Scribd
Upload
1 views

ANN Lab Assignment

The lab assignment focuses on implementing and analyzing neural network algorithms, including perceptron training and backpropagation for various problems such as digit recognition and sentiment analysis of tweets. Students are required to perform feature engineering, utilize classification benchmarks like IRIS and MONK, and apply N-fold cross validation in their experiments. The assignment emphasizes studying convergence, local minima, and the effects of learning parameters on neural network performance.

Uploaded by

sparshdavra0795
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
1 views

ANN Lab Assignment

The lab assignment focuses on implementing and analyzing neural network algorithms, including perceptron training and backpropagation for various problems such as digit recognition and sentiment analysis of tweets. Students are required to perform feature engineering, utilize classification benchmarks like IRIS and MONK, and apply N-fold cross validation in their experiments. The assignment emphasizes studying convergence, local minima, and the effects of learning parameters on neural network performance.

Uploaded by

sparshdavra0795
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

Lab assignment: Neural Network

A. Basic assignments:
1. Implement perceptron training algorithm (PTA). Run it on NAND, XOR, 5-input
palindrome, 5-input majority, 5-input parity problems. Observe convergence when it
happens. Look for cycle when it does not. Record the convergence time (# iterations) as
function of the initialization point.
2. Give perceptrons for recognizing digits 0-9. Assume 7-segment display. Each perceptron
K (K=0…9) outputs 1 when K is input, else outputs 0.
3. Implement backpropagation (BP) on feedforward neural n/w (FFNN). Give FFNNs for all
the above problems including digit recognizer. Choose the learning rate judiciously.
Study convergence time, local minima, saturation, effect of initialization, effect of
learning rate and momentum factor. 1 and 0 decisions are based on the output being
above the high water mark or being below the low water mark.

B. Applications:
a. Give a neural network for recognizing the sentiments of tweets. Download tweets,
do feature engineering on them. A naïve feature vector is the set of words in the
tweets. Collect all the words in the tweets, sort them, remove duplicates. Each
tweet will be represented by a 1/0 vector depending on the presence/absence of
the word in the tweet. We will supply you with some sentiment marked tweets. You
will have to annotate some. All the annotated tweets will be used by all the groups.
b. Download any classification benchmark data from ML repositories (Look up, e.g.,
University of California at Irwine). Train and test FFNN on such data. Of particular
note is a classic problem called IRIS data. Read up on the internet. For any classifier
IRIS and MONK serve as benchmark data. You should surely show results on these
two data sets.
c. Apply FFNN to IR. Download TREC data sets. Apply the procedure of a. to classify
documents into relevant and irrelevant sets.

For part B, it is essential that you perform N-fold cross validation. Typically N is 5. That
means you divide your classification data into 5 partitions. Use 4 partitions for training the
neural n/w and the remaining for measuring accuracy.

You might also like