1

Week 3 Assignment

Week 3 Assignment

Navya Gurijala

University of the Cumberland's

Business Intelligence

ITS-531-A01

Dr. Abiodun Adeleke (Abbey)

05/21/2024
2

Week 3 Assignment

Chapter 5

Discussion Question

Question 1
Artificial neural network abbreviated as ANNs is a computational based intelligence

model which is based on the structure of the brain. It contains the status of layers of

networked computing components known as neurons that in turn analyze data and make

decisions. It takes an input, applies the weights to these inputs, arrays it in the required form,

and then applies an activation function to get an output (Shen et al. , 2021). This structure

makes it possible for ANNs to bring out patterns and dependencies in data that would be

impossible for a programmer to see or quantify; hence, they work great for applications such

as image and speech recognition, natural language processing, and predictive modeling.

ANNs are most effective in computation-heavy tasks and for problems requiring data

classification or pattern detection. For instance, they can perform better than traditional

algorithms especially in image or voice recognition where the algorithms have high

variabilities and complexities of the data (Shen, Wang, & Navlakha, 2021). Further, ANNs

are applied to medical diagnostics, financial prediction, and in controllers for self-acting

systems to discern patterns and relationships obscured to man or profit by linear methods.

Question 2
ANNs and BNNs originate from the same idea but differ by the complexity and the

capability of the networks. Both kinds of networks contain nodes (neurons) which in turn,

perform computations across layers; each neuron accepts inputs, performs operations, and

then forwards the result to the following layer (Shen et al., 2021). This similarity enables

ANNs to work on tasks like pattern identification and decision making, like the brain does.
3

Week 3 Assignment

Nonetheless, ANNs are strikingly reduced versions of the biological neurons, with which

they are intended to be comparable.

ANNs are much simpler with only a few layers as compared to the billions of neurons

that exist in biological neural networks that are also dynamic in terms of connectivity (Shen,

Wang & Navlakha 2021). Some of these include the brain’s healing mechanisms,

neuroplasticity, consciousness, and capability to generate emotions are not reproduced in

artificial networks. Nevertheless, both types of networks employ quite similar strategies of

learning and signal processing which underlines the fact that these Networks are built based

on the same principles.

Question 3
Some of the most widely used ANNs are Classified into feedforward neural networks

(FNNs, CNNs, and Recurrent neural networks (RNNs). The FNNs are the basic structure;

data flows in a unidirectional manner starting from input to the output without any feedback

loop (Shen et al. , 2021). They are generally utilized to solve tasks such as face and speech

recognition tasks where the input data maybe in the form of blacks and whites and do not

require data in the future.

CNNs are particularly well-suited for filtering highly structured grid data such as

images through convolutional layers where spatial hierarchies of features are learned in an

adaptive manner (Shen, Wang, & Navlakha, 2021). This makes them highly suitable for

scenarios, for instance, in image and video analysis. Specifically, some types of Recurrent

Neural Networks (RNNs) are designed specifically for sequential data due to their ability to

contain loops whereby information can be retained (Shen, Wang, & Navlakha, 2021). Most

often they are applied in applications of natural language processing and time series

forecasting techniques where the order of data and context matters.

Week 3 Assignment

Question 4
The typical learning model in ANNs is the supervised mode in which text training

data is labelled in that for every given input there must exist a specific output. The

backpropagation of error is used to adapt the weights of the network during training, such that

the differences between the estimated results and actual labels can be minimized, usually with

Popular gradients descent techniques (Shen, Wang & Navlakha, 2021). This training

continues back and forth and helps to minimize on some error allowing the ANN to perform

well on new unseen data. In addition, the most effective type of machine learning is applied

for tasks such as image classification, speech recognition, and, to an extent, medical

diagnosis.

In the unsupervised learning, the ANNs operate on data wherein there is no provision

for output data to map against to this data and the ANN seeks to determine the underlying

structure. Clustering and dimensionality reduction are common approaches utilized in the

analysis (Shen, Wang, & Navlakha, 2021). The classification occurs and new patterns are

found making the unsupervised learning even more effective in chores such as detection of

outliers, categorizing customers, and data compression chores. Unlike supervised learning,

this approach assist in identifying hidden data structures without previous suspension of

classes.

Exercise

Questions 6
The last few advancements in machine learning have all been instrumental in

responding to numerous difficulties in mainstream science. It can thus be desirably

informative to compare and contrast the approaches and results of two similar research
5

Week 3 Assignment

studies, Dou et al. , 2023 and Psaros et al. , 2023, which present reviews and surveys of

existing literature on machine learning with focus on various specific problem domains.

The paper by Dou et al. (2023) is devoted to the problems with such a kind of data as

small data in molecular science and engineering. It deals with the popular and fundamental

set of machine learning algorithms including linear regression, support vector machines

(SVM), random forest (RF), convolutional neural networks (CNN), and generative

adversarial networks (GAN) amongst others. The work also highlights the possibility of

approaching the investigated issues in a twofold manner: by discussing the applicability of

the discrete, traditional machine learning techniques and the more elaborate deep learning

methods for high dimensional data, noises, and small size imbalanced databases. The authors

note major advances in the methodological domain towards combining both sophisticated

machine learning strategies with conventional solutions for tackling small data, thereby

enhancing the efficiency of scientific discovery tasks in chemical and biological sciences

(Dou et al. , 2023).

On the other hand, Psaros et al. (2023) present an overview on the issue that indicates

how uncertainty quantification (UQ) can be addressed in scientific machine learning based on

neural networks. They briefly explore why quantifying of uncertainties by aspects such as

noisy data, limited data, hyperparameter tuning, and misspecified models is challenging. The

given study to encompasses the different UQ methods and assessments it proposes an

elaborate set of comparative studies on the prototype problems to exemplify how these

approaches can be used. As it was mentioned by Psaros et al. (2023) the neural networks are

also much helpful in the integration of the data with the mathematical laws in physics and

engineering but the consequences of the uncertainties remain quite difficult to measure and

which can only be estimated by a systematic approach.

Week 3 Assignment

The two works also emphasize the need for the development of appropriate methods

to match the aims of machine learning and problem domains. Dou et al. (2023) concentrate

themselves on small data challenges, suggesting that it is possible and efficient to adopt

integrated strategies that include not only the conventional machine learning methods, but

also the AMMs to cope with data scarcity more efficiently. However, it is necessary to

recognize that more extensive UQ frameworks have to be employed to increase the

confidence level of the methods based on the use of neural networks in various scientific

practices, according to Psaros et al. (2023). Além disso, ambas as investigações apresentam

importantes lições para melhorar a eficiência e a integridade das técnicas de aprendizado de

máquina na varie-Dade científica em geral.

Internet Exercise

Questions 7
On its website, NeuroShell gives several examples of its applying the facility in the

sphere of financial forecasting and in the variety of other situations; the model seems to work

well in all the cases. For instance, the ‘Deere Neural Net’ provides an example of a neural

network that has been trained and tested using data relating to Deere (DE) over two years

where the return on trades was 38. 1% compare to the purchase and hold at 13. 2% in the

similar period. This model included basic percentage indicators, such as Averages Directional

Movement Plus DI (+DI), Commodity Channel Index CCI, Relative Strength Index RSI, and

stochastic %K: this approach demonstrated commendable out-of-sample results (NeuroShell,

2023).

One more example of it’s application was a “USG Corporation” where a free-time

neural network was designed to determine the probability of a buy/sell signal instead of the

signal itself. Using training data the proposed this model delivered a return of 202. 1%
7

Week 3 Assignment

against the mere 184. 8% return provided by the buy-and-hold strategy. The out-of-sample

period has also a great return of 21. 4% in average as compared to the buy-and-hold

technique with an average return of 13. 8% (NeuroShell, 2023). These examples indicate that

the concept of trading and giving a potential signal of a neural network model of trading

opportunities is feasible as well as providing an adequate possibility of creating significant

revenue. Yet, reaching the similar results in realistic setting may be fraught with a need to

adjust the model and test it for variation to obtain the most accurate and viable solution in the

context of the marketplace.

The possibility of replicating such research is highly reliant on such factors as quality

of the data used as well as the choice of appropriate indexes that are to be used in the given

context of the market. These applicative examples in NeuroShell prove that neural networks

are a useful method in predicting prices in the stock market; however, some problems not

present in this simulation of the financial market which might affect real data like noise,

overfitting and the dynamic change of market behavior. However, the use of neural networks

has, with the right setting and calibration, the potential to be an invaluable resource for

financial and trading analysis.

Chapter 6

Discussion Question

Questions 1
Convolutional neural network is a type of deep learning which is defined by

avoidance of using simple feedforward networks and the utilization of deep artificial neural

networks. Such could afford to learn data representations that are hierarchical simply by

transforming input data through one or more layers of abstraction (Coenen, Verbeke, & Guns,
8

Week 3 Assignment

2022). This capability of portraying intricate structural attributes and dependencies of large

data sets and hence, the key characteristic that differentiates deep learning from conventional

machine learning techniques that require distinct features to be identified and simpler models

to be trained on.

Deep learning is the new and advanced level of machine learning as it enjoys success

even when it is faced with unstructured data like image, audio, and text. It performs

extraordinarily well in context such as image and speech recognition, natural language

processing, and playing video games whereby the traditional methods are quite unfit due to

the problems associated with high dimensionality and nonlinearity of data. On the other hand,

deep learning models can also enhance their performance once they receive more information

while traditional methods usually have their learning rate reduced and reach a performance

ceiling quicker, thus stressing on the applicability and versatility of DLMs in various

domains.

Questions 2
In the current artificial intelligence regime, there are many learning paradigms that

they use in the system, and among them are the supervised learning, unsupervised learning,

and reinforcement learning. Supervised learning can be described as a machine learning

technique that focuses on supervised data where the input-output pairs are used to train the

model in order to reach a certain hypothesis on the prediction of the unseen data. It can be

employed in many intelligent tasks including, but not limited to, classification and regression.

This model continuously updates the model complexes thus is useful where there are data that

has already been seen and have known values.

Conversely, unsupervised learning is the type of learning that falls under the category

of classification learning since it works with unlabeled data with the primary goal of finding
9

Week 3 Assignment

hidden structure in the input. The latter include clustering and dimensionality reduction

among other techniques (Coenen, Verbeke, & Guns, 2022). It is useful for open-ended

modeling and data mining, when there are no clearly defined classes and the normal goal is to

discover the underlying distribution of the data for such applications as anomaly detection

and customer grouping. The other learning techniques are reinforcement learning, where the

agent is involved in learning decision-making processes by for estimation of cumulative

utility.

Questions 3
Reproduction learning is a paradigm of the machine learning algorithm that aims at

identifying representations that define the data for the purposes of a certain task in the

process of learning from raw data. Compared to the previous approaches to feature

representation, the current approach enables the model to learn the feature representations

directly from the input data, and reduce the need for human (expert) intervention in this

process, which leads to the better performance of models (Coenen, Verbeke & Guns, 2022).

This process helps in identifying and mapping the complex structures involved in data

relationships, which makes it vital for operations that deal with many dimensions.

Representation learning is strongly linked with deep learning since deep neural

networks work perfectly for the extraction of hierarchical representations that are achieved

through multiple layers of abstraction. .Every layer in a deep learning model operates on an

input data and transforms it in a way that higher layers get to derive sophisticated and

important features in the data, which makes deep learning models suitable for solving

complicated problems like image and voice recognition (Coenen, Verbeke, & Guns, 2022). It

is, therefore, vital to uncover meaningful features by utilizing representation learning in deep
10

Week 3 Assignment

learning models since it can lead to automating or improving the entire feature engineering

course, greatly enhancing the efficiency of machine learning.

Questions 4
ANNs employ many activation methods to introduce non-linearities within the model

so as to enable them get a handle on the data patterns. Some of the most widely used

activation functions are S-activation function or sigmoid function, hyperbolic tangent

function or tanh function, Rectified Linear Unit or ReLU, and Softmax functions. It should be

noted that due to its nature, the sigmoid function returns outputs in the range [0, 1], which

makes it particularly useful in binary classification tasks; In terms of disadvantages, it has

been shown that sigmoid activation can lead to issues with vanishing gradients (Coenen,

Verbeke, & Guns, 2022). The function tanh maps input values in the range (-∞, ∞) to the

range (-1, 1), and it can also give better training performance than sigmoid, though it can be

still susceptible to vanishing gradients.

The ReLU- Rectified Linear Unit function is used frequently because of its

effectiveness: it returns the object as is if it is a positive number and returns zero otherwise

(Coenen et al. , 2022). These do help in reducing the vanishing gradient problem, and also

help in achieving faster convergence of the network. The softmax activation function is used

commonly in the output layer of the classification neural networks The softmax function

fixes the outputs between 0 and 1 and normalizes them if computing scores for multiple

classes. All of these function are essential in allowing ANNs to capture relationships in the

data while being efficient and poised to converge.

Questions 5
A multilayer perceptron (MLP) as an ANN architecture model comprises an input

layer, one or several hidden layers, and an output layer. Each layer of the network is made up
11

Week 3 Assignment

of neurons and these neurons are connected and structured in a way that allows the network

to learn more enhanced patterns of compression (Coenen, Verbeke, & Guns, 2022). MLPs are

widely used because of their high ability when it comes to classification and regression since

they have multiple layers that allow them to create non-linear representation of data.

In MLPs, the summation function sum up the another activity of inputs of the weight

for every neuron. This sum is further processed through the activation function, which makes

the model non-linear and helps it learn such relations that may exist in the data (Coenen,

Verbeke, & Guns, 2022). During training process, the weights are developed to reduce the

differences between the outputs calculated by the model and the targeted outputs. Thus,

backpropagation allows the MLP to learn and adapt the weight values of its connections for

subsequent operations and become a worthy instrument in a broad spectrum of the machine

learning field.

Exercise

Question 4
Cognitive computing, as witnessed by a recent ‘war’ of IBM Watson against people

on Jeopardy!, has been successfully applied to a wide range of sophisticated real-world

problems. One specific and an important use case is its application in the medical field,

precisely in the field of oncology to better define personalized treatment for cancer patients

(Esteva et al. , 2017). This capability ensures that for any particular phenomena, appropriate

clinical trials and measures for treatments are selected hence a positive impact on patient’s

results.

Another application that is quite prominent is in the field of Finance and Banking,

where cognitive computation tools are applied in improving and facilitating the identification

of frauds and assessment of risks. It is possible for cognitive systems to review the
12

Week 3 Assignment

transaction data and patterns, and in so doing, Village can identify patterns of fraudulent

occurrences (Davenport & Ronanki). This way, the financial institutions carry out regular

preventive measures in order to reduce risks and strengthen security to protect their assets and

follow the regulations.

In customer service, cognitive computing has significantly changed the landscape in

how it is created today across organizations. Cognitive computing enabled virtual assistants

and Chatbots can interpret the explicit and latent semantic of the customer query in natural

language and supply the query with relevant and personified answers (Diederich, Brendel,

Kolbe, 2019). Such systems can support millions of queries, adapt to interaction and increase

with time thus bringing improvements to clients and productivity.

Week 3 Assignment

References
Coenen, L., Verbeke, W., & Guns, T. (2022). Machine learning methods for short-term

probability of default: A comparison of classification, regression and ranking

methods. The Journal of the Operational Research Society, 73(1), 191–206.

https://doi.org/10.1080/01605682.2020.1865847

Davenport, T. H., & Ronanki, R. (2018). "Artificial intelligence for the real world." Harvard

Business Review, 96(1), 108-116. https://doi.org/10.1145/3144177

Diederich, S., Brendel, A. B., & Kolbe, L. M. (2019). "Towards a taxonomy of platforms for

conversational agent design." International Journal of Information Management, 48,

263-282. https://doi.org/10.1016/j.ijinfomgt.2018.10.005

Dou, B., Zhu, Z., Merkurjev, E., Ke, L., Chen, L., Jiang, J., ... & Wei, G. W. (2023). Machine

learning methods for small data challenges in molecular science. Chemical

Reviews, 123(13), 8736-8780. https://doi.org/10.1021%2Facs.chemrev.3c00189

Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017).

"Dermatologist-level classification of skin cancer with deep neural networks." Nature,

542(7639), 115-118. https://doi.org/10.1038/nature21056

NeuroShell. (2023). Examples. Retrieved from http://www.neuroshell.com/examples.html

Psaros, A. F., Meng, X., Zou, Z., Guo, L., & Karniadakis, G. E. (2023). Uncertainty

quantification in scientific machine learning: Methods, metrics, and

comparisons. Journal of Computational Physics, 477, 111902.

https://doi.org/10.1016/j.jcp.2022.111902

Shen, Y., Wang, J., & Navlakha, S. (2021). A Correspondence Between Normalization

Strategies in Artificial and Biological Neural Networks. Neural Computation, 33(12),

3179–3203. https://doi.org/10.1162/neco_a_01439
14

Week 3 Assignment