Scribd
Upload
45 views2 pages

Neural Network Classification Problems

The document contains a tutorial sheet with various exercises related to neural networks and perceptrons, including classification problems, decision boundaries, weight adjustments, and activation functions. It covers topics such as the design of neural networks, steepest descent methods, and backpropagation for weight updates. Each exercise requires the application of theoretical concepts to practical scenarios in neural network design and optimization.

Uploaded by

happy nain
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
45 views2 pages

Neural Network Classification Problems

The document contains a tutorial sheet with various exercises related to neural networks and perceptrons, including classification problems, decision boundaries, weight adjustments, and activation functions. It covers topics such as the design of neural networks, steepest descent methods, and backpropagation for weight updates. Each exercise requires the application of theoretical concepts to practical scenarios in neural network design and optimization.

Uploaded by

happy nain
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

Tutorial Sheet-1

1. Can a single neuron perceptron learn to correctly classify the following data, where each consists of
three binary input values and a binary classification value: (111,1), (110,1), (011,1), (010,0), (000,0)?
Either show that this is impossible or construct such a Perceptron.
2. A perceptron with single neuron uses a unipolar binary function, weights w1 = 0.2 and w2 = -0.5, and a threshold
= -0.2. The input vector is x = [0, 1]T and the desired output 0.6. Learning rate = 0.4. What is the equation of the
decision boundary after one iteration of PLR?
3. Consider a single sigmoid threshold unit with three inputs, x1, x2, and x3. The output y is
1
y = g ( w0 + w1 x1 + w2 x2 + w3 x3 ) where g ( z ) = . We input values of either 0 or 1 for each of
1 + e−z
these inputs. Assign values to weights w0, w1, w2 and w3 so that the output of the sigmoid unit is greater
than 0.5 if an only if (x1 AND x2) OR x3.
4. In the network shown, the output of each unit is a constant C multiplied by the weighted sum of inputs.
Can any function represented by this network also be represented by a single unit ANN (Perceptron)? If
so, draw the equivalent perceptron, showing the weights and activation function. Otherwise, explain
why not. _ _
4 + +
w1
X1
_ _ + +
w5 3
w2
2 _ _ _ _
w3 w6
_ _ _ _
w4 1
X2
1 2 3 4
5. Design a two layer neural network (with unipolar binary transfer function) to classify the region shown into
+ and – regions.
6. Consider the simple, single-input, single-output neural network shown in Figure 1 below. Assuming
sigmoidal hidden-unit and linear output-unit activation functions what values of the weights will
approximate the function in Figure 2?

w1 w5
w2
x
w3 w6

1 w7
w4

Figure 1 1 Figure 2
7. In the figure shown below, three lines passing through the origin split the x-y plane into six regions. Is it
possible to design a two layer network with single neuron in the output layer which classifies the shaded
regions from the other regions? Activation functions of the neurons are unipolar binary activation functions
(hardlim). If possible, then design the network. If not, describe why not in one or two sentences

-1
8. Apply steepest descent method to the following function F(x) and find the value of x after one iteration. The
starting value of x is − 2 1T and the learning rate is 0.05
F ( x) = x12 + 4 x1 x2 + 20x2 2 + x1 + x2 + 1
Will this algorithm converge to a solution or not? Justify your answer.

9. What is the maximum stable learning rate for the function


F ( x) = 5x12 − 6 x1 x2 + 5x2 2 + 4 x1 + 4 x2 Sketch a contour plot of the function.

Consider a function F ( x) = x1 + x2 . Find the derivative of the function at the point 0.5 0.5 in the direction of
2 2 t

a vector p = − 3 4 . Also find the second derivative along the vector p.


t

10. The following network uses logsig activation functions in the hidden layer and purelin in the output layer.
The initial weights are shown in figure. Training samples are x=1, t1=0.8 and t2= 0.6. Use backpropagation
to determine the hidden layer weights after one epoch. Assume a learning rate of 0.8

1 w3=0.1 3
y1
w1=0.5
x

w2=0.3 w5=0.2 w4=0.4


w6=0.1
y2
2 4

11. For the neural network shown, the hidden layer neurons have tansig activation functions. The output unit is
linear. The training set is shown in the table. The initial values of all weights are 0.5. Compute the value of
weight c after the first update, using backpropagation in batch mode, with a learning constant of 0.1

12. Figure shows a two-layer neural network. Activation functions of the hidden layer neurons are tansig and
output layer are linear (identity function). Learning rate is 0.4, w1=w2=w3=w5=0.2, w4=w6=w8=w10=0.1,
w7=0.5, w9=0.4. Find the modified error of neuron 4. Find the weight w4 after applying backpropagation
for one epoch. (Penalty for showing steps which are not required)

Common questions

Powered by AI

To ensure a single sigmoid outputs greater than 0.5 for (x1 AND x2) OR x3, set weights such that w0 = -1.5, w1 = 1, w2 = 1, w3 = 1. For input condition (1, 1, 0), the weighted sum 1 + 1 - 1.5 = 0.5, yields sigmoid output exceeding 0.5. Also, for (0, 0, 1), it becomes 1 - 1.5 = -0.5, ensuring total output exceeds 0.5 only for true logical combinations .

For backpropagation with logsig activations and linear outputs, forward propagate to compute outputs, calculate error per output, distribute deltas backward per activation derivatives derived from logsig characteristics. Update follows error-weighted derivative traces correcting weights inversely to reduce output error markedly. Through sample iteration with epoch setting, focus corrections specific to task increased output accuracy .

To approximate a target function with given constraints, synaptic weights need to modulate input towards sigmoid outputs reflecting the target shape. Small, positive weights ensure minute variations in sigmoid outputs, fine-tuning outputs aggregated by weights to linearize overall behavior per target expectations .

Factors include: 1) configuration and capabilities of the hidden layer to remap inputs onto linearly separating decision planes, 2) ability to match shading specifications with weights enabling boundary placement, 3) correct thresholding by hardlim function aligned with partition demands. If concealed separations exceed boundaries hidden layers generate, the task is computationally challenging .

After one iteration of PLR with a learning rate of 0.4, the weight update ∆w can be determined as ∆w = learning_rate * (desired_output - actual_output) * input. Given input [0, 1], actual_output = -0.5 * 1 + (-0.2), and desired_output = 0.6, the adjustment for w2 becomes ∆w2 = 0.4 * (0.6 - (-0.5)) * 1 = 0.44, therefore new w2 = -0.5 + 0.44 = -0.06. The decision boundary is 0.2x1 - 0.06x2 - 0.2 = 0 .

Design the first layer to map input data onto a new feature space where non-linear boundaries can become linear. Suppose weights for hidden units are adjusted based on data distribution; then choose output weights such that summation points align with region boundaries. Transfer function thresholds should strategically align with boundaries, allowing the network to differentiate between + and - zones .

Convergence depends on function curvature and step magnitude: rapid convergence requires well-behaved quadratic curvature and adequately chosen step sizes. Given F(x) containing quadratic terms susceptible to steep local descents, with an initial point optimal via derivative convergence per parabolic logic, convergence is expected provided gradients maintain consistent decrease .

A single neuron perceptron can classify linearly separable data. For the specified points, construct a weight vector w = [w1, w2, w3, b] where w1, w2, w3 represent the weights for inputs and b is the bias. Through experimentation, w1 = 1, w2 = 1, w3 = -1, and b = -1 can correctly classify the given dataset. This shows that the data is linearly separable, aligning with perceptron capabilities .

Compute ∇F([0.5, 0.5]) = [dF/dx1, dF/dx2] = [1.0, 1.0] and directional derivative via −3x1 + 4x2 = 1.0(−3) + 1.0(4) = 1. Anticipate second derivative challenges as directional variations amplify changing slope extensions into steep isolated changes, adding complexity when linearity assumptions are surpassed .

A function represented by a constant C multiplied by the weighted sum of inputs can be equivalent to a function computed by a single perceptron only if it is linearly separable. If the network involves non-linear separability, a single perceptron cannot replicate it due to its limitations on linear decision boundaries .

You might also like