this is a question why you say insuffcient context? please attention: only answe

Question

this is a question why you say insuffcient context?

please attention: only answer question (c) and (d)

3. a) Derive the update rule for the weights in the output layer of a neural network using gradient descent rule. Assume that the sigmoid function is used as an activation function, the quadratic loss as the error function and L1 regularisation is applied. b) Assume the network's error function is Eo. How is it modified when L2 regularisation is applied? Describe how this type of regularization works and what is the difference with LI regularisation. c) Assume that you wish to train a classifier on a large dataset. How would you estimate its generalization performance and optimize its pararneters? Describe briefly the procedure that you would follow d) Compute the classification rate for the given confusion matrix. Do you think the classification rate is a suitable performance measure in this case? Explain your reasoning and the alternatives. Class 1 . Predicted Class 2 - Class 3 Class 1 - Actual 1000 Class 2 - Actual20 Class 3 - Actual Predicted 100 0 10 Predicted 50 10 0 10 e four parts carry, respectively, 40%, 20%, 20%, 20% of the marks.

Explanation / Answer

c.I would split the data set into a validation set and training set for example say 20 percent validation set and rest training set .train my model on the training set and test it on both validation set and training set.

Now i may get the following cases

1.training accuracy is low say less than 80 percent or something(depends upon problem) and so is validaton accuracy. this mean you have to train your model more or adjust parameters(number of nodes or layers(try increasing them)) to increase your accuracy.

2.training accuracy is high say close to 95 percent and validation accuracy is low say close to 80 percent this means your model is overfitting and not generalizing to unseen data then you need to employ regularization or similiar techiniques to reduce overfitting.

remember an ideal model would be one which gives you a good training and validation accuracy (>85 percent) and both accuracies are close to each other.

d.classification rate=true predictions/(all predictions)=(1000+0+0)/(1000+100+50+20+0+10+10+10+0)=0.83

this is an imbalanced data as there is a huge data of class 1 as compared to other two therefore classification rate is not not a suitable performance measure as it could be biased.ROC (Receiver Operating Characteristics) would be the most ideal performance metric in this case.

Related Questions

Navigate

• Browse (All)
• Browse T
• Subjects

---

---

Integrity-first tutoring: explanations and feedback only — we do not complete graded work. Learn more.