings, and other diagnostics during training (Correct answer)
Compiling model code to native binaries
Managing distributed training across multiple GPUs

Correct answer: Visualizing training metrics, model graphs, embeddings, and other diagnostics during training

TensorBoard is a visualization toolkit for TensorFlow (and PyTorch via SummaryWriter) that displays training curves, model architecture graphs, weight histograms, and embedding projections.

Question 55: What distinguishes an automated vehicle?

Reinforcement learning
Active learning
Unsupervised learning
Supervised learning (Correct answer)

Correct answer: Supervised learning

Explanation: In supervised learning, an educator (for example, a system designer) oversees the artificial neural network and prepares it with labeled data sets using his or her expertise of the system.

Question 56: What distinguishes a capsule network from a standard CNN?

It uses recurrent connections
It uses no pooling operations at all
It encodes pose information as vectors rather than scalar activations (Correct answer)
It applies self-attention to feature maps

Correct answer: It encodes pose information as vectors rather than scalar activations

Capsule networks represent features as vectors (capsules) that encode both the presence and spatial relationships/pose of entities, unlike scalar activations in CNNs.

Neural Networks Fundamentals Assessment

A comprehensive assessment covering the theory and practical application of artificial neural networks, including architecture design, training algorithms, optimization techniques, and modern deep learning frameworks.

Questions

75 min

Time Limit

70%

To Pass

Checking for saved progress...

Exam Rules

You can skip questions and return to them later
Flag questions for review before submitting
No feedback shown until you submit the entire exam
Unanswered questions count as wrong — answer everything
10 pretest questions are mixed in and don't affect your score
Timer auto-submits when time runs out
Your progress is auto-saved every 30 seconds