1/97
Looks like no tags are added yet.
In machine learning, learning a function f(x) is best described as a process that
Maps inputs to outputs learned from data
Which best describes function approximation in ML?
Fitting a curve or surface to data
In supervised learning, the true function f(x) is
Unknown but estimated from examples
Why is machine learning often considered function approximation rather than exact mapping?
Real data is noisy so we estimate f(x) approximately
Which of these performs non-linear function approximation?
Both Polynomial Regression and Decision Tree
Overfitting occurs when the approximated function:
Fits the training data too closely and fails to generalize
A model that fits the training data too closely and fails to generalize is said to be ____.
Overfitting
A model that is too simple and captures too little structure in the data is said to be ____.
Underfitting
If a model has high bias, it usually means that it
Underfits and approximates the function poorly
What are true about the bias-variance tradeoff.
Which of the following typically has the highest function approximation capacity?
Deep Neural Network
A function approximator is called universal if it can
Represent essentially any target function
The Universal Approximation Theorem applies specifically to
Neural networks with one hidden layer
Why might a deep network outperform a shallow one on complex function approximation?
It can model hierarchical relationships
Regularization helps function approximation mainly by
Penalizing large weights
What is the key difference between model capacity and model generalization?
Capacity is possible complexity, generalization is accuracy on new data
What correctly describe overfitting and generalization
Why is data normalization important before function approximation with neural networks?
It keeps features on comparable scales for stable training
A biological neuron primarily communicates using
Electric and chemical signals
Which part of a biological neuron corresponds to the weights in an artificial neuron?
The synapses
The part of a biological neuron that receives incoming signals is the ____.
Dendrites
The part of a biological neuron that transmits signals away from the cell body is the ____.
Axon
The junction where one neuron connects and passes signals to another is the ____.
Synapse
The biological neuron’s axon is analogous to what in an artificial neuron?
The output signal
What are parts of a biological neuron.
What are all components that make up an artificial neuron.
What are the correct mappings from biological neuron parts to artificial neuron parts.
In an artificial neuron, the weighted sum of inputs plus bias is immediately followed by
A non-linear activation function
In an artificial neuron, each input is multiplied by an adjustable parameter called a ____.
Weight
The constant term added to the weighted sum in an artificial neuron is the ____.
Bias
Why are activation functions like ReLU or sigmoid crucial in neural networks?
They introduce non-linearity
Why can’t a neural network without activation functions solve non-linear problems?
A stack of linear layers is still just a linear function
The main limitation of comparing artificial and biological neurons is that
Biological networks are far more complex and adaptive
Who first proposed the neuron model that inspired artificial neurons?
McCulloch and Pitts
What is true of the McCulloch-Pitts and early neuron models.
A single perceptron can only solve problems that are:
Linearly separable
The logical function that a single perceptron cannot learn because it is not linearly separable is ____.
XOR
The perceptron decision rule outputs class 1 when:
The weighted sum plus bias exceeds the threshold
Perceptron learning updates the weights only when:
The output differs from the target
What is true about a single perceptron.
The classic perceptron uses a ____ (threshold) activation that outputs 0 or 1.
Step
The single-neuron model with learnable weights introduced by Rosenblatt is called the ____.
Perceptron
The Greek letter commonly used to denote the learning rate is called
Eta
What does ADALINE stand for?
Adaptive Linear Neuron
Which concept did ADALINE contribute to modern neural networks?
Continuous weight updates and a differentiable loss
ADALINE adjusts its weights to minimize the mean squared ____ between its linear output and the target.
Error
If the learning rate is set too high in ADALINE, what typically happens?
It will fail to converge
When ADALINE’s weights stop changing significantly, the model has:
Converged
What is true about ADALINE.
Which algorithm generalizes both Perceptron and ADALINE ideas into multi-layer networks?
Backpropagation Neural Network
A neural network is mainly designed to
Approximate functions
The smallest computational unit in a neural network is called a
node
The bias term in a neuron helps by
Shifting the activation threshold
Which of the following architectures allows information flow from input to output without loops?
FNN
What is a “layer” in a neural network?
Group of neurons processing inputs
Which of the following is not an activation function?
MSE
Why do we use deep neural networks?
To learn hierarchical features
The capacity of a neural network increases when
number of layers increases
In neural networks, “epoch” refers to
One full pass of the data
Forward propagation refers to
Passing input data
Which activation function can help avoid vanishing gradients?
ReLU
Why does the sigmoid function sometimes cause vanishing gradients?
Gradients become small
Which of these can be used as an output activation for binary classification?
sigmoid
What is the primary role of the output layer?
Transform hidden data into final predictions
When performing forward propagation in batch mode
Inputs are processed in mini-batches
What determines the number of parameters in a layer?
Number of neurons
Backpropagation is used for
Updating weights
The loss function in training neural networks measures:
Prediction error
If learning rate is too high, what might occur?
Oscillations
What is the vanishing gradient problem?
Gradients shrink to near-zero
Backpropagation algorithm updates weights in which order?
From output layer to input layer
The derivative of the loss with respect to weights is used to
Adjust weights for minimization
Which of these is not required for backpropagation?
random sampling
Why does a neural network need nonlinear activation functions like ReLU or tanh?
They prevent stacked layers from collapsing into one overall linear mapping
Which expression gives the mathematical form of forward propagation for a single neuron?
output = f(Wx + b), a weighted sum plus bias passed through an activation
What is the key difference between softmax and sigmoid activations?
Softmax normalizes scores across classes, while sigmoid squashes each value independently
How does forward propagation differ between the training and inference phases?
The layer computations match, but training also stores activations for the backward pass
Which sequence correctly summarizes one training step with backpropagation?
Forward pass, compute loss, backward pass for gradients, then update the weights
Select all statements that correctly contrast single-layer and multi-layer networks.
Select all true statements about weights and biases in a neuron.
Select all statements that are true of ReLU.
Select all statements that are true of the sigmoid activation.
Select all steps that occur in one mini-batch training iteration.
Select all true statements about the learning rate.
Select all correct pairings of a task with its standard loss function.
Select all true statements about backpropagation.
The key mathematical tool used by backpropagation to compute gradients through layers is the chain ____.
rule
A model that fits its training data almost perfectly but performs poorly on new data suffers from ____.
overfitting
The activation function whose output range is 0 to 1 is the ____ function.
sigmoid
The activation function whose output ranges from -1 to 1 is ____.
tanh
For any negative input value, ReLU outputs the number ____.
0
The standard output activation for multiclass classification, producing class probabilities that sum to 1, is ____.
softmax
A fully connected layer with 4 inputs and 3 neurons has a total of ____ parameters, counting both weights and biases.
15
the hyperparameter that controls the step size of each gradient descent update is the learning ____.
rate
The optimizer that combines momentum with per-parameter adaptive learning rates is ____.
Adam
When gradients shrink toward zero as they flow backward through many layers, this is called the ____ gradient problem.
vanishing
The abbreviation for the standard regression loss, mean squared error, is ____.
MSE
The subset of training examples used for a single weight update in network training is called a ____.
mini-batch