Questions & explanations
1. Give an example of an unbalanced triad and why it might cause stress.
Suppose you like your friend, but your friend hates your enemy, yet you also hate that enemy. That triad is balanced? Actually, let's correct: You like friend (positive), you hate enemy (negative), friend hates enemy (negative) => two negatives, one positive => balanced. Unbalanced: you like friend, friend likes enemy (positive), you hate enemy (negative) => two positives, one negative? Wait: Heider's balanced triads: all positive, or one positive and two negatives. So unbalanced: one positive and two negatives? No: balanced: either 3 positives, or 1 positive and 2 negatives. Unbalanced: 2 positives and 1 negative, or 3 negatives. Example: You like friend (P), you like enemy (P? that's weird), friend likes enemy (P) => you like both, but enemy? Actually simpler: three positive is balanced. Two positive one negative is unbalanced. So: You like your boss (P), your boss likes your rival (P), but you dislike your rival (N). This is unbalanced because you and your boss agree on rival? No, you disagree. That's unbalanced. Such imbalance creates tension because you feel conflict between you
2. Compare federated learning with traditional machine learning in terms of data location and model update frequency.
In traditional machine learning, all data is collected in one central place, and the model is updated in one big batch every time new data arrives. In federated learning, data stays on each device, and the model is updated many times by different devices over weeks. Traditional updates can be done any time, while federated updates happen when devices are idle and charging to save battery. Federated learning requires many more communication rounds to converge. Also, traditional learning can use powerful servers, while federated learning must work on weak mobile devices. The trade-off is privacy vs. simplicity and speed.
3. Compare discriminative and generative models for recognizing a situation like 'running' from accelerometer data.
A discriminative model learns the boundary between 'running' and 'not running' directly from labeled data. It is good at telling apart situations but does not understand how the data looks. A generative model learns the probability distribution of both situations; it can also generate new data. For example, it could create fake accelerometer signals that look like running. Generative models can handle missing data better, but they need more training data and are slower. Discriminative models are simpler and often more accurate for pure recognition tasks. Both can be used, but discriminative is more common in ubicomp.
4. How might the meaning of a hand gesture differ between the dominant culture and a marginalized community?
In the dominant culture, a hand gesture might have a neutral or positive meaning, but in a marginalized community, the same gesture could be used as a secret signal of identity or solidarity. For example, a subtle hand sign might help members recognize each other in spaces where their identity is not accepted. This difference arises because marginalized groups often develop alternative communication codes to avoid detection or oppression. The dominant culture may misinterpret these gestures, sometimes leading to conflict or stereotyping. Understanding this context is important when studying gesture across cultures.
5. Compare quantum computing and classical computing for solving optimization problems in ubicomp.
Classical computing solves optimization problems by trying one solution at a time, which can be very slow for large problems. Quantum computing uses quantum bits that can be in many states at once, allowing it to explore many possibilities simultaneously. This makes quantum computers much faster for certain problems like scheduling or routing. However, quantum computers are error-prone and require special cooling. Classical computers are reliable and already available in small sizes for ubicomp. For now, classical computers handle most ubicomp tasks, but quantum could help with the hardest parts in the future.
6. What is a trade-off between accuracy and power consumption in situation identification algorithms?
A trade-off is that more accurate algorithms often need more complex models and frequent sensor readings, which drain battery. For example, using a deep neural network to detect activities requires constant CPU use, while a simple rule-based method uses less power but may misclassify. To save power, the algorithm can run only when a simple trigger (like coarse motion) is detected. Also, sensors can sample at lower rates, trading precision for energy. In ubicomp, designers often choose a balanced approach: accurate enough for the task but not consuming too much battery, so devices last days instead of hours.
7. Compare template-based vs machine learning gesture recognition.
Template-based recognition uses fixed patterns; you store one example of each gesture and compare user input to those patterns. It is simple but less flexible: if the user makes a slightly different circle, it may fail. Machine learning uses many examples to create a model that can handle variation. For example, a machine learning algorithm can learn that people draw circles with different sizes and speeds. Machine learning is more accurate for complex gestures but needs more data and computing power. On mobile, simple gestures often use templates, while complex ones like handwriting use machine learning.
8. What is the Model Human Processor (MHP) for mobile?
The Model Human Processor (MHP) is a simple way to think about how a person processes information when using a computer. It has three parts: the perceptual processor (taking in what you see or hear), the cognitive processor (thinking and deciding), and the motor processor (moving your body, like tapping a screen). On a mobile device, the MHP helps predict how long a task will take, such as reading a message and tapping a reply button. It considers delays like how fast your eyes see something or how quickly your finger moves. This model is useful for designing mobile apps that feel fast and easy to use.
9. What does responsive design mean and what role do breakpoints play?
Responsive design means a website adjusts its layout to look good on any screen size, like a phone, tablet, or desktop. Breakpoints are specific screen widths where the design changes, for example at 768 pixels the layout switches from two columns to one. Designers set these breakpoints so the content reflows and stays readable. Without breakpoints, the site might look broken or too small on a different device. Breakpoints are defined in CSS using media queries. They help create a smooth experience across devices. Using common breakpoints like 480px for phones and 768px for tablets is a good practice.
10. How does federated learning protect user privacy compared to centralized machine learning?
Federated learning protects privacy by training a shared model across many devices without moving their personal data. Each device improves the model using its own data, then sends only the updated model weights, not the raw data, to a central server. This means your photos or messages never leave your phone. In centralized learning, all data is collected on one server, which can be hacked or misused. Federated learning also lets users keep control of their data. However, the model updates can still leak some information, so extra techniques like differential privacy add noise for stronger protection.
11. What is a challenge of bio-integrated devices, such as sensors implanted in the body, for ubicomp?
A big challenge is safety and biocompatibility. The body may reject the device or cause infection. The device must be very small, use little power, and not heat up. Also, data from inside the body can be sensitive, so privacy is a concern. Another challenge is communication: radio waves do not travel well through tissue, so new methods like ultrasound are needed. Power is also tricky because changing batteries requires surgery. These devices could monitor health continuously, but they must be reliable for many years without failure. Ethical questions about data ownership and hacking also arise.
12. Describe a situation where a Kalman filter would fail in a mobile device and how a particle filter would handle it.
Situation: You spin your phone very fast. The Kalman filter assumes motion is smooth, so it reacts slowly – the orientation estimate lags behind. The screen shows the virtual world turning late. Particle filter, with many particles, can jump to the new orientation faster because some particles will match the fast spin. Also, if the phone is dropped, Kalman filter might assume it continues falling smoothly, but the actual impact is sudden. Particle filter handles the impact because particles can spread to different accelerations. So for quick, unpredictable motion, particle filter works better.