Designing How Machines Learn: Learning Rate, Iterations and Convergence Through HCI and Human-Centered AI.

The goal is to design the invisible itself— maths is equations, but we can visualise them with good design and control them with good data sampling and model tuning through hyperparameterisation.

1. Introduction: Machine Learning as an Interactive Learning System

When we train a machine learning model, we are not simply executing mathematical operations. We are configuring a system, observing its behaviour, interpreting feedback, and making decisions about how it should learn.

This creates an interesting connection between Machine Learning, Human-Computer Interaction (HCI), and Human-Centered AI Design.

In traditional machine learning, we focus on mathematical questions:

• How does the model update its parameters?
• How does the cost function change?
• How do we find an optimal solution?

From an HCI perspective, we can ask additional questions:

• How does a person understand what the model is doing?
• How can we communicate the model’s learning progress?
• How can users control the training process?
• How can we make the system’s behaviour understandable and predictable?
• How can we help users recognise when the model is learning effectively or behaving unexpectedly?

To explore these questions, let’s understand three fundamental concepts: hyperparameters, learning rate, and iterations, followed by the concept of convergence.

We will use a simple neural network, specifically a single-layer perceptron with sigmoid activation, as our example.

2. Hyperparameters: The Controls That Shape Learning

Hyperparameters are configuration settings that we define before or during the training setup. They influence how a machine learning model learns.

Think about using a camera application.

A photographer can adjust exposure, shutter speed, and ISO to influence the final photograph. These controls do not directly create the photograph, but they influence how the camera produces it.

Similarly, hyperparameters act as controls that influence how a machine learning model learns.

In our experiment, we are working with two hyperparameters:

1. Number of iterations.
2. Learning rate.

From a Human-Centered AI perspective, these parameters are more than values in a Python function. They are controls through which a person influences the behaviour of an intelligent system.

For example, consider a training interface that allows a researcher to adjust the learning rate and number of iterations.

The interface should not simply expose two numerical input fields. It should help the researcher understand what each parameter controls, what consequences a change might have, and how the model responds.

This is an important HCI principle: a good interface makes the relationship between an action and its consequence understandable.

3. Number of Iterations: How Many Learning Opportunities Does the Model Get?

The number of iterations determines how many times the model repeats its training process.

To understand this, let’s first examine what happens during one training iteration.

A simple neural network follows four major steps:

1. Forward propagation: The model receives input data and generates predictions.
2. Cost calculation: The model calculates how far its predictions are from the actual answers.
3. Backward propagation: The model calculates gradients that indicate how its parameters should change to reduce the cost.
4. Parameter update: The model updates its weights and bias using gradient descent.

One complete execution of these steps is called one iteration.

Consider this example:

parameters = nn_model(
    X_larger,
    Y_larger,
    num_iterations=100,
    learning_rate=1.2
)

Here, the model performs 100 iterations.

Understanding iterations through an interaction analogy

Imagine a person learning to play a musical instrument.

During the learning process:

• The person performs a musical passage.
• They listen to the result.
• They identify mistakes.
• They adjust their technique.
• They try again.

Each practice cycle provides an opportunity to improve.

A neural network follows a comparable feedback process:

• It makes predictions.
• It calculates its errors.
• It determines how its parameters should change.
• It updates its parameters.
• It repeats the process.

The number of iterations determines how many such learning cycles the model performs.

What happens when we change the number of iterations?

Few iterations

The model gets fewer opportunities to adjust its parameters. It may stop training before finding a suitable decision boundary.

More iterations

The model gets more opportunities to reduce the cost and improve its predictions.

Excessive iterations

Once the model has converged, additional iterations may produce very little improvement. In some situations, continued training can also contribute to overfitting, where the model performs well on training data but poorly on unseen data.

HCI perspective: Making training progress observable

Imagine designing a dashboard for machine learning researchers.

Instead of displaying only:

Iterations: 100

A more informative interface could communicate:

• Current iteration: 65 of 100.
• Current training cost.
• Change in cost over recent iterations.
• Training progress.
• Whether the cost is still decreasing significantly.

Why does this matter?

Because users need feedback to understand the consequences of their actions.

In HCI, this relates to visibility of system status. The system should continuously communicate what is happening so that users do not have to guess whether an operation is progressing, completed, or stuck.

4. Learning Rate: How Aggressively Does the Model Learn?

The learning rate determines how much the model changes its weights and bias during each parameter update.

It is commonly represented by the Greek letter alpha (α).

The gradient descent equations are:

Weight update:

W(new) = W(old) − α × dW

Bias update:

b(new) = b(old) − α × db

Here:

• W represents the model’s weights.
• b represents the bias.
• α represents the learning rate.
• dW represents the gradient of the cost with respect to the weights.
• db represents the gradient of the cost with respect to the bias.

The gradient tells us the direction of increasing cost. We subtract it to move towards lower cost.

The learning rate determines the magnitude of this movement.

Understanding learning rate through interaction

Imagine navigating a person through an unfamiliar environment using a digital navigation system.

The navigation system continuously receives feedback about the person’s current position and suggests adjustments to the route.

There are different ways to respond to this feedback:

• Make very small adjustments.
• Make moderate adjustments.
• Make extremely large adjustments.

A similar principle applies to gradient descent.

Let’s examine three situations.

Situation 1: Learning rate is too small

Imagine moving towards a destination using extremely small steps.

You are moving in the correct direction, but reaching your destination takes a long time.

In gradient descent:

• Parameter updates are very small.
• The cost may decrease very slowly.
• Training requires more iterations.
• The model may not reach a suitable solution within the available training time.

From a human-computer interaction perspective, this can create a perception of slow system responsiveness.

If a training interface does not communicate progress properly, users might assume that the model has stopped working.

This is why progress indicators and cost visualisations are important.

Situation 2: Learning rate is appropriate

Imagine taking reasonably sized steps towards your destination.

You make consistent progress without repeatedly moving past the target.

In gradient descent:

• Parameter updates are appropriately sized.
• The cost generally decreases.
• The model learns efficiently.
• Training can converge relatively quickly.

From a system design perspective, this represents a more predictable interaction between the training configuration and the model’s behaviour.

Situation 3: Learning rate is too large

Now imagine taking extremely large steps.

You might repeatedly move past your destination without reaching it.

Something similar can happen during gradient descent.

• Parameter updates become too large.
• The cost may increase instead of decreasing.
• The model may repeatedly overshoot the minimum.
• Training can become unstable or diverge.

This is an important Human-Centered AI consideration.

A system should not simply allow users to configure parameters without communicating the consequences of potentially unsuitable values.

For example, a training interface could provide warnings when the cost increases repeatedly or when the optimisation process becomes unstable.

The goal is not to remove user control. It is to make that control more informed.

5. Convergence: How Do We Know When Learning Has Stabilised?

Convergence occurs when the model’s parameters approach stable values and further training produces very little improvement in the cost function.

To understand this, imagine learning a new skill.

Initially, every practice session produces noticeable improvement. Over time, improvements become smaller.

Eventually, your performance becomes relatively stable.

Machine learning follows a similar pattern.

Initially, the neural network starts with randomly initialised weights and bias. Its predictions may therefore be inaccurate.

During training:

1. The model generates predictions.
2. It calculates the cost.
3. It calculates gradients.
4. It updates its weights and bias.
5. It repeats the process.

The cost generally decreases as the model improves, although individual updates do not always guarantee a decrease.

Consider these hypothetical cost values:

IterationCost10.6950.55100.38200.21300.12400.08500.075600.074700.074

Notice the pattern.

Between iterations 1 and 40, the cost decreases considerably.

After iteration 40, the improvement becomes much smaller.

The model is approaching convergence.

Does convergence mean that the cost becomes zero?

No.

Convergence does not necessarily mean that the cost becomes zero.

It means that the optimisation process has reached a relatively stable point where additional updates produce very little improvement.

The final cost can still be greater than zero.

Furthermore, convergence does not automatically mean that the model is accurate or useful.

A model can converge to a solution that performs poorly on unseen data.

This distinction is important in Human-Centered AI because the objective is not merely to build a model that completes its mathematical optimisation.

The objective is to build a model that performs reliably for the task and people it is intended to support.

6. The Relationship Between Learning Rate, Iterations and Convergence

These three concepts are connected.

Think of training as a journey towards a destination.

• Learning rate: Determines the size of each step.
• Iterations: Determine how many steps the model takes.
• Cost: Measures the model’s prediction error.
• Convergence: Describes when further steps produce very little improvement.

Consider the following combinations:

Learning rateIterationsPossible outcomeVery small10Training may be too slow.Very small100Better learning, but more iterations may be needed.Appropriate10The model may not have converged.Appropriate100The model may converge successfully.Very large100The model may oscillate or diverge.

There are three important terms to understand.

Undertraining: The model has not performed enough updates to learn a suitable solution.

Convergence: The cost has become relatively stable, and additional updates produce little improvement.

Divergence: The optimisation process moves away from a solution instead of approaching it.

One important clarification: these outcomes depend on the dataset, initial parameter values, model architecture, and optimisation method. There is no universally optimal learning rate or iteration count.

7. Experimenting with Hyperparameters: From Mathematics to Interface Design

Now let’s connect the concepts to the actual Python implementation.

Our original code is:

parameters_larger = nn_model(
    X_larger,
    Y_larger,
    num_iterations=100,
    learning_rate=1.2,
    print_cost=False
)

We can perform three experiments.

Experiment A: Change the number of iterations

Keep the learning rate fixed at 1.2.

Try:

num_iterations = 5
num_iterations = 20
num_iterations = 50
num_iterations = 100
num_iterations = 500

Observe:

• How does the cost change?
• Does the decision boundary change?
• After how many iterations does the decision boundary become relatively stable?

Experiment B: Change the learning rate

Keep the number of iterations fixed at 100.

Try:

learning_rate = 0.001
learning_rate = 0.01
learning_rate = 0.1
learning_rate = 1.2
learning_rate = 5.0

Observe:

• How quickly does the cost decrease?
• Does the decision boundary approach a suitable position?
• Does the cost increase or fluctuate at higher learning rates?

These are experimental values, not guaranteed optimal settings.

Experiment C: Make the learning process observable

Our original code contains:

print_cost=False

Change it to:

print_cost=True

This allows the function to print the cost during training, assuming the nn_model implementation is configured to print it at the relevant intervals.

From an HCI perspective, this experiment introduces an important concept: feedback and observability.

A person cannot make informed decisions about a system if the system does not communicate its behaviour.

The cost function acts as a feedback signal. It helps the researcher understand whether the model is improving, stagnating, or behaving unexpectedly.

A well-designed machine learning interface could make this feedback more useful through:

• A live cost-versus-iterations graph.
• A learning-rate control with a clear explanation.
• A visual representation of the decision boundary.
• A comparison between different training experiments.
• Warnings when the cost increases significantly.
• An indication that training has stabilised.

Such an interface transforms an otherwise abstract mathematical process into something a person can observe, interpret, and control.

8. Human-Centered AI: Designing for Understanding, Control and Trust

The concepts we have discussed reveal a broader principle of Human-Centered AI.

An intelligent system should not be treated as a black box that simply produces an output.

People need ways to understand its behaviour, evaluate its performance, and make informed decisions about its use.

Consider three design principles.

8.1 Transparency: Help people understand what the system is doing

The system should communicate its current state and behaviour.

For example, instead of displaying only a message saying “Training completed”, it could communicate:

• The number of iterations performed.
• The final training cost.
• Whether the cost stabilised.
• Whether training showed signs of instability.

This helps users understand what happened during training.

8.2 User control: Allow meaningful intervention

Users should be able to configure the training process and understand the consequences of their decisions.

For example, a researcher might increase the number of iterations when the cost is still decreasing.

Alternatively, they might reduce the learning rate if the optimisation process appears unstable.

The interface should support these decisions without assuming that users always know the correct configuration.

8.3 Feedback: Make the consequences of actions visible

When users change a hyperparameter, they should be able to observe how the model responds.

For example:

Changing the learning rate → changes parameter updates → influences the cost trajectory → affects the optimisation process.

This creates a feedback loop between the person and the intelligent system.

The user makes a decision, the model responds, and the interface communicates the outcome.

This is a fundamental idea in interactive system design.

9. The Bigger Picture: From Training a Model to Designing an Intelligent System

Understanding learning rate, iterations, and convergence is not only useful for implementing a neural network.

These concepts also help us understand how to design interfaces for complex AI systems.

Consider a machine learning dashboard used by researchers, engineers, or domain experts.

Such users may need to:

1. Configure model parameters.
2. Start a training process.
3. Monitor the model’s behaviour.
4. Interpret performance indicators.
5. Identify unexpected behaviour.
6. Compare different experiments.
7. Decide whether the model is suitable for deployment.

The mathematical model performs the computation, but the interface shapes how people understand and interact with that computation.

This is where HCI becomes particularly valuable.

A well-designed interface does not merely expose technical parameters. It helps users build an accurate mental model of the system.

It connects an action to its consequence, makes internal processes more observable, and supports informed human judgement.

This is especially important when AI systems are used in complex environments such as healthcare, transportation, agriculture, disaster management, and autonomous systems.

In these environments, understanding system behaviour is not simply a usability concern. It can influence operational decisions, safety, and trust.

10. Final Takeaway

Let’s bring everything together.

Imagine that you are designing an interface through which a researcher trains a neural network.

The researcher needs to understand three things:

Learning rate — How aggressively is the model updating its parameters?

Iterations — How many learning cycles does the model perform?

Convergence — Has the optimisation process stabilised, or is the model still improving?

The mathematical relationship is:

Learning rate determines the magnitude of each parameter update.

Iterations determine how many parameter updates are performed.

Convergence describes the stabilisation of the optimisation process.

From an HCI and Human-Centered AI perspective, the additional challenge is to make these processes understandable, observable, and controllable.

The objective is not just to make machines learn efficiently. It is also to help people understand how machines learn, evaluate their behaviour, and interact with them responsibly.

The central idea: Machine learning defines how the model learns; Human-Centered AI design defines how people understand, interact with, and make decisions about that learning process.

Feel free to read my article on UX Research for HCAI Design (Human Centred AI design): Classification and Attention in Knowledge Representation.

If you’d like to learn more about my work and experiences, follow me here on Medium or visit my portfolio at https://abmincodecreations.github.io/.

Designing How Machines Learn: Learning Rate, Iterations and Convergence Through HCI and… was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.

As we continue to witness the transformative impact of artificial intelligence on the world, another revolutionary technology is quietly approaching maturity: quantum computing. While AI dominates headlines and product design conversations today, the eventual rise of quantum computing will bring its own wave of disruption and innovation. For UX designers and product creators, understanding this shift is essential not only for innovation but also for creating user experiences that are ethical, accessible, and human-centered.

Understanding Quantum Computing’s Potential Impact

Quantum computing represents a fundamental change in how computers process information. Unlike classical computers, which use bits that are either 0 or 1, quantum computers use quantum bits (qubits) that can exist in multiple states simultaneously (superposition) and be entangled with one another. This allows quantum computers to perform certain computations exponentially faster than classical computers.

Famous physicist Richard Feynman once said, “Nature isn’t classical… and if you want to make a simulation of nature, you’d better make it quantum mechanical.” This quote captures the spirit of what’s possible when quantum technology becomes accessible to everyday tools.

Use Case 1: Personalized Healthcare with Quantum Insights

Meet Sara, a patient managing a chronic condition like diabetes. Today, she uses a mobile health app powered by AI to track her glucose levels, predict trends, and get medication reminders. But with quantum computing, this system could simulate her metabolic responses in real-time based on her genome, current physical state, and dietary habits.

Imagine Sara waking up to a soft voice from her AI assistant: “Your glucose is trending toward a dip by noon. Consider a protein-rich breakfast.” She feels cared for but also vulnerable. The UX must balance insight with empathy — explaining why recommendations are made, offering choices, and protecting sensitive genomic data. When data is so personal and predictive, interfaces must become more emotionally intelligent.

Use Case 2: Quantum Logistics Optimization for Small Businesses

Ali, a logistics manager for a small e-commerce company, currently uses route optimization software powered by AI. It works well but struggles with last-minute changes in weather or traffic.

With quantum computing, Ali’s dashboard becomes alive. It shows hundreds of simulations instantly — one for each route scenario, fuel efficiency level, and delivery urgency. Ali sees a shimmering overlay of possible future states, and the system recommends the top three choices.

Ali is impressed but initially overwhelmed. The interface provides layers of clarity: a simple default view, with expandable overlays for more detail. Good UX here means turning quantum noise into intuitive guidance.

Use Case 3: Drug Discovery and Molecular Design

Dr. Lin, a chemist working in pharmaceutical research, often spends weeks running simulations on how new compounds might bind to target proteins. With quantum computing, simulations that once took weeks now take seconds.

Dr. Lin enters a molecular structure and watches a real-time visualization of its quantum interactions. She sees color-coded cloud patterns representing probabilistic bonding zones. The UX needs to make these abstract visuals actionable — offering annotations, probability sliders, and integrated team sharing.

This revolution will enable scientists to discover medicines faster than ever, but only if they can comprehend the complex outputs. UX design becomes the translator between human intuition and quantum probability.

Use Case 4: Citizens in a Quantum Smart City

In 2035, Maya lives in a city where quantum computing runs infrastructure. Her daily commute, energy consumption, and public service access are optimized by quantum algorithms.

Her phone buzzes: “Traffic accident ahead. Alternate route saves 15 minutes.” Her energy bill reads: “This week’s consumption reduced by 20% thanks to predictive cooling.”

Maya feels safe but sometimes suspicious. “How does it know all this about me?” she wonders. UX must empower users with visibility into why decisions are made, offering override options and ensuring that tech doesn’t erode trust or autonomy.

Use Case 5: Financial Decisions in the Quantum Market

Raj is an investor using a quantum-enhanced financial advisor. Instead of simple graphs, his dashboard now shows a multiverse of possible market futures. Each potential scenario glimmers with a probability rating.

Raj gets a suggestion: “70% chance of portfolio A outperforming B in the next six months, based on geopolitical quantum modeling.” He is intrigued but skeptical. The system explains: “Model uses quantum simulation of economic variables, updated every 10 minutes.”

Raj appreciates the transparency. Here, the UX must make incredibly complex backends comprehensible, blending confidence indicators, narrative explanations, and ethical warnings.

Quantum Computing in Scientific Research: Chemistry and Biology

Quantum computing’s biggest near-term promise may be in accelerating discovery:

• In chemistry, quantum simulation can unlock design of new materials and drugs by accurately modeling electron interactions.
• In biology, it enables high-fidelity protein folding models and genomic analysis in seconds, advancing personalized medicine.

Famed physicist David Deutsch once said, “Quantum computers will help us understand the fabric of reality.” But UX designers will ensure people can interact with that reality meaningfully.

Security and Privacy: A New Frontier

Quantum computing doesn’t just solve problems — it creates new ones, especially around encryption. Data that is secure today could be easily decrypted by a quantum computer tomorrow. Identity management, secure communication, and digital trust must be reimagined.

Users may feel either invincible or exposed. Designers must clarify what’s protected and how, using design cues that communicate quantum-safe protocols. Transparency, reassurance, and user control become fundamental pillars.

Conclusion: Human-Friendly Innovation in the Quantum Era

The rise of quantum computing is not just a technological leap — it’s a leap in how humans interact with information, uncertainty, and possibility. UX professionals must rise to the occasion by:

• Crafting interfaces that make probabilistic outcomes understandable
• Championing ethical and inclusive design in uncertain territories
• Redefining metaphors for how humans engage with future machines

“Everything we call real is made of things that cannot be regarded as real,” said Niels Bohr, a pioneer of quantum theory. In a future where machines process unreality to improve our real lives, the job of the UX designer will be to ground, translate, and humanize the intangible.

Understanding Qubit Collapse — The Quantum Measurement Paradox

To fully grasp the implications of quantum computing for user experience, it helps to understand one of its most counterintuitive phenomena: the collapse of a quantum state upon measurement. This principle is foundational to how quantum systems behave and, by extension, how future interactions with such systems must be designed.

Consider a qubit encoded in the polarization of a photon. Polarization is the orientation of the photon’s electromagnetic field — commonly described in terms like horizontal (|H⟩) and vertical (|V⟩) polarization. Unlike classical bits that exist in a definitive state of 0 or 1, a quantum bit — or qubit — can exist in a superposition of both:

|ψ⟩ = α|H⟩ + β|V⟩, where |α|² + |β|² = 1.

This means that until we observe the photon’s polarization, it simultaneously holds a probability of being both horizontal and vertical. But here’s where the strange quantum magic kicks in: the moment we measure the photon to determine its polarization, the superposition collapses into one of the basis states — either |H⟩ or |V⟩ — with a probability governed by the square of the respective coefficients.

Measurement, in this context, is not a passive process. It fundamentally alters the state of the system. As if reality itself only “decides” what the outcome is when we ask the question. This has major implications for user interactions in quantum systems.

Imagine a user operating a quantum visual analytics interface that simulates multiple probable futures of a system. Each interaction — a click, a query, a simulation run — could cause a collapse of quantum information into a definitive outcome. This introduces a radically different interaction model: not just exploring a fixed dataset but influencing the dataset’s resolution by the act of observation.

Here’s a metaphor often used: a quantum system is like a spinning coin, simultaneously heads and tails. Measurement is like catching the coin — it must land on one side, and only then do we “know” what it was. But in quantum mechanics, it wasn’t either until that moment.

In practical use, take a photon polarized at 45°. When passed through a vertical polarizer, there’s a 50% chance it will pass (collapsing into |V⟩) and a 50% chance it will be absorbed (collapsing into |H⟩). This randomness isn’t due to ignorance — it’s built into the fabric of quantum mechanics.

For UX designers, this concept demands an entirely new paradigm of thinking. Interfaces need to be designed with uncertainty, probability, and irreversibility in mind. Systems must clearly communicate the probabilistic nature of their outputs while offering users intuitive metaphors for understanding and interacting with these strange behaviors.

As physicist Niels Bohr once said:

“Everything we call real is made of things that cannot be regarded as real.”

Understanding and translating this quantum weirdness into meaningful, human-friendly interactions is the next great challenge in UX design.

Designing for the quantum future starts now.

If you’d like to learn more about my work and experiences, follow me here on Medium or visit my portfolio at https://abmincodecreations.github.io/.

The Future of UX in the Age of Quantum Computing was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.