1

Week 3 Assignment

Week 3 Assignment

Sri Ram Kommana

Business Intelligence (ITS-531-A04)

Dr. Jason Hutcheson

University of the Cumberlands

May 21st, 2024

Week 3 Assignment

Chapter 5

Discussion Question

Question 1
Specifically, an Artificial Neural Network (ANN) is modeled based on the human

brain neurons configuration concept. It comprises layers of interconnected nodes (neurons)

where each node takes an input and formats it before passing it to the subsequent layer and

these layers on applying activation functions (Tabbussum & Dar, 2021). The training phase

involves tweaking the weights of the connections so that the output of the network is closer to

the target during training. This process, termed backpropagation , enables the ANN to

enhance its ability in classification and regression tasks.

The flexibility of ANNs is that they can solve diverse problems and perform functions

such as image and speech recognition, natural language processing, and predictive analysis

(Tabbussum & Dar, 2021). For instance in environmental monitoring where the aim is to

predict cases of flood, the ANNs works hand-in-hand to analyse data and develop ability to

predict flood event succeeded in proving its capacity in handling non linear relationship in

data. Because of this flexibility and performance, HL are helpful in various fields including

health, business, and finance, where predictive and classification abilities play a critical

function.

Question 2
Although there are some similarities as to their structure and functionality in that both

are neural networks, artificial neural networks (ANN) and biological neural networks have a

world of differences in terms of their complexity and functionality. Both concepts are

organized in a way that includes interconnected nodes (neurons) that work through layers and

where each node has an activation function applied to the information it receives (Tabbussum
3

Week 3 Assignment

& Dar, 2021). This similarity helps in ANNs to perform such processes like pattern

recognition and decision making, thus learning ability like human beings learn from past

experiences. However, they are abstract networks that do not fully emulate the complexities

of the biological neurons network.

The biological neural networks are vastly different from the those described above as

there are billions of neurons that are connected through synapsing which have a provision to

modify and alter with time, thus showing plasticity. They contain power to self-heal, build

new relations, and ANNs can have consciousness and feel emotions which ANNs cannot

(Tabbussum & Dar, 2021). Also, biological networks work at a much higher energy

utilization and network reliability as compared to the artificial networks in use today.

However, the basic concepts of learning and signal processing in these environments also

reveal the underlying common objective of mimicking intelligent systems.

Question 3
There are three most extensive types of artificial neural networks: Feed forward

neural networks (FNNs) convolutional neural networks (CNNs) and Recurrent neural

networks (RNNs). They are precisely the simplest FNNs, in which the data does not cycle

and moves solely in the forward direction from the input layer to the output layer, making

them appropriate for image or speech recognition and regression type problems (Tabbussum

& Dar, 2021). Structural diagramming tools can actually be widely used mainly because of

their simplicity and practicality in terms of managing structured data. CNNs are initiated for

the perceptions of predefined grid-like data structures or images. They also incorporate

convolutional layers to extract spatial hierarchies of features, making them suitable to work

on images and video feeds, budding objects, and detection, and classification (Tabbussum &

Dar, 2021). Due to their capability to model temporal and spatial dependencies in the data, it
4

Week 3 Assignment

is almost impossible to implement computer vision without the use of these transforms.

Meanwhile, RNNs have been designed specifically for processing sequential data, employing

loops that enable memory retention across the spatial dimension of time steps. Some

advantages of neural networks include that they are suitable for time series data, natural

language processing, speech recognition (Tabbussum & Dar, 2021). Obviously, RNNs are

able to work with inputs of different lengths more flexibly and are effective in cases when the

temporal context is important for making the correct prognosis.

Question 4
As mentioned, ANNs have two modes – supervised and unsupervised learning which

are employed in various processes and employ different techniques. In supervised learning

the ANNs are trained with the help of labeled data, whichmeans each of the input have an

associated output. The network learns by modifying weights in order to reduce the error

between the calculated value and the target labeled value and this is made possible through

the help of some algorithms such as back propagation (Sarkar & Zhang, 2021). This method

is very effective for most problems like classification such as figure two and regression where

the goal is to predict an output from input features and this can be applied in flood prediction

models where the ANN was trained with previous flood data with an aim of predicting future

events.

While in supervised learning, the ANNs are trained on the data by having labels that

guide the learning process, unsupervised learning entails training of the ANNs by feeding

them data with no labels. It tries to capture large scale features of the input data by learning

the probability distribution of the particular structures present in the network. Unlike

supervised learning, clustering and dimensionality reduction are popular methods in

unsupervised learning (Tabbussum & Dar, 2021). For instance, ANNs have their application
5

Week 3 Assignment

even in the field of environmental sciences including possibilities to categorize similar

weather conditions or to simplify climate data by reducing the number of attributes to

uncover other dependencies and structures that could be hard to detect. It is especially helpful

when one wants to examine the dataset and look for patterns in a general sense.

Exercise

Questions 6
Dealing with the comparison of machine learning methods in one or another problem

field is crucial for understanding the efficiency of the corresponding methods in particular

domains. There is no doubt that the contributions of both studies are notable especially when

their comparative analyses in resource provisioning in edge-cloud computing and feature

selection for data classification are considered.

The first paper by Duc et al (2019) provides a systematic literature review of the

machine learning techniques used for dependable resource allocation in edge-cloud

computing scenarios. The authors highlight the complexity of orchestrating edge-cloud

applications and categorize the problem into three main areas: Far before the elastic

application and remediation and the placement of components and system consolidation and

workload characterization and prediction. Both the authors describe numerous applications of

machine learning algorithms and discuss the merits of different types of algorithms in terms

of resource utilization factors. For instance, workload prediction utilizes approaches like the

neural network work load estimation and regression models to provide a correct estimation of

resources demands. In contrast, the enhancement of component placement and system

consolidation gains optimization algorithms and techniques for clustering that helps to

address resource allocation for managing considerable distributed systems (Duc et al. , 2019).

Finally, the paper discusses the above-mentioned challenges and outlines future suggestions
6

Week 3 Assignment

for further research, socio-economic implications, and the integration of more comprehensive

machine learning methods, which can improve the confidence level and real-time

performance of the edge-cloud computing framework.

The second paper by Chen et al. (2020) also relates to the process of data

classification and provides an approach to feature selection for classification purpose. The

study sets aside datasets that have a high number of variables as the basis for comparing the

several feature selection approaches that were explored, namely: Random Forest (RF),

Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Linear Discriminant

Analysis (LDA). The authors further reveal that Random Forest, especially when coupled

with various methods like varImp(), Boruta, and Recursive Feature Elimination (RFE), is

powerful in selecting significant attribute and further boosting the accuracy of a given model.

It shows how feature selection works to reduce the complexity of models; minimizes training

period, overfitting and generalize the curse of dimensionality thus enhancing the performance

of classifiers as illustrated by Chen et al. , 2020. Therefore, it is advisable to select the most

appropriate feature selection techniques to come up with an efficient approach to feature

selection for enhancing the performance of machine learning when analyzing large datasets.

As it can be seen in both papers, it is crucial to find proper algorithms of the machine

learning which should apply to the given problems. Duc et al. (2019) give an example of how

a proper method selection affects the resource management of the edge-cloud architecture,

while Chen et al. (2020) put under discussion the effect of feature selection for utilizing an

accurate and efficient classification. Although the two papers are applied in different

contexts, they share similarities, which include: Achieving optimized model performance and

the use of more elaborate Machine Learning procedures in solving huge problems. The

knowledge obtained from this paper is a great fit to draw the general picture on how machine

learning algorithms can be adopted in various fields to obtain accurate and optimal results.
7

Week 3 Assignment

Internet Exercise

Questions 7
The newly released Gee whiz examples on NeuroShell show the use of Neural

Network models for trading stocks using parameter settings that produced impressive annual

returns, as compared to buy and hold investment. For example, the Deere Neural Net

example suggests that this model generated 38. 1 percent profit within two years of training

while B&H generated 13. 2 percent only. Likewise, their specified model for training USG

Corporation reached a 202. 1 % return and a 21. 4% return on out-of-sample testing, which is

higher than the buy-and-hold strategy (NeuroShell, 2023).

It is therefore possible to parse the feasibility of these results based on few factors.

However, these neural networks can uncover patterns in the financial data that other

mechanical methods may not see, and therefore, might lead to more higher returns. However,

actual trading entails random distribution of other variables that govern the operations of the

market and therefore the past performance of one strategy cannot be used as a yardstick to

measure the performance of another strategy in the future. It is necessary to acknowledge that

the real performance of these neural networks to be deployed in real-life applications can be

affected by factors such as market influence; the quality of data; and possible overtraining on

past data (NeuroShell, 2023). Also, it is crucial that a trading strategy is fine-tuned and

constantly proven for its efficacy in changing market environments, which may entail a fair

amount of effort and challenging work.

Overall, it is remarkable to see the results that can be achieved by using NeuroShell

and the Gee Whiz examples presented in the present document are outstanding. The high

returns that may be realised can only be assumed under the best of environments and with no

external pressure, a reality that is quite different from theory, practice has always been
8

Week 3 Assignment

accompanied by a countless number of factors and possibilities that make or break any entity.

These tools should be used in concert with other tools like backtesting, regular recalibration,

and understanding of the various inputs and outputs as well as risks in using the NeuroShell

(NeuroShell, 2023).

Chapter 6

Discussion Question

Questions 1
Deep learning is that aspect of machine learning that involves deep learning neural

networks or deep learning artificial neural network to identify such representations from data.

This differs from previous machine learning methods which often involve hand-engineering

of features from raw input data because deep learning models can learn these features at

multiple levels simultaneously, based on the input data themselves (Janiesch, Zschech, &

Heinrich, 2021). This capability allows them to process large volumes of data, including

pictures, sound, words, etc because these inputs undergo a sequence of abstraction processes

to produce the desired output.

Deep learning is superior to more conventional methods which employ other

machine-learning algorithms in a number of ways. Over the last several years, it is preferred

to methodical patterns of distinct intricacy and quantity of high-dimensional datasets,

including image and voice identification, viz. natural language processing, and playing games

(Janiesch et al. , 2021). The traditional ways of implementing machine learning models are

not very effective when the data size is large and complex, while deep learning models on the

other side are more capable of handling large and complex datasets as the amount of data

increases, the performance of deep learning models increases as well. Moreover, such

techniques can identify fine-grained patterns by analyzing large, high-dimensional datasets,

Week 3 Assignment

which tend to be oblivious to traditional algorithms due to the reliance on hand-designed

features.

Questions 2
Supervised learning is one of the forms of learning in artificial intelligence, whereby

the computer is programmed to learn from the existing data set. Supervised learning: it is a

process where the AI model learns the relationship between an input and an output through

the use of a dataset that contains both results. This is a common technique used for

classification, and it is also useful in regression, where the model can develop a pattern and

can forecast based on the input data (Janiesch, Zschech & Heinrich, 2021). Some of them

include: Spam detection in emails, and identification of objects or things in images. In

contrast to this, unsupervised machine learning works with the data that has no predefined

classes and focuses on finding patterns or structures inherent in data sets. Some approaches,

common in this paradigm, are clustering and dimensionality reduction (Janiesch et al. , 2021).

This method is great for working in the exploratory data analysis, market basket analysis, and

in the customer segmentations, in which the data is not labeled in any way, yet the algorithm

needs to group or cluster the data naturally. Reinforcement learning is another type of AI that

focuses on the way an agent learns within an environment to gain maximum benefits or

rewards in the form of reinforcement. This method is especially useful in applications where

there is a process that involves making decisions in a step-wise manner, and has been widely

used in game AI, robotics, and self-driving cars (Janiesch et al. , 2021). Over the course of

many occasions, the practical guidelines of the particular construct are modified by the agent,

in the given case, in order to enhance one’s performance.

Week 3 Assignment

Questions 3
Representation learning is an aspect of machine learning in which the model identifies

the characteristics needed for a task from a raw data stream on its own. In contrast to feature-

based methods that require designers to select useful features by hand, representation learning

let the models find useful features from given data for enhancing the performance of the

dimensional learning modes, (Janiesch, Zchech & Heinrch, 2021). The strengths can be

illustrated through unstructured data like images, audio, and text since this type of big data is

easier to analyze using a cleverer approach rather than a brute-force approach.

Representation learning means a lot to not only machine learning, but deep learning

too. In machine learning, it improves the precision and quality of the model’s work by

ensuring and improving the features used in the model rather than manually creating features,

which makes it more efficient (Janiesch et al. , 2021). Hence for deep learning, representation

learning is crucial because deep neural networks are able to perform multiple levels of

abstraction starting from raw input data using the layers. This is useful in deep learning

models for advanced tasks such as image identification, voice transcriptions, and self-driving

cars where the approximate outcome of the relationships is very important.

Questions 4
Decision trees are a supervised learning algorithm in essence which incorporate

different types of decision based functions in order to produce non-linearity in neural

networks. The two major types of activation functions used in neural networks include: The

sigmoid function which delivers outputs only within the range of zero and one. This is

particularly valuable when working with binary classification tasks but it is known to have a

vanishing gradient problem, which makes it somewhat less helpful when dealing with deeper

architectures (Janiesch, Zschech, & Heinrich, 2021). The tanh function also outputs values in
11

Week 3 Assignment

the range [-1,1], and it typically results in higher performance when training than the sigmoid

function; however, it can also suffer from gradient issues.

The Rectified Linear Unit (ReLU) is another commonly used activation function

which gives an output equal to input if it is a positive value otherwise the output from the

ReLU activation function is 0 (Janiesch et al. , 2021). ReLU in particular contributes to

solving the vanishing gradient problem and increases the rate of convergence thus suitable for

deep networks. The softmax function which is explained in detail in the next section is widely

utilized in the output layer of the classification networks to map the scores of hypotheses into

probabilities so that the total probability equals to 1. These activation functions have very

important roles of increasing the performance as well as learning of the ANNs in all the tasks

they are designed for.

Questions 5
A Multilayer Perceptron (MLP) is a class of feedforward artificial neural networks

(ANNs) consisting of multiple layers of neurons: There are sharp differences in the

architecture and the number of layers as well as in the number of neurons for each layer, and

these include the input layer, one or more hidden layers, and the output layer. Every node in a

layer is connected to every node in the subsequent layers to pass information and enhance the

learning process and the ability to generate and identify complicated patterns from these

connections (Janiesch, Zschech, & Heinrich, 2021). MLPs are learning in the context of

Supervised learning and during the training phase the weights of the network are changed in

order to minimize the error of the output of the network, in order to classify input and give

accurate prediction.

The summation function in an MLP is generally of the form where each neuron sums

its inputs each weighted by its respective connection weight. This weighted sum, in turn, is
12

Week 3 Assignment

translated through an activation function, which is the key to introducing non-linearity and

determine intricate associations between the variables found in the data (Janiesch et al. ,

2021). Thus, the weights are very important because of the fact that each input contributes to

the output of the neuron and they undergo a process of training in an attempt to reduce the

difference between the two. This process of consecutive adjustment is called backpropagation

and it guarantees the MLP will learn the information inputted in it.

Exercise

Question 4
The application of cognitive computing is well exemplified by IBM Watson’s

performance in Jeopardy!problem-solving valid and effective, it is widely used in many fields

to solve multifaceted issues. In the healthcare sector, IBM Watson helps doctors, show

different amount of data, and come up with an accurate diagnosis so as to help the patient to

be treated by the right method thus helping in improving the quality of service that is being

offered Topol, 2019). This capability to process and interpret unstructured data offers a level

of insight into identifying desirable clinical trials and therapies. In the related financial field,

the application of cognitive computing increases the effectiveness of the fight against fraud

and the provision of risk management by conducting an analysis of transaction information to

identify similarities and differences that indicate fraudulent actions (Ghosh & Reilly, 2021).

This sophisticated approach enables a financial institution to reduce risks factor`s and

enhance secure environment and compliance with rules of the increased security level to

protect the asset. Furthermore, cognitive computing has transformed the customer service

field by allowing the use of virtual assistant and chatbots that are usually effective and

attentive (Huang & Rust, 202 1). These systems increase the value of customer satisfactions
13

Week 3 Assignment

and the companies’ performance by managing several questions, using the efficiency from

interactions, and evolving over time.

Week 3 Assignment

References
Chen, RC., Dewi, C., Huang, SW. et al. Selecting critical features for data classification

based on machine learning methods. J Big Data 7, 52 (2020).

https://doi.org/10.1186/s40537-020-00327-4

Duc, T., Leiva, R., Casari, P., & Östberg, P.-O. (2019). Machine Learning Methods for

Reliable Resource Provisioning in Edge-Cloud Computing: A Survey. ACM

Computing Surveys, 52(5), 1–39. https://doi.org/10.1145/3341145

Ghosh, S., & Reilly, D. L. (2021). "Credit card fraud detection with a neural network."

International Journal of Computer Science and Information Security, 14(9), 20-29.

https://doi.org/10.1016/j.cose.2020.102007

Huang, M. H., & Rust, R. T. (2021). "A strategic framework for artificial intelligence in

marketing." Journal of Marketing, 85(1), 30-54.

https://doi.org/10.1177/0022242920953847

Janiesch, C., Zschech, P., & Heinrich, K. (2021). Machine learning and deep learning.

Electronic Markets, 31(3), 685–695. https://doi.org/10.1007/s12525-021-00475-2

Tabbussum, R., & Dar, A. Q. (2021). Performance evaluation of artificial intelligence

paradigms—artificial neural networks, fuzzy logic, and adaptive neuro-fuzzy

inference system for flood prediction. Environmental Science and Pollution Research

International, 28(20), 25265–25282. https://doi.org/10.1007/s11356-021-12410-1

Topol, E. J. (2019). "High-performance medicine: The convergence of human and artificial

intelligence." Nature Medicine, 25(1), 44-56.

https://doi.org/10.1016/j.cell.2019.03.032