CNNs use convolution layers to detect spatial features, making them ideal for images.
Incorrect! Try again.
10Which type of neural network is best suited for sequential data such as text or time series?
types of ANN
Easy
A.Autoencoder
B.Feedforward Network
C.Convolutional Neural Network (CNN)
D.Recurrent Neural Network (RNN)
Correct Answer: Recurrent Neural Network (RNN)
Explanation:
RNNs have loops that let them retain memory of previous inputs, suiting them to sequential data.
Incorrect! Try again.
11In a feedforward neural network, information flows in which direction?
types of ANN
Easy
A.Randomly between layers
B.Only forward, input to output
C.Only backward, output to input
D.In circular loops
Correct Answer: Only forward, input to output
Explanation:
In a feedforward network, signals move in one direction from the input layer to the output layer without cycles.
Incorrect! Try again.
12Which network type is designed to learn a compressed representation of data for tasks like dimensionality reduction?
types of ANN
Easy
A.Autoencoder
B.GAN
C.Perceptron
D.CNN
Correct Answer: Autoencoder
Explanation:
An autoencoder encodes input into a compressed form and reconstructs it, useful for dimensionality reduction.
Incorrect! Try again.
13What does GAN stand for in deep learning?
types of ANN
Easy
A.Generative Adversarial Network
B.Grouped Activation Node
C.Graph Aligned Network
D.General Analog Network
Correct Answer: Generative Adversarial Network
Explanation:
A GAN consists of a generator and a discriminator competing with each other to produce realistic data.
Incorrect! Try again.
14The simplest type of artificial neural network, consisting of a single neuron, is called a:
types of ANN
Easy
A.Autoencoder
B.Perceptron
C.Transformer
D.Boltzmann machine
Correct Answer: Perceptron
Explanation:
The perceptron is the simplest neural model, using a single neuron for binary classification.
Incorrect! Try again.
15Which specialized RNN variant is designed to handle long-term dependencies in sequences?
types of ANN
Easy
A.LSTM (Long Short-Term Memory)
B.Autoencoder
C.Perceptron
D.CNN
Correct Answer: LSTM (Long Short-Term Memory)
Explanation:
LSTM networks use gates to remember information over long sequences, solving the vanishing gradient problem in RNNs.
Incorrect! Try again.
16Deep learning is commonly used in healthcare to analyze medical images such as:
case studies for the application of deep learning in biology and health care research
Easy
A.Weather reports
B.Text messages
C.Bank statements
D.X-rays and MRI scans
Correct Answer: X-rays and MRI scans
Explanation:
CNNs are widely applied to X-rays, MRI, and CT scans to detect diseases like tumors and fractures.
Incorrect! Try again.
17In biology, deep learning has been famously applied to predict the 3D structure of:
case studies for the application of deep learning in biology and health care research
Easy
A.Volcanoes
B.Bridges
C.Planets
D.Proteins
Correct Answer: Proteins
Explanation:
Tools like AlphaFold use deep learning to accurately predict protein 3D structures from amino acid sequences.
Incorrect! Try again.
18Which healthcare task uses deep learning to identify cancerous cells in tissue slides?
case studies for the application of deep learning in biology and health care research
Easy
A.Payroll processing
B.Digital pathology image analysis
C.Appointment scheduling
D.Inventory tracking
Correct Answer: Digital pathology image analysis
Explanation:
Deep learning models analyze histopathology images to detect and classify cancerous cells.
Incorrect! Try again.
19In genomics, deep learning is used to analyze which type of data?
case studies for the application of deep learning in biology and health care research
Easy
A.Stock prices
B.Satellite orbits
C.DNA and gene sequences
D.Traffic signals
Correct Answer: DNA and gene sequences
Explanation:
Deep learning helps interpret DNA and gene sequences to predict gene function and disease risk.
Incorrect! Try again.
20Which of the following is a healthcare application of deep learning?
case studies for the application of deep learning in biology and health care research
Easy
A.Predicting disease from patient records
B.Painting house walls
C.Playing background music
D.Cooking recipes automatically
Correct Answer: Predicting disease from patient records
Explanation:
Deep learning models analyze electronic health records to predict disease onset and support clinical decisions.
Incorrect! Try again.
21A researcher is designing a network to classify pixel images of handwritten digits while preserving spatial relationships between pixels. Which type of ANN is most appropriate?
CNNs use convolutional filters that preserve spatial locality, making them ideal for image data where neighboring pixels are correlated.
Incorrect! Try again.
22For modeling a time-series of gene expression measured over sequential days, where the value at each step depends on previous steps, which architecture is best suited?
types of ANN
Medium
A.Feedforward Neural Network
B.Convolutional Neural Network (CNN)
C.Radial Basis Function Network
D.Recurrent Neural Network (RNN)
Correct Answer: Recurrent Neural Network (RNN)
Explanation:
RNNs maintain a hidden state that carries information across time steps, allowing them to model sequential dependencies in time-series data.
Incorrect! Try again.
23What is the key limitation of a standard RNN that motivated the development of LSTM networks?
types of ANN
Medium
A.The vanishing gradient problem over long sequences
B.Excessive memory requirements for small inputs
C.Lack of a nonlinear activation function
D.Inability to process any sequential data
Correct Answer: The vanishing gradient problem over long sequences
Explanation:
Standard RNNs suffer from vanishing gradients, making it hard to learn long-range dependencies. LSTMs use gating mechanisms to preserve gradients over longer sequences.
Incorrect! Try again.
24An autoencoder is primarily used for which of the following tasks?
types of ANN
Medium
A.Sorting data into predefined clusters using labels
B.Supervised classification of labeled images
C.Unsupervised learning of compressed data representations
D.Reinforcement learning of game strategies
Correct Answer: Unsupervised learning of compressed data representations
Explanation:
Autoencoders learn to compress (encode) input into a lower-dimensional representation and reconstruct (decode) it, making them useful for unsupervised dimensionality reduction and denoising.
Incorrect! Try again.
25In a Generative Adversarial Network (GAN), what are the two competing components?
types of ANN
Medium
A.A teacher and a student
B.A generator and a discriminator
C.An encoder and a decoder
D.A forward pass and a backward pass
Correct Answer: A generator and a discriminator
Explanation:
A GAN consists of a generator that creates synthetic samples and a discriminator that tries to distinguish real from fake, trained in an adversarial min-max game.
Incorrect! Try again.
26Which statement best distinguishes a feedforward neural network from a recurrent neural network?
types of ANN
Medium
A.Both have identical topologies and differ only in activation functions
B.A feedforward network has no cycles, while an RNN contains feedback loops that feed outputs back as inputs across time steps to model sequential dependencies
C.A feedforward network can only handle images
D.A feedforward network uses cycles while an RNN does not
Correct Answer: A feedforward network has no cycles, while an RNN contains feedback loops that feed outputs back as inputs across time steps to model sequential dependencies
Explanation:
Feedforward networks pass data in one direction with no cycles, whereas RNNs include recurrent connections that allow information to persist across sequence steps.
Incorrect! Try again.
27Which type of ANN is specifically designed for graph-structured data such as molecular structures or protein interaction networks?
types of ANN
Medium
A.Graph Neural Network (GNN)
B.Long Short-Term Memory (LSTM)
C.Multilayer Perceptron (MLP)
D.Convolutional Neural Network (CNN)
Correct Answer: Graph Neural Network (GNN)
Explanation:
GNNs operate on nodes and edges, aggregating information from neighbors, making them well-suited for graph-structured data like molecules and biological networks.
Incorrect! Try again.
28Why is a nonlinear activation function such as ReLU essential in a deep neural network?
introduction of deep learning algorithms
Medium
A.It reduces the number of parameters needed
B.It removes the need for a loss function
C.It guarantees the network will never overfit
D.Without nonlinearity, stacked layers collapse into a single linear transformation
Correct Answer: Without nonlinearity, stacked layers collapse into a single linear transformation
Explanation:
Composing linear functions yields another linear function. Nonlinear activations let the network approximate complex, nonlinear mappings across multiple layers.
Incorrect! Try again.
29The ReLU activation function is defined as which of the following?
introduction of deep learning algorithms
Medium
A.
B.
C.
D.
Correct Answer:
Explanation:
ReLU (Rectified Linear Unit) outputs the input if positive and zero otherwise: , which helps mitigate vanishing gradients.
Incorrect! Try again.
30During backpropagation, what is being computed and used to update the network weights?
introduction of deep learning algorithms
Medium
A.The raw input values fed forward through the network
B.The number of neurons in each hidden layer
C.The gradient of the loss function with respect to each weight
D.The random initial weight values
Correct Answer: The gradient of the loss function with respect to each weight
Explanation:
Backpropagation applies the chain rule to compute the gradient of the loss with respect to each weight, which gradient descent then uses to update the weights.
Incorrect! Try again.
31A model achieves 99% accuracy on training data but only 70% on test data. What is this a clear sign of?
introduction of deep learning algorithms
Medium
A.Underfitting
B.A learning rate that is too small
C.Vanishing gradients
D.Overfitting
Correct Answer: Overfitting
Explanation:
High training accuracy with much lower test accuracy indicates overfitting: the model memorized training data instead of learning generalizable patterns.
Incorrect! Try again.
32Which technique randomly deactivates a fraction of neurons during training to reduce overfitting?
introduction of deep learning algorithms
Medium
A.Batch normalization
B.Dropout
C.Gradient clipping
D.Data augmentation
Correct Answer: Dropout
Explanation:
Dropout randomly sets a fraction of neuron activations to zero during each training step, preventing co-adaptation and improving generalization.
Incorrect! Try again.
33In gradient descent, if the learning rate is set too high, what is the most likely consequence?
introduction of deep learning algorithms
Medium
A.The network automatically adds more layers
B.The model always converges faster with no downside
C.The loss may diverge or oscillate instead of converging
D.The gradients become exactly zero
Correct Answer: The loss may diverge or oscillate instead of converging
Explanation:
A learning rate that is too high causes weight updates to overshoot minima, leading to oscillation or divergence of the loss rather than smooth convergence.
Incorrect! Try again.
34Which loss function is most appropriate for a multi-class classification problem with a softmax output layer?
introduction of deep learning algorithms
Medium
A.Mean absolute error
B.Hinge loss for regression
C.Categorical cross-entropy
D.Mean squared error
Correct Answer: Categorical cross-entropy
Explanation:
Categorical cross-entropy measures the divergence between predicted class probabilities (from softmax) and the true one-hot labels, making it standard for multi-class classification.
Incorrect! Try again.
35What is the primary purpose of the softmax function in the output layer of a classification network?
introduction of deep learning algorithms
Medium
A.It computes the gradient of the loss function directly
B.It reduces the number of output neurons to one
C.It converts raw scores into a probability distribution over classes that sums to 1
D.It normalizes the input images before training
Correct Answer: It converts raw scores into a probability distribution over classes that sums to 1
Explanation:
Softmax exponentiates and normalizes the output logits so they form a valid probability distribution across classes summing to 1.
Incorrect! Try again.
36DeepMind's AlphaFold made a major breakthrough in biology by accurately predicting which of the following?
case studies for the application of deep learning in biology and health care research
Medium
A.The 3D structure of proteins from their amino acid sequences
B.The stock prices of pharmaceutical companies
C.The exact age of a patient from a blood sample
D.The complete DNA sequence of the human genome
Correct Answer: The 3D structure of proteins from their amino acid sequences
Explanation:
AlphaFold uses deep learning to predict protein 3D structures from amino acid sequences, solving a decades-old grand challenge in structural biology.
Incorrect! Try again.
37In a case study applying CNNs to diagnose diabetic retinopathy, what type of input data is used?
case studies for the application of deep learning in biology and health care research
Medium
A.Genomic sequence text
B.Patient audio recordings
C.Tabular billing records
D.Retinal fundus images
Correct Answer: Retinal fundus images
Explanation:
Diabetic retinopathy screening applies CNNs to retinal fundus images to detect lesions and grade disease severity from visual features.
Incorrect! Try again.
38A hospital deploys a deep learning model to detect tumors in medical scans. Why is high recall (sensitivity) often prioritized over high precision in this context?
case studies for the application of deep learning in biology and health care research
Medium
A.Recall requires less computation than precision
B.Missing a true tumor (false negative) can be far more harmful to the patient than a false alarm
C.Precision is impossible to compute for medical images
D.High recall always guarantees the highest overall accuracy
Correct Answer: Missing a true tumor (false negative) can be far more harmful to the patient than a false alarm
Explanation:
In tumor detection, a false negative may delay life-saving treatment, so maximizing recall (catching as many true cases as possible) is often prioritized over precision.
Incorrect! Try again.
39When applying deep learning to medical imaging with limited labeled data, which strategy leverages a model pre-trained on a large general image dataset?
case studies for the application of deep learning in biology and health care research
Medium
A.Increasing the learning rate
B.Random weight initialization from scratch
C.Transfer learning
D.Removing all hidden layers
Correct Answer: Transfer learning
Explanation:
Transfer learning reuses features learned from a large dataset (e.g., ImageNet) and fine-tunes them on the smaller medical dataset, improving performance when labeled data is scarce.
Incorrect! Try again.
40In genomics, RNNs and CNNs are applied to DNA sequences primarily to accomplish which task?
case studies for the application of deep learning in biology and health care research
Medium
A.Replace the need for laboratory sequencing entirely
B.Identify functional motifs and predict regulatory elements from nucleotide sequences
C.Store patient records in a distributed database
D.Physically synthesize new DNA molecules in the network
Correct Answer: Identify functional motifs and predict regulatory elements from nucleotide sequences
Explanation:
Deep networks learn patterns in nucleotide sequences to predict binding sites, splice sites, and regulatory motifs, aiding genomic annotation and functional prediction.
Incorrect! Try again.
41In a Recurrent Neural Network (RNN) trained with Backpropagation Through Time (BPTT), the gradient of the loss with respect to early time steps involves a product of Jacobians. If the largest eigenvalue of the recurrent weight matrix's Jacobian is consistently , what phenomenon dominates?
types of ANN
Hard
A.Saturation of the output softmax layer only
B.Exploding gradients that destabilize weight updates
C.Perfect gradient flow across all time steps
D.Vanishing gradients that hinder learning of long-term dependencies
Correct Answer: Vanishing gradients that hinder learning of long-term dependencies
Explanation:
Multiplying many Jacobians with spectral radius shrinks the gradient exponentially toward zero, so contributions from distant time steps vanish. This is why vanilla RNNs struggle with long-term dependencies, motivating LSTM/GRU gating.
Incorrect! Try again.
42A Convolutional Neural Network uses a kernel with stride 1 and no padding on a input. What is the spatial dimension of the resulting feature map?
types of ANN
Hard
A.
B.
C.
D.
Correct Answer:
Explanation:
Output size . The map is therefore .
Incorrect! Try again.
43Which architectural property most fundamentally distinguishes an autoencoder used for anomaly detection from a standard feedforward classifier?
types of ANN
Hard
A.It applies recurrent connections to model temporal sequences
B.It is trained to reconstruct its input via a bottleneck, using reconstruction error as the signal
C.It uses labeled anomaly classes to minimize cross-entropy loss
D.It relies on max-pooling to downsample spatial features
Correct Answer: It is trained to reconstruct its input via a bottleneck, using reconstruction error as the signal
Explanation:
Autoencoders learn to reconstruct normal data through a compressed latent bottleneck. Anomalous inputs yield high reconstruction error, making the error itself the detection signal—no anomaly labels are required.
Incorrect! Try again.
44In a Generative Adversarial Network (GAN), the generator and discriminator play a minimax game with value function . At the theoretical global optimum, what value does output for real and generated samples?
types of ANN
Hard
A. everywhere, since the generator wins
B. for real and for generated samples
C. oscillates between and indefinitely
D. everywhere, as the two distributions become indistinguishable
Correct Answer: everywhere, as the two distributions become indistinguishable
Explanation:
At the global optimum , so the discriminator cannot distinguish real from fake and outputs , corresponding to the Jensen–Shannon divergence being minimized.
Incorrect! Try again.
45Why do Long Short-Term Memory (LSTM) networks mitigate the vanishing gradient problem more effectively than vanilla RNNs?
types of ANN
Hard
A.The cell state provides a nearly linear path allowing gradients to flow largely unattenuated
B.They replace all sigmoid activations with ReLU to prevent saturation
C.They eliminate recurrent connections entirely, using only feedforward layers
D.They increase the learning rate adaptively at every time step
Correct Answer: The cell state provides a nearly linear path allowing gradients to flow largely unattenuated
Explanation:
The LSTM cell state acts as a gated 'constant error carousel' where the forget gate controls a near-identity recurrence. This preserves gradient magnitude across long sequences far better than the repeated nonlinear multiplications in vanilla RNNs.
Incorrect! Try again.
46A Radial Basis Function (RBF) network differs from a multilayer perceptron primarily in how its hidden units respond. Which statement is correct?
types of ANN
Hard
A.RBF units compute weighted sums followed by softmax normalization
B.RBF units produce localized responses based on distance to a center, giving local approximation
C.RBF networks require at least three hidden layers to function
D.RBF units use global sigmoidal activations identical to MLP hidden units
Correct Answer: RBF units produce localized responses based on distance to a center, giving local approximation
Explanation:
RBF hidden units activate based on the distance between the input and a prototype center (e.g., a Gaussian), yielding localized receptive fields. MLPs use global activations of weighted sums, giving a fundamentally different, distributed approximation.
Incorrect! Try again.
47During backpropagation, the ReLU activation has a derivative that is for . What practical failure mode can this cause?
introduction of deep learning algorithms
Hard
A.Vanishing gradients across all layers regardless of input sign
B.The 'dying ReLU' problem, where neurons output zero and stop updating permanently
C.Exploding gradients due to unbounded positive activations
D.Guaranteed convergence to the global minimum
Correct Answer: The 'dying ReLU' problem, where neurons output zero and stop updating permanently
Explanation:
If a neuron's weighted input stays negative, its gradient is and weights never update, so the neuron 'dies.' Leaky ReLU and variants introduce a small negative slope to keep gradients flowing.
Incorrect! Try again.
48Adam optimizer combines momentum and RMSProp. Given first moment and second moment , why is bias correction applied in early iterations?
introduction of deep learning algorithms
Hard
A.It forces the second moment to equal the first moment
B.It converts the gradients into probabilities for stability
C.The moment estimates are initialized at zero and are biased toward zero early in training
D.It prevents the learning rate from ever decreasing
Correct Answer: The moment estimates are initialized at zero and are biased toward zero early in training
Explanation:
Because , the running averages are biased toward zero during the first steps. Dividing by corrects this bias, giving accurate estimates especially when is small.
Incorrect! Try again.
49Batch Normalization normalizes layer inputs using mini-batch statistics, then applies learnable parameters and . Why are and necessary?
introduction of deep learning algorithms
Hard
A.They replace the need for any activation function
B.They let the network recover representational capacity, including undoing normalization if optimal
C.They compute the gradient of the loss directly
D.They fix the mean and variance to and permanently
Correct Answer: They let the network recover representational capacity, including undoing normalization if optimal
Explanation:
Strictly forcing zero mean and unit variance could limit expressiveness. The learnable scale and shift allow the network to restore any needed distribution—even identity—so normalization does not reduce representational power.
Incorrect! Try again.
50Dropout with rate is applied during training. At inference time, why must activations be scaled (or weights adjusted)?
introduction of deep learning algorithms
Hard
A.To increase the variance of predictions at test time
B.To zero out half the neurons during inference as well
C.To convert the network into a recurrent architecture
D.To match the expected activation magnitude, since all units are active at test time
Correct Answer: To match the expected activation magnitude, since all units are active at test time
Explanation:
During training only a fraction of units are active, so activations are scaled to preserve expected values. At test time all units fire, so either activations are multiplied by (or 'inverted dropout' scales during training) to keep magnitudes consistent.
Incorrect! Try again.
51Consider training with cross-entropy loss and a softmax output. Why is the combination of softmax + cross-entropy numerically and computationally preferred over softmax followed by a separate MSE loss for classification?
introduction of deep learning algorithms
Hard
A.It always produces sparser weight matrices
B.Its gradient simplifies to , giving stable, well-scaled updates
C.It removes the need for any activation in hidden layers
D.It guarantees the loss surface is globally convex
Correct Answer: Its gradient simplifies to , giving stable, well-scaled updates
Explanation:
The gradient of softmax cross-entropy w.r.t. logits reduces elegantly to , avoiding saturation-induced tiny gradients that MSE with softmax suffers from. This yields faster, more stable classification training.
Incorrect! Try again.
52Xavier (Glorot) initialization sets weight variance to . What problem is this designed to prevent?
introduction of deep learning algorithms
Hard
A.The optimizer becoming stuck at a saddle point
B.Signal and gradient variance shrinking or growing exponentially across layers
C.Overfitting caused by too many training epochs
D.Class imbalance in the training dataset
Correct Answer: Signal and gradient variance shrinking or growing exponentially across layers
Explanation:
Xavier initialization keeps the variance of activations and gradients roughly constant across layers by accounting for fan-in and fan-out, preventing the vanishing/exploding signal that poor initialization causes in deep nets.
Incorrect! Try again.
53In transfer learning, a pretrained CNN is adapted to a small medical imaging dataset. Which strategy best reduces overfitting while leveraging learned features?
introduction of deep learning algorithms
Hard
A.Freeze the classifier head and train only the input layer
B.Remove all convolutional layers and use only dense layers
C.Freeze early convolutional layers and fine-tune only the later layers and classifier head
D.Randomly reinitialize all layers and train from scratch
Correct Answer: Freeze early convolutional layers and fine-tune only the later layers and classifier head
Explanation:
Early layers capture generic features (edges, textures) transferable across domains, so freezing them preserves useful representations and limits parameters to train. Fine-tuning later, task-specific layers adapts the model with less overfitting risk on small datasets.
Incorrect! Try again.
54AlphaFold predicts protein 3D structure from amino acid sequences. Which deep learning innovation was most central to its breakthrough accuracy in CASP14?
case studies for the application of deep learning in biology and health care research
Hard
A.Unsupervised clustering of protein images without sequence data
B.Recurrent networks processing one amino acid at a time without pairwise features
C.A single fully connected network mapping sequence directly to coordinates
D.Attention-based modeling of residue-residue relationships and evolutionary (MSA) information
Correct Answer: Attention-based modeling of residue-residue relationships and evolutionary (MSA) information
Explanation:
AlphaFold2 leverages attention (the Evoformer) to jointly reason over multiple sequence alignments and pairwise residue representations, capturing co-evolutionary and geometric constraints that drive its unprecedented structural accuracy.
Incorrect! Try again.
55A CNN trained to detect diabetic retinopathy achieves 99% accuracy on its training hospital's data but performs poorly at a new clinic. What is the most likely cause?
case studies for the application of deep learning in biology and health care research
Hard
A.Cross-entropy loss is inappropriate for image classification
B.The model has too few parameters to memorize training data
C.The learning rate was set too low during training
D.Domain shift, where imaging equipment and population differences alter the data distribution
Correct Answer: Domain shift, where imaging equipment and population differences alter the data distribution
Explanation:
Different cameras, acquisition protocols, and patient demographics shift the input distribution at the new site. The model overfits site-specific artifacts, so generalization fails—a key reason clinical DL requires multi-site validation and domain adaptation.
Incorrect! Try again.
56In genomics, a CNN scans DNA sequences to predict transcription factor binding sites. Why are convolutional filters particularly well suited to this task?
case studies for the application of deep learning in biology and health care research
Hard
A.They require the motif to appear at a fixed position to be detected
B.They act like position weight matrices detecting motifs regardless of location in the sequence
C.They inherently normalize GC content across the genome
D.They model long-range temporal dependencies better than any other layer
Correct Answer: They act like position weight matrices detecting motifs regardless of location in the sequence
Explanation:
Convolutional filters function like learnable motif detectors (analogous to position weight matrices), and weight sharing gives translation invariance—so binding motifs are recognized wherever they occur along the sequence.
Incorrect! Try again.
57When deploying a deep learning model for cancer diagnosis, why is model calibration (not just accuracy) critical?
case studies for the application of deep learning in biology and health care research
Hard
A.Calibration increases the raw accuracy above 100%
B.Calibration converts the classifier into a regression model
C.Calibration replaces the need for a validation dataset
D.Predicted probabilities must reflect true likelihoods to support clinical risk decisions
Correct Answer: Predicted probabilities must reflect true likelihoods to support clinical risk decisions
Explanation:
A well-calibrated model outputs probabilities matching real-world frequencies (e.g., 0.8 means 80% likely). Clinicians rely on these for risk-based decisions; a highly accurate but overconfident model can mislead treatment choices.
Incorrect! Try again.
58A recurrent model (LSTM) is applied to electronic health record (EHR) time series to predict sepsis onset. What key advantage does the LSTM provide over a static feedforward classifier here?
case studies for the application of deep learning in biology and health care research
Hard
A.It removes the need to handle missing values in the data
B.It eliminates class imbalance between septic and non-septic patients
C.It guarantees interpretability of every prediction to clinicians
D.It captures temporal patterns and irregular dependencies across sequential clinical measurements
Correct Answer: It captures temporal patterns and irregular dependencies across sequential clinical measurements
Explanation:
Sepsis prediction depends on how vital signs and labs evolve over time. LSTMs model these sequential dependencies and trends, which static classifiers ignore—though missing data and imbalance still require separate handling.
Incorrect! Try again.
59In drug discovery, Graph Neural Networks (GNNs) represent molecules as graphs. Why is this representation superior to a fixed-length fingerprint vector for property prediction?
case studies for the application of deep learning in biology and health care research
Hard
A.It requires no training data because chemistry is deterministic
B.It ignores molecular topology to focus only on atom counts
C.Message passing over atoms and bonds learns task-specific structural features directly
D.It converts every molecule into a single scalar automatically
Correct Answer: Message passing over atoms and bonds learns task-specific structural features directly
Explanation:
GNNs propagate information along bonds (message passing), letting the model learn representations tailored to the target property from the molecular topology, rather than relying on hand-crafted, fixed fingerprints that may miss relevant structure.
Incorrect! Try again.
60A U-Net architecture is widely used for biomedical image segmentation. What structural feature most directly enables its precise localization of boundaries?
case studies for the application of deep learning in biology and health care research
Hard
A.Purely recurrent connections between encoder and decoder
B.Absence of any downsampling to preserve full resolution throughout
C.Skip connections that fuse high-resolution encoder features with decoder upsampling
D.A single global average pooling layer at the output
Correct Answer: Skip connections that fuse high-resolution encoder features with decoder upsampling
Explanation:
U-Net's skip connections pass fine-grained spatial details from encoder layers directly to the corresponding decoder layers. This recovers boundary precision lost during downsampling, which is critical for accurate biomedical segmentation.
Incorrect! Try again.
Did this save you a night before the exam?
LPU Notes is free, and it stays free. Ads cover part of the server bill.
The rest comes out of a student's own pocket: the domain, the storage,
and keeping the site up through the weeks everyone needs it at once.
The payment button didn't load. An ad blocker or a filtered network is the usual reason.
to try again.
Nothing here is ever locked, and nothing unlocks. Chip in only if it was worth it.
What it pays for →