Project #1: Artificial Neural NetworksAmir SadovnikCOSC 424/525: Deep Learning (Spring 2023)1 OverviewIn this project you will be implementing a simple artificial neural network (ANN) library....

1 answer below »
all parts need to be completed


Project #1: Artificial Neural Networks Amir Sadovnik COSC 424/525: Deep Learning (Spring 2023) 1 Overview In this project you will be implementing a simple artificial neural network (ANN) library. I will ask you to implement it in a specific object oriented manner. Although this will make the library less efficient it will allow you to gain a better understanding of how ANN’s predict and learn. Make sure to use your own solution and your own code. 2 Problem Description Your are tasked with writing the following functions: 1. Write a Neuron class with the following properties: (a) Should be initialized with an activation function, the number of inputs, the learn- ing rate, and possibly a vector of weights (if not, set to random). (b) Treat bias as an additional weight - where the input to that weight is set to 1 (that is the weight vector’s length could simply be |w|+ 1). (c) Needs to store the the input, output and partial derivative for each of its weights for back-propagation. (d) Should have an activate method which given a value returns its value after activation (depending on the activation function used). (e) Should have a calculate method which given an input calculates the output. (f) Should have a activationderivative method which returns the derivative of the activation function with respect to the net. (g) Should have a calcpartialderivative method which calculates the partial deriva- tive with respect to each weight (using the derivative method) and returns the vector w × δ. 1 (h) Should have a updateweights which actually changes the weights using the par- tial derivative values (and the learning rate). 2. Write a FullyConnectedLayer class. (a) Should be initialized with the number of neurons in the layer, the activation function for all the neurons in the layer, the number of inputs, the learning rate, and possibly a vector of weights (if not set to random). (b) Note: Neurons should be created once in the FullyConnectedLayer constructor and not each time you call calculate. That way the neurons can store the input and output to be used in backpropagation. (c) Should have a calculate method which given an input calculates the output of all the neurons in the layer. (d) Should have a calculatewdeltas which given the ∑ w × δ from the next layer, goes through all the neurons in the layer to calculate their partial derivative (using the calcpartialderivative method), updates all the weights (using the updateweights method) and returns its own ∑ w × δ. 3. Write a NeuralNetwork class. (a) Should be initialized with the number of layers, number of neurons in each layer, a vector of activation functions for each layer, the number of inputs to the network, the loss function used, the learning rate, and possibly a vector of weights (if not set to random). (b) Should have a calculate method which given an input calculates the output of the network. (c) Should have a calculateloss method which given an input and desired output calculates the loss. (d) Should have a lossderivative method which returns the value of the loss deriva- tive (depending on the loss). (e) Should have a train method, which given a single input and desired output takes one step of gradient descent. This method should first do a forward pass (give its own method (f) Any additional methods/instance variables needed for back-propagation calculate, then calculate the derivative of the loss (using lossderivative) and finally go through the layers backwards calling calculatewdeltas for each and passing∑ w × δ. as it goes. Note that the first time calculatewdeltas is called it should be passed the derivative loss. 4. Your main method should take the following command line argument to show the following: (a) The first argument should state the learning rate. 2 (b) The second argument should state the type of problem to train on (for COSC424 you only need to implement the first one): i. If given example, simply run a single step of back-propagation using the example we did in class. Print out the weights after the 1-step update to show that you have gotten the same results as we did in class. ii. (only COSC525) If given and should train a single perceptron to produce the output of the “and” logic gate. Train it for enough steps so that it converges. Show the results of the prediction for all 4 possible inputs. iii. (only COSC525) If given xor should train two types of networks to produce the output of the “xor” logic gate. One is a single perceptron, while the other adds a hidden layer. Train it for enough steps so that it converges. Show the results of the prediction for all 4 possible inputs for each of them. 3 Additional Information You must do so under the following constraints: 1. Your library should support at least two activation functions: logistic and linear 2. Your library should support at least two loss functions: square error and binary cross entropy loss (for COSC424 you just need to support square error). 3. You must use Python3 with only the numpy and sys libraries allowed 4. Make sure to comment your code 5. Work plan tip: Start with implementing a feed-forward network. First implement the neuron methods for feed forward, then the fullyconnectedlayer for feed forward, and finally neuralnetwork for feed forward. For each of them you can do some manual calculations to ensure you are getting the correct results. Only then start working on back-propagation. Do so in the same order checking manually along the way. 4 Report You should submit a short PDF report with the following (if you do not have anything to add for a section, simply put the section title and then state there is nothing to add): 1. A short introduction to the problem. 2. Assumptions/choices you have made. 3. Problems/Issues you were not able to solve. 4. How to run you code (if there is any difference from what is stated) 3 5. Show a graph of the loss dropping as a function of learning for different learning rates (you can choose one of the example setups). Try to find learning rates that are too low (Do not worry about finding a learning rate that is too high - this error function is a bit weird). Describe what you see in the graph in words. 5 Submission You are required to submit 3 files: 1. pdf file for the report 2. pdf file for the code 3. A zip file with your code. 4 Overview Problem Description Additional Information Report Submission
Answered 4 days AfterFeb 02, 2023

Answer To: Project #1: Artificial Neural NetworksAmir SadovnikCOSC 424/525: Deep Learning (Spring 2023)1...

Aditi answered on Feb 05 2023
39 Votes
Introduction
This project implements a simple artificial neural network (ANN) library. The library is object-oriented, with three classes, namely Neuron, FullyConnectedLayer, an
d NeuralNetwork, to provide three abstraction levels for an ANN:
An array of FullyConnectedLayer particles makes up a NeuralNetwork object.
An array of Neuron objects is contained in each FullyConnectedLayer object.
The library performs a forward pass by determining the neural network output and a weights update utilizing gradient descent in each iteration.
1. Assumptions/choices made
· A neuron's bias is treated with just an extra input data of 1. As a result, the user must ensure that each input array contains an extra input value of 1.
· Only when the output ranges from 0 to 1 can the binary cross entropy loss function be used. When the activation function used (e.g., linear function) doesn't really guarantee that the output value is between 0 and 1, an error may occur.
· This library can currently support two convolution layers (linear/logistic) as well as two loss functions.
2. Problems/Issues
Nothing to add.
3. How to run the code
You can run the code with examples using the following command:
$ python pro je ct 1. py [ example | and | xor ]
You can choose to run either the class example problem, the AND problem or the XOR problem.
4. Results
5.1. The class example problem. One step backpropagation Number of hidden layer: 1
Activation function: Sigmoid Loss function: Squared error Learning rate: 0.5
Number of iteration: 1
The weights after the 1-step update:
w1 = 0.14978732, w2 = 0.19957463, b1 = 0.34574630
w3 = 0.24975842, w4 = 0.29951684, b2 = 0.34516843
w5 = 0.35895608, w6 = 0.40870723, b3 = 0.53075162
w7 = 0.51181741, w8 = 0.56188906, b4 =...
SOLUTION.PDF

Answer To This Question Is Available To Download

Related Questions & Answers

More Questions »

Submit New Assignment

Copy and Paste Your Assignment Here