Appearance
Topic Drill 3.4.1: Structure and Working of a Deep Neural Network
About this drill
LO AI-3.4.1 (K2): Explain the structure and working of a deep neural network.
Lesson: Ch 3, section 3.4.1
This drill covers every point in section 3.4 / 3.4.1:
- The perceptron (binary, linearly separable); deep networks; fully connected networks = multi-layer perceptrons
- Input, hidden and output layers; neurons (nodes); increasingly abstract representations
- The activation value: weighted sum + bias → non-linear activation function; neuron bias vs. bias in 5.1.2
- Training: initialization, forward pass, loss, feeding the loss back to adjust weights and biases, epochs, stopping
Topic Drill: 3.4.1
Q1. In a deep neural network for a camera-based system, what does the INPUT layer do?
AIt adjusts the weights during training
BIt provides results to the outside world
CIt receives the inputs, for example pixel values from a camera
DIt calculates the loss
Q2. What are the hidden layers of a deep neural network made of, and what do they do?
AArtificial neurons (nodes) that compute and pass information on, transforming the input into increasingly abstract representations
BRules written by domain experts
CCopies of the training data
DActivation functions only, without weights
Q3. Which is an example the syllabus gives of what an OUTPUT layer might provide?
ARaw pixel values
BThe weights of the first layer
CThe number of epochs
DA value indicating the likelihood that the input image is of a cat
Q4. A neuron receives activation values 1.0 and 0.5 from two connected neurons, with weights 0.4 and 0.6. Its bias is 0.1. Following the syllabus description, what value is passed to its activation function?
A0.7
B0.8
C1.0
D2.5
Q5. How does the "bias" of a neuron relate to the bias discussed in section 5.1.2 (testing for bias)?
AThey are the same thing: a neuron's bias causes unfair treatment
BA high neuron bias always indicates discrimination
CThey are unrelated: a neuron's bias is an individual value added to its weighted sum, while 5.1.2 is about unfair differences in treatment based on sensitive attributes
DNeuron bias is only used in perceptrons
Q6. How are the weights and biases typically set at the START of training?
AInitialized to small random values (biases sometimes to zero)
BCopied from the test dataset
CSet to 1 for all connections
DChosen manually by the tester
Q7. What kind of problem can a single-layer perceptron solve, according to the syllabus?
AAny problem a deep neural network can solve
BClustering of unlabeled data
CGenerating new images
DSupervised learning of binary classifiers for linearly separable problems
Q8. What does the syllabus say fully connected networks can be considered to be?
ASingle-layer perceptrons
BMulti-layer perceptrons
CExpert systems
DRecurrent neural networks
Q9. Which TWO statements about training a deep neural network are CORRECT?
Select 2 options.
AEach pass through the training dataset is called an epoch
BThe activation functions are changed after every epoch
CThe weights are reset to random values after each batch of data
DTraining always stops after exactly one epoch
EThe error (loss) is fed back through the network to adjust the weights and biases
0 / 9 answered
← 3.3.1 Performance Metrics · Next drill: 3.4.3 Coverage Measures →