Designing and Building

AI Products
and Services
Workbook
Your AI Mascot will guide you through
this week, so watch out for these icons:

Read

Write

Consider

Visit the website

Design/Create

1
 Workbook 8: Final Capstone Project

Design and develop a product using AI with the aid of the

Week 8

resources you have learned about throughout the course. The

application you decide to develop in your project is completely
arbitrary; it can be something that has to do with your interest or
hobbies or something that is closely connected with your
educational and/or working background. To help you, we have
listed the main points below or parts that will help you
throughout each phase of your design project.

2
 1 Step One: AI Design Practices

As you have learned in Module 8, AI design practices involve four stages,

namely boundary setting, value formation resulting from explanation, and
bias remediation. Briefly explain these four steps.

AI systems have come under scrutiny lately because of potential bias

issues. Answer the following scenarios in terms of model bias.

Scenario One:

Assume that you implement a resume filtering tool using machine learning
models. However, your model starts rejecting applicants based on their
gender. What is the potential reason(s) why your model biases against a
particular gender, and how can you solve it?

Type your response here

If the training data used to develop the model is biased, reflecting historical gender disparities in
rejecting, the model may learn and perpetuate these biases. First of all, check if imporve data
imbalance or model drift issues can fix this problem.

Data Imbalance:
Biased features in resumes, such as specific keywords or historical patterns, can unintentionally
contribute to gender bias in the model's decision-making.

Model Drifts:
Certain model architectures might inadvertently amplify gender-related patterns in the data,
leading to biased predictions.

How to fix:
1. Ensure that the training dataset is diverse, representative, and free from gender bias.
2. Regularly review and update features to eliminate gender-specific indicators.
3. Use data augmentation techniques to balance gender-related features.

3
 1 Step One: AI Design Practices(Cont.)

Scenario Two:

Assume that you want to detect attack samples over the internet
by implementing an AI system. However, your training dataset is
imbalanced where it consists of a small number of attack instances.
In such circumstances, it is challenging to classify samples as an attack
or non-attack using traditional machine learning algorithms since
state-of-the-art machine learning models cannot learn the characteristic
behavior of the minority attack class. As a result, models are easily
biased to the majority class. One naïve solution is to remove the
necessary amount of non-attack samples and make the dataset
balanced with a small size. What is the limitation of this solution?
How can the GAN idea be integrated into this imbalanced training
dataset issue to produce a more efficient solution?

Type your response here

Removing non-attack samples to balance the dataset might lead to a loss of important
information. It can result in oversimplification and inadequate representation of the majority class,
this action may improve data imbalance but limit the model's ability to generalize effectively.

Use GANs to generate synthetic attack samples, and augment the minority class, can learn the
underlying patterns of attacks and produce realistic samples that enhance the dataset, improve
data imbalnce by GANs can can learn features from both attack and non-attack instances,
preventing loss of important information, preserve the diversity and complexity of the attack class.

By integrating GANs, we can adapt to evolving attack strategies by continuously generating new
synthetic attack samples. Use a dynamic adaptation and strategy for the model's robustness over
time.

4
 Step Two: Transfer Learning and
2 Online Learning

Assumption One:

Assume that you want to create an AI system to improve production of

your company. However, you don’t have enough training samples to
implement an AI system. You propose using a pre-trained model to
reduce the computational demands of learning. Such an approach is
often associated with the idea of transfer learning. Provide a thorough
description of transfer learning and how and why it works including a
concrete example?

Type your response here

Transfer learning is a machine learning technique where a model trained on one task is
repurposed for a related but different task. We leverage the knowledge gained from a source task
to improve the learning of a target task, so it can benefit my tasks and pre-trained preparation
before proceeding to proceed with the real-time verification of production improvement.

A typical example of transfer learning is like one introduced by this week's course. It gives a
model that is pre-trained on a large dataset containing images of various animals (cats, dogs,
birds). The learned features include edges, textures, and shapes relevant to animal classification.
A case of reusing the model is used for recognizing dogs in photos, for example, transfer learning
for solving stray dogs problems via analyzing images from street montors.

5
 Step Two: Transfer Learning and
2 Online Learning(Cont.)

Assumption Two:

Assume that you generate an AI system to create a fully automated model

for your company's needs. Your model initially works fine as it is. However,
your company produces new training data daily that is rapidly or slightly
different from your initial training samples. Therefore, your AI model is not
suitable to solve the problem for new samples in a dynamic setting.
Provide a thorough description of online learning and how and why it
solves the problem under the a forementioned circumstances.

Type your response here

Online learning, also known as incremental or streaming learning, is a machine learning paradigm
that enables a model to adapt and update its knowledge continuously as new data becomes
available. It is well-suited for dynamic environments with evolving data like the given problem
example in the above description.

Online learning has key features:

- Continuous Training -- trained incrementally, allowing them to adapt to new observations without
retraining on the entire dataset.
- Sequential Updates -- new data arrives sequentially, and the model updates its parameters
based on each incoming sample.

Online learning helps capture and adapt to dynamical changes related to the problem of Concept
Drift, by updating the model continuously, it helps maintain the relevance and accuracy in
scenarios and provides model robustness, with the usage of allocating resources focusing on
recent and relevant data it also lifts the resource usage efficiency.

6
 3 Step Three: Develop A Plan

This question consists of various sub-steps.

Part One:

Choose a problem that you want to solve with the help of AI. As an
AI-based model solution to your problem, you need to think about the
data-gathering efforts, infrastructure, storage, ETL, and cleaning you
may want to conduct.

Think about what your project may require to complete this step. You can
also search various resources like research papers and other publications
that might help you in this process.

Type your response here

Topic: Developing a hand gesture recognition program Developing a hand gesture recognition
program with AI support involves several steps in data collection, infrastructure, and storage.
Here's my structured approach.

1. Data Gathering:
Collect gestures data through user input when they perform input operations via various devices
like laptops and mobiles. Give the input data features labeled with annotation on images or
sequences with the corresponding hand gesture. Performing user demographics on diverse user
groups to account for variations in hand shapes, and sizes. Also, corresponding data sources
including laptops and mobile devices, to ensure the model's adaptability across platforms.

2. Infrastructure:
Consider cloud-based solutions for scalable and efficient infrastructure, to use AWS, Azure, or
Google Cloud.

3. Data processing:
Develop ETL pipelines for preprocessing data. Steps may include resizing images, normalizing
pixel values, and augmenting data for increased diversity. Divide the dataset into training,
validation, and testing sets.

7
 3 Step Three: Develop A Plan (Cont.)

Part Two:

This step involves incorporating the higher steps in the pyramid that we
have defined as the AI hierarchy of needs. This step involves, among
other actions, employing data analytics, aggregating the data, and
training the data. Other parts of this step involve running A/B testing
and simple ML algorithms to understand what to expect from your data.
Finally, you will need to think about which AI deep learning method to
apply to achieve your goals. Depending on your project, this step may
also include designing an HCI.

Now, write a short paragraph about the second part.

Type your response here

1. Data Analytics:
Analyze the distribution of gesture classes, identify patterns, and explore statistical measures.
Use data visualization techniques to understand relationships between features and classes.
Examining correlations between features helps in understanding how different elements in the
data relate to each other.

2. Aggregating Data:
Aggregate data based on spatial features across different time periods for each user and user
group.

3. Training the Model:

Choose CNN as an appropriate deep learning architecture for hand gesture recognition, and train
the model on the preprocessed and augmented dataset. Utilize the training, validation, and
testing sets for accurate evaluation.

4. A/B Testing and Simple ML Algorithms:

Comparing different versions of the model or variations in the recognition algorithm helps in
identifying the most suitable approach and evaluating their effectiveness. Define and measure
relevant performance metrics (accuracy, precision, recall) for both A/B testing and simple ML
algorithms.

8
 3 Step Three: Develop A Plan (Cont.)

Part Three:

Another fundamental step when designing an AI product involves thinking

about the humans in the loop, not in terms of a gathering of data, as you
have described in the previous step—but more as a way of thinking about
how humans may be involved in decision making (think about the example
presented in Week 7 about autonomous vehicles).

Describe the number of ways you think that humans and your idea will be
interacting and document them in a point-wise manner.

Type your response here

This AI model will actively help human user input efficiency in a contiguous approach.
1. Humans can provide accurate and nuanced annotations for training data. For tasks like input
gesture processing, human annotators can label complex patterns that AI might struggle with.

2. Integrate feedback from human evaluators to continuously refine and improve the AI model.
Humans can identify false positives/negatives and provide insights that algorithms might miss.

3. Humans are good at understanding context and handling ambiguous situations. In scenarios
where AI might struggle due to uncertainty, involving humans in decision-making can lead to
more contextually aware outcomes.

With these additional user information contiguous filling to the dataset as a contiguous online
learning approach. The model has a good chance to be improved efficiently to a robust model.

9
 4 Step Four: Facing Possible Challenges

The last step of designing an AI product may involve thinking about

facing the possible challenges, solving the problems, and extending
the ongoing research to improve your product.

Describe the points mentioned above based on the AI product you

want to design.

Type your response here

When the input program to reach a mature production stage with AI assistant power. The future
anticipated possible challenges would widely include legal and regulatory compliance issues, user
adaptation problems, data explainability and transparency, and also cybersecurity robustness
considerations. So some approaches should be placed before problems happen:

1. Implement robust cybersecurity measures to protect against potential attacks and regularly
update security protocols.

2. Regularly review and update the AI product to ensure compliance with evolving legal and
regulatory frameworks. Stay informed about data protection laws, user privacy rights, and
industry-specific regulations.

3. Use user feedback to adapt the AI system iteratively, ensuring it stays aligned with user
expectations.

4. Develop educational materials to help users understand the capabilities and limitations of the
AI product.

5. Allocate resources for ongoing research and development to stay at the forefront of AI
advancements. Explore emerging technologies, algorithms, and methodologies to enhance the
product's capabilities.

Congratulations on designing your first AI product!

10
 Designing Artificial
Intelligence Products
Workbook