Unit 2: Artificial Intelligence & Machine Learning - Practice Quiz

Artificial Intelligence is a broad field of computer science concerned with building smart machines capable of performing tasks that typically require human intelligence.

The core concept of Machine Learning is to allow algorithms to use data to learn patterns and make predictions or decisions, rather than having every rule hard-coded.

Deep Learning is a specific field within Machine Learning that uses deep neural networks (neural networks with multiple hidden layers) to analyze complex patterns in large datasets.

Expert systems are designed to solve complex problems by reasoning through bodies of knowledge, represented mainly as if–then rules, just as a human expert would.

Fuzzy Logic is a form of many-valued logic that deals with reasoning that is approximate rather than fixed and exact. It allows for values between 0 (false) and 1 (true).

AR enhances a user's perception of the real world by adding digital elements to it, often through a smartphone camera or special glasses. This differs from Virtual Reality (VR), which creates a fully immersive virtual world.

Natural Language Processing (NLP) is a branch of AI that enables computers to understand, interpret, and generate human language.

AI, particularly computer vision, is highly effective at identifying patterns in medical scans, assisting doctors in making faster and more accurate diagnoses.

AI can analyze images and data from drones and sensors to monitor crop health, detect diseases, and optimize resource use, a practice known as precision agriculture.

AI tools can perform sentiment analysis on a massive scale, helping businesses understand public opinion and feedback about their products and services from social media.

Python is the dominant language in AI/ML due to its simple syntax and a vast ecosystem of powerful libraries like TensorFlow, PyTorch, and Scikit-learn.

These are essential software libraries that provide pre-built modules and functions, making it easier for developers to build and train machine learning models.

Google Translate relies heavily on NLP and a sub-field called Machine Translation to understand the grammar, context, and meaning of text to translate it accurately.

Computer Vision is the field of AI that enables machines to interpret and understand information from images and videos, which is essential for a self-driving car to navigate its environment.

These assistants use Speech Recognition and Natural Language Processing to interpret spoken requests and perform actions like playing music, setting alarms, or answering questions.

ChatGPT is built on a Generative Pre-trained Transformer (GPT) architecture, which is a type of Large Language Model designed to understand and generate human-like text.

ChatGPT was created by OpenAI, an AI research and deployment company.

Generative AI, which includes models that can create text, images, music, and code (like ChatGPT and DALL-E), is one of the fastest-growing and most impactful trends in the field today.

A Machine Learning Engineer is a professional who designs, builds, and deploys ML models to solve business problems. The other roles belong to different areas of information technology.

Machine Learning is fundamentally based on mathematical concepts like linear algebra, calculus, and probability, along with statistical methods for data analysis and model evaluation.

This is an example of Supervised Learning because the model learns from a dataset where each data point (image) is explicitly labeled with the correct output ('cat' or 'dog'). The goal is to learn a mapping function to predict the output for new, unlabeled data.

The Inference Engine is the 'brain' of an expert system. It uses reasoning methods (like forward or backward chaining) to apply rules from the Knowledge Base to the user's query (patient symptoms) to arrive at a conclusion or diagnosis.

Fuzzy logic excels in situations that involve ambiguity and continuous values, mirroring human reasoning. It allows for smooth transitions between states (e.g., fan speed can gradually increase as it gets 'warmer') rather than abrupt on/off switches, making it ideal for control systems like this.

The key distinction is the relationship with the real world. Augmented Reality (AR) overlays computer-generated images and information onto the user's view of the real world. Virtual Reality (VR) replaces the user's real-world environment with a completely simulated one, typically requiring a headset to block out external stimuli.

Text Classification (or categorization) is the process of assigning predefined labels or categories to a piece of text. In this scenario, the goal is to classify each email into one of the given categories, which is a classic application of this NLP task.

Convolutional Neural Networks (CNNs) are specifically designed for processing grid-like data, such as images. Their architecture, with convolutional layers, is highly effective at automatically learning and detecting spatial hierarchies of features (e.g., edges, shapes, textures) in images, making them the standard for medical image analysis tasks.

While both PyTorch and TensorFlow are excellent for deep learning, PyTorch is often favored in research for its 'Pythonic' nature, flexibility, and dynamic computation graphs, which are very useful for developing and debugging complex models like transformers. Scikit-learn is better for classical ML, and Pandas is for data manipulation.

Sensor Fusion is the process of integrating data from different sensors to produce more accurate, complete, and reliable information than could be obtained from any single sensor. This is crucial for autonomous vehicles to perceive their surroundings safely and accurately under various conditions (e.g., rain, fog, darkness).

The Transformer architecture, and its core 'self-attention mechanism', is the breakthrough that enables models like ChatGPT to understand long-range dependencies and context in text. This allows it to remember earlier parts of the conversation and generate coherent, relevant responses, unlike simpler models that process requests in isolation.

While a Data Scientist often focuses on research, analysis, and model creation, a Machine Learning Engineer specializes in the practical aspects of deploying, scaling, monitoring, and maintaining ML models in production environments. Their skillset bridges software engineering and data science.

Explainable AI (XAI) is a field focused on developing methods and models that can explain their decisions in a way that is understandable to humans. For 'black box' models like complex neural networks, XAI techniques are crucial for transparency, accountability, and debugging in high-stakes domains like finance and healthcare.

The significant leap in performance came from adopting Neural Machine Translation (NMT), specifically using deep neural networks (like LSTMs and later, Transformers). Unlike SMT which translated phrases in isolation, NMT could process entire sentences as input, allowing it to capture context and produce much more fluent and accurate translations.

This is a regression problem because the goal is to predict a continuous numerical value (crop yield, e.g., in tons per hectare). Classification predicts a discrete category (e.g., 'healthy' or 'diseased'), while clustering groups similar data points without predefined labels.

Sentiment Analysis is a specific application of Natural Language Processing (NLP) that focuses on identifying and extracting subjective information from text. Its primary goal is to determine the emotional tone or attitude of the writer, making it perfect for gauging public opinion on social media.

Generative AI models, like ChatGPT, learn the underlying patterns and structure of their training data. They then use this learned knowledge to generate new, synthetic content that is similar to, but not a direct copy of, the data they were trained on. This is different from discriminative models which simply classify or predict labels for given inputs.

The relationship is hierarchical. AI is the broad concept of creating intelligent machines. Machine Learning (ML) is a subfield of AI that focuses on giving computers the ability to learn from data without being explicitly programmed. Deep Learning (DL) is a further subfield of ML that uses multi-layered neural networks to learn from vast amounts of data.

Overfitting occurs when a model learns the training data too well, including its noise and random fluctuations, to the point where it fails to generalize to new, unseen data. The classic symptom is excellent performance on the data it was trained on, but poor performance on a separate validation or test set.

Domain knowledge (understanding the industry/field where AI is being applied, e.g., finance, healthcare) is crucial for framing the right problem, selecting relevant features, and interpreting results correctly. Strong communication skills are essential to explain complex technical concepts and findings to non-technical stakeholders.

Edge AI involves running AI algorithms directly on a local device. This significantly reduces latency (delay) because data doesn't need to travel to a cloud server and back. It also enhances privacy and security, as sensitive data (like a video feed from a security camera) can be processed on the device itself without being transmitted over the internet.

Deep neural networks are characterized by their depth (many layers), which results in millions or even billions of parameters. To properly adjust these parameters and avoid overfitting, the model needs to be exposed to a massive and diverse dataset. With small datasets, these complex models can easily memorize the data instead of learning generalizable patterns.

While overfitting (A) is related, the core issue described is a 'domain shift' or 'distributional shift,' a specific and difficult type of generalization failure. The model learned the statistical distribution of the training domain (e.g., studio photos) but failed to generalize to the target domain (e.g., outdoor, real-world photos). The vanishing gradient problem (B) is a training issue, not a deployment generalization issue. Insufficient resources (D) would affect speed, not accuracy due to environmental changes.

The fundamental limitation of rule-based expert systems is the 'closed-world assumption.' The system assumes its knowledge base is complete. It cannot infer or hypothesize about entities or rules that are not explicitly defined. Therefore, it cannot identify a new disease, even if the symptoms are present, because the disease itself is outside its defined world. While the knowledge acquisition bottleneck (B) is a real problem, the immediate cause of the diagnostic failure is the closed-world assumption (D).

In fuzzy logic, the 'AND' operator (conjunction) is typically implemented using a T-norm, the most common of which is the min operator. It takes the minimum of the membership grades of the antecedents (Temperature IS Hot and Humidity IS Humid). This is a conservative choice because the rule's overall truth value is limited by its weakest premise. Fuzzy OR uses max. Probabilistic SUM is a different type of aggregation. Centroid Defuzzification is a method for converting the final fuzzy output set back to a crisp number, not for evaluating a rule's antecedent.

The key advantage of the Transformer architecture is its self-attention mechanism. Unlike RNNs (B), which process text sequentially and struggle to remember context over long distances (even with gates, A), self-attention allows a word to directly 'look at' and weigh the importance of all other words in the sequence simultaneously, regardless of their distance. This parallel processing and direct contextual linking are what solve the long-range dependency problem so effectively. While Transformers are deep (D), the mechanism itself is the core innovation.

This is a key distinction of Level 4 autonomy. While the system is fully autonomous within its Operational Design Domain (ODD), it is not expected to handle all possible edge cases, especially those requiring nuanced human interaction like interpreting hand signals (A is a goal of Level 5). When a Level 4 system encounters a situation it cannot confidently navigate, it is designed to achieve a 'minimal risk condition' (D), which is typically to stop safely. It does not require the human to take over immediately (that's Level 3), but it recognizes its own limitations and defaults to a safe state.

LLMs are fundamentally sequence prediction models. They learn the statistical patterns of language. A hallucination occurs because the model generates a sequence of tokens (words) that is highly probable given the context, but it lacks a genuine understanding or a mechanism to verify the factual accuracy of the generated statement. It is not 'looking up' facts from a database (A) but rather predicting the next most likely word. The output is a plausible-sounding continuation, not a verified fact.

This question differentiates the core roles. The Data Scientist's work often ends with a validated model in a research/prototyping environment. The ML Engineer's primary role is 'MLOps'—operationalizing the model. This involves taking the proof-of-concept code and refactoring it for performance, reliability, and scalability; containerizing it for consistent deployment; and exposing it via an API so other applications can use it. A, C, and D are typically responsibilities of a Data Scientist or AI Researcher.

SLAM is the fundamental technology for markerless AR. It allows a device to build a map of an unknown environment while simultaneously keeping track of its own location within that map. This is what enables the AR system to understand the geometry of the room and 'anchor' virtual objects to real-world surfaces. Object detection (A) would identify objects like 'chair' or 'table' but doesn't handle the spatial mapping and device tracking required to anchor new virtual objects.

In high-stakes fields like healthcare, a correct prediction is not enough. Medical professionals need to understand the reasoning behind a diagnosis to trust it, verify it, and take legal and ethical responsibility for the treatment plan. Deep learning models are often 'black boxes,' making it very difficult to trace the specific features (e.g., which pixels or textures) that led to its conclusion. This lack of explainability, a field known as XAI (Explainable AI), is a major hurdle for adoption, even with high accuracy.

The core limitation of SMT was its reliance on phrase-based translation. It would break sentences into chunks, translate them independently, and then stitch them together, often resulting in unnatural and grammatically flawed output. NMT, particularly with architectures like RNNs and Transformers, processes the entire source sentence to generate a translation, allowing it to capture context, fluency, and more complex grammatical structures, leading to significantly more human-like translations.

The defining characteristic of Federated Learning is that it decentralizes the training process to protect user privacy. Instead of collecting raw data onto a central server, the central server sends the model to local devices (like smartphones). Each device trains the model on its local data, and only the resulting model parameters (updates or gradients) are sent back to the server for aggregation into an improved global model. The raw, sensitive data never leaves the user's device.

This scenario specifically calls for on-device (edge) deployment. TensorFlow Lite (and similarly, PyTorch Mobile) is a specialized framework designed to optimize, convert, and run deep learning models on resource-constrained devices like mobile phones. It reduces the model size and optimizes for mobile hardware. The other options are ill-suited: Scikit-learn is for classical ML, not deep learning for object detection (A); Spark is for big data processing, not edge deployment (C); a standard Python server is for cloud deployment, not on-device (D).

The Vanishing Gradient Problem is a major challenge in training deep networks, especially RNNs. During backpropagation, the gradients are repeatedly multiplied by the weights of each layer. If these weights are small (less than 1), the gradient signal can shrink exponentially as it propagates back to the initial layers, becoming so small that it no longer contributes to updating the weights, thus stopping the learning process for those layers. LSTMs and GRUs with their gating mechanisms were specifically designed to mitigate this issue.

The fuzzy logic process has distinct steps. 1) Fuzzification: Convert crisp inputs to fuzzy sets. 2) Rule Evaluation: Apply fuzzy rules. 3) Aggregation: Combine the outputs of all rules into a single fuzzy set. 4) Defuzzification: Convert that final aggregated fuzzy set back into a single, crisp output number. Common defuzzification methods include Centroid, Bisector, and Mean of Maximum.

Traditional voice assistants are primarily designed for NLU (Natural Language Understanding). Their goal is to understand a user's command by classifying their 'intent' (e.g., 'play music') and extracting key 'entities' (e.g., 'artist: Daft Punk'). Their responses are often templated. ChatGPT, on the other hand, is a generative model. While it also performs NLU, its defining capability is open-ended Natural Language Generation (NLG), creating novel, contextually relevant sentences from scratch rather than just filling in a template.

Identifying sarcasm is an extremely difficult NLP task. It requires the AI to go beyond the literal meaning of words ('Great', 'SoImpressed' are positive) and understand the context, common sense (3-hour battery life is bad), and cultural cues to recognize the intended negative sentiment. This is a higher order of complexity than simple sentiment classification (B), which often fails on sarcastic text, or basic data extraction (A, D).

This is a classic 'lab-to-field' problem in AI for agriculture. The controlled environment of the lab creates a very narrow data distribution. The real world ('the field') has immense variability. This difference between the training data distribution and the deployment data distribution is called a 'domain shift'. The model overfit the pristine lab conditions and did not learn the robust features needed to handle real-world variations, causing its performance to collapse. This is a much more fundamental issue than just camera resolution (A).

The adversarial nature of GANs makes them exceptionally good at creating synthetic data that is nearly indistinguishable from real data. While this has positive applications like augmenting medical datasets (D), its most prominent and discussed risk is the creation of 'deepfakes.' These are malicious manipulations of media used to create convincing but false evidence, spread propaganda, and engage in targeted harassment, posing a serious threat to information integrity.

Cameras rely on visible light and are severely degraded by fog, making them the least reliable for any task. LiDAR uses light (laser pulses) and is also significantly affected by fog, as the light scatters off water droplets. Radar, however, uses radio waves that can penetrate fog, rain, and snow with minimal interference. While it provides a lower-resolution output than LiDAR, it is extremely robust for detecting the presence, distance, and velocity of objects (especially metallic ones like other cars), making it the most reliable sensor in these adverse weather conditions.

The core function of an AI Researcher is to push the boundaries of the field by creating new knowledge and techniques. This requires a profound theoretical foundation in mathematics to understand why existing methods work and to formulate new ones. While ML Engineers (A) focus on productionizing models and Data Scientists (B) focus on applying models to solve business problems, the AI Researcher's main currency is the invention and validation of novel algorithms and architectures.