0% found this document useful (0 votes)
9 views4 pages

Problems

The document presents a set of numerical practice problems related to artificial neural networks, covering topics such as single perceptrons, multi-layer perceptrons, error gradient calculations, and various activation functions. Each problem includes specific inputs, weights, biases, and the required calculations, with final answers provided for outputs and errors. The problems aim to enhance understanding of forward and backpropagation processes in neural networks.

Uploaded by

letsthinkk.tank
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
9 views4 pages

Problems

The document presents a set of numerical practice problems related to artificial neural networks, covering topics such as single perceptrons, multi-layer perceptrons, error gradient calculations, and various activation functions. Each problem includes specific inputs, weights, biases, and the required calculations, with final answers provided for outputs and errors. The problems aim to enhance understanding of forward and backpropagation processes in neural networks.

Uploaded by

letsthinkk.tank
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

Artificial Neural Networks: Forward and Backpropagation Practice Set

Numerical Practice Problems for 4th Semester Curriculum

Problem 1: Single Perceptron with Sigmoid Activation


Consider a single perceptron with two inputs x1 = 0.6 and x2 = 0.4. The initial weights are w1 = 0.5 and
w2 = −0.3, with a bias b = 0.2. The activation function is the Sigmoid function σ (z) = 1+e1 −z . Calculate the
final output y and the Mean Squared Error E = 12 (t − y)2 given a target output t = 1.

x1
w1
=0
.5 b = 0.2
y
∑ |σ
.3
= −0
w 2

x2

Final Answer:
Output y ≈ 0.5939
Error E ≈ 0.0825

Problem 2: Error Gradient Calculation


For the network below, the output neuron uses a Sigmoid activation. Input x = 0.8, weight w = 0.5, bias
b = −0.1, and the target t = 0. Using the loss function E = 12 (t − y)2 , find the gradient of the error with
respect to the weight, ∂∂ Ew , and calculate the updated weight after one step of Gradient Descent with learning
rate η = 0.1.

b = −0.1

w = 0.5 y
x = 0.8 σ

Final Answer:
Gradient ∂∂ Ew ≈ 0.1123
Updated weight w ≈ 0.4888

1
Problem 3: Multi-Layer Perceptron (MLP) Forward Pass
A neural network has a 2-2-1 architecture. All hidden and output nodes use the ReLU activation f (z) =
max(0, z). Inputs: x1 = 1.0, x2 = −1.0.
Weights (Input to Hidden): w13 = 0.4, w23 = 0.6, w14 = −0.2, w24 = −0.5.
Biases (Hidden): b3 = 0.5, b4 = 0.1.
Weights (Hidden to Output): w35 = 0.8, w45 = 0.9.
Bias (Output): b5 = −0.2.
Calculate the final output y of the network.
w13
X1 3
w3
5

w 45
X2 w24 4

Final Answer:
Output y = 0.04

Problem 4: Hidden Layer Delta Calculation


In a network using Sigmoid activation for all nodes, the error term (delta) at the output node O is given as
δO = 0.15. The weights from the two hidden nodes to the output node are w1O = 0.5 and w2O = −0.4. The
outputs of the hidden nodes during the forward pass were h1 = 0.7 and h2 = 0.3. Calculate the error term δ1
for hidden node 1.

h1
w1O
= 0.
5

O δO = 0.15

0.4
=−
w 2O
h2

Final Answer:
δ1 = 0.01575

2
Problem 5: Full Forward Pass with Biases (Matrix Representation)
Calculate the final output y for the following network using Sigmoid activation for all nodes. Inputs: X1 =
0.5, X2 = 0.8.  
(1) 0.2 −0.1 (1)
Weights from input to hidden layer W = (where Wi j is the weight from input i to hidden
0.4 0.3
j), b(1) = [0.1, −0.2].
Weights from hidden to output W (2) = [0.5, −0.5], b(2) = 0.2.
(1)
W11
X1 H1
W (2)
1
W (
12 1)

)
O
(1
W 21

(2)
W2
X2 H2
(1)
W22

Final Answer:
Output y ≈ 0.5658

Problem 6: ReLU Dead Neuron Scenario


Consider a single neuron with three inputs x1 = 2, x2 = −3, x3 = 1 and weights w1 = 0.2, w2 = 0.5, w3 =
−0.1. The bias is b = 0. Calculate the output y using the ReLU activation. If the target is t = 1 and learning
rate η = 0.1, what is the updated weight for w2 ?
x1
w1

w2 y
x2 ReLU

w3
x3

Final Answer:
Output y = 0
Updated weight w2 = 0.5 (Gradient is 0 due to inactive ReLU)

3
Problem 7: Cross-Entropy Loss Calculation
A network with a 2-2-1 architecture produces a final predicted probability y = 0.8 at the output node (using
a Sigmoid activation). If the true target label is t = 0, calculate the Cross-Entropy Loss E = −[t ln y + (1 −
t) ln(1 − y)].

Final Answer:
Error E ≈ 1.6094

Problem 8: Linear Perceptron Update


A neuron with a linear activation function (identity) has inputs x1 = 0.5, x2 = 0.5 with weights w1 = 0.4, w2 =
0.4 and bias b = −0.2. Given a target t = 1.0 and learning rate η = 0.2, calculate the updated weight w1
using Stochastic Gradient Descent. Loss function E = 12 (y − t)2 .

Final Answer:
Updated weight w1 = 0.48

Problem 9: Sigmoid Derivative Evaluation


During backpropagation, the weighted sum at a hidden node is found to be z = 1.5. Evaluate the output of
the node h = σ (z) and the derivative of the activation function with respect to z, given by σ ′ (z).

Final Answer:
h ≈ 0.8176
σ ′ (z) ≈ 0.1491

Problem 10: Tanh Activation Forward Pass


z −e−z
A hidden neuron uses the hyperbolic tangent activation function tanh(z) = eez +e −z . It receives inputs x1 =

0.8, x2 = −0.4 with respective weights w1 = 1.5, w2 = 2.0. The bias is b = 0.1. Calculate the final output h
of this neuron.

x1
w1 =
1.5 b = 0.1

h
tanh
2.0
w2 =
x2

Final Answer:
Output h ≈ 0.4621

You might also like