Unit 1 - Practice Quiz

Artificial Intelligence is a field of computer science focused on creating systems that can perform tasks like learning, reasoning, problem-solving, and understanding language, all of which are characteristic of human intelligence.

Narrow AI, also called Weak AI, is focused on a single, specific task. Most of the AI we use today, like virtual assistants and image recognition software, falls into this category.

AI, particularly computer vision, is widely used in healthcare to analyze medical scans and images, assisting doctors in diagnosing conditions like cancer or fractures more accurately and quickly.

TensorFlow (developed by Google) and PyTorch (developed by Meta/Facebook) are two of the most popular software libraries that provide tools and frameworks for building and training machine learning and deep learning models.

Transparency, or Explainability, is a key principle of Responsible AI. It focuses on making the 'black box' of AI decision-making more understandable, so that humans can trust, audit, and debug the system's outputs.

A state space consists of all possible configurations of a problem. A single 'state' is a snapshot of one of those configurations, such as the positions of all pieces on a chessboard at a particular turn.

AGI refers to a form of AI with human-like cognitive abilities, capable of solving unfamiliar problems and learning new tasks without being explicitly programmed for them. It is still a theoretical concept.

AI is a branch of computer science. Its foundations are also deeply rooted in mathematics, psychology, linguistics, and neuroscience, but computer science provides the primary framework for its implementation.

Natural Language Processing (NLP) is the area of AI focused on enabling computers to understand, interpret, and generate human language, both text and speech.

The foundation of any successful AI model is high-quality data. Therefore, the first step is always to collect, clean, and prepare the data that will be used for training and testing.

One of the biggest hurdles in AI is data acquisition. Many state-of-the-art models are 'data-hungry' and need vast datasets to learn effectively, and creating this data can be expensive and time-consuming.

Clustering is an unsupervised learning technique where the algorithm groups data points based on their similarities, without being told what the groups are beforehand. An example is grouping customers based on purchasing behavior.

Learning from experience is a fundamental aspect of intelligence. It allows an entity to adapt its behavior and improve its performance on future tasks based on past outcomes.

A data-centric approach prioritizes improving the data used to train a model over tweaking the model's code or architecture. The philosophy is that high-quality data is often more impactful for performance than a slightly better algorithm.

The problem space, or state space, encompasses all possible states of the problem. For a route-finding problem, this includes all potential paths and intermediate locations between the start and goal.

Bias in AI refers to skewed results that are prejudiced against certain groups or outcomes. This often originates from biased training data that doesn't accurately represent the real world.

Learning is a hallmark of AI. Machine learning algorithms, for example, can analyze data and 'learn' patterns from it, allowing them to improve their accuracy and decision-making over time as they are exposed to more data.

AI is excellent for automating repetitive tasks. Using Natural Language Processing (NLP), an AI can categorize incoming emails and send automated responses to common queries, freeing up human agents for more complex issues.

AI toolkits and libraries abstract away much of the low-level complexity, providing developers with high-level building blocks for tasks like creating neural network layers, calculating gradients, and training models.

Fairness is a critical pillar of Responsible AI. It aims to mitigate and correct for biases in data and algorithms to ensure that the AI system's decisions are equitable and just for all user groups.

In a state space representation, a 'state' is a complete snapshot of the world relevant to the problem. For routing, the agent's current location (an intersection) is a state. An 'action' transitions the agent from one state to another; in this case, driving down a road to the next intersection.

This is a classic example of algorithmic bias. The AI system learns and amplifies existing biases present in the training data, leading to unfair outcomes for certain groups. Responsible AI practices focus on identifying and mitigating such fairness issues.

Artificial Narrow Intelligence (ANI), also known as Weak AI, refers to AI systems that are designed and trained for a particular task. They operate within a limited, pre-defined range. AGI, in contrast, would possess human-like intelligence across a wide range of tasks.

Data-centric AI is an approach to AI development that focuses on systematically improving the quality and quantity of the dataset to enhance model performance, rather than iterating only on the model code or architecture (which is model-centric AI).

The Turing Test does not evaluate the internal thought processes (thinking humanly) or optimal decision-making (acting rationally). It focuses purely on external behavior – can the machine act in a way that a human interrogator cannot distinguish it from another human? This falls under the 'Acting Humanly' paradigm of AI.

PyTorch is famous for its 'define-by-run' approach, which creates a dynamic computational graph. This means the graph is built on the fly as operations are performed, making it more intuitive and Pythonic, which is especially useful for debugging and building complex, dynamic models. In contrast, early TensorFlow used a 'define-and-run' static graph approach.

Combinatorial explosion refers to the extremely rapid growth of the complexity of a problem due to how the components can be combined. In problems like chess or the Rubik's Cube, the number of possible states becomes astronomically large, making exhaustive search methods impractical.

Classification is a supervised learning task where the goal is to predict a discrete label. Since there are more than two possible categories ('Primary,' 'Social,' etc.), it is a multi-class classification problem. Clustering is unsupervised, and regression predicts a continuous value.

Computer Vision is the field of AI that trains computers to interpret and understand the visual world. Analyzing medical images like MRIs to identify patterns, objects (like tumors), and anomalies is a core task in computer vision.

In a partially observable environment, the agent cannot access the complete state of the world at all times. In Poker, the opponents' hidden cards are a critical part of the state that the agent cannot see, forcing it to reason under uncertainty.

A good heuristic for A* search must be admissible (never overestimates the cost to reach the goal). The Manhattan distance is a classic admissible and effective heuristic for grid-based problems like the 8-puzzle, as it provides an informed estimate of the minimum number of moves required to solve the puzzle from the current state.

Explainability (XAI) is the ability to explain the reasoning behind an AI's decision in human-understandable terms. In high-stakes domains like healthcare, this is crucial for trust, accountability, and debugging. A 'black box' model lacks this characteristic.

Bayesian networks are probabilistic graphical models. Their entire framework for representing knowledge and reasoning under uncertainty is built directly on the principles of probability theory, specifically Bayes' theorem.

This scenario describes a system with broad, human-like cognitive abilities, including creativity, emotional intelligence, and abstract reasoning. These are hallmark characteristics of AGI, which aims to perform any intellectual task that a human being can.

The test set must be kept separate and used only once after all training and hyperparameter tuning (on the training and validation sets) is complete. This ensures an honest, unbiased measure of how the model will generalize to new, real-world data.

The core feature highlighted in this example is the system's ability to modify its behavior based on new data (your changing routines). This capacity to learn from experience and adapt its actions over time is a fundamental characteristic of modern AI.

This task requires the AI to understand and categorize human language in text format. This is the central purpose of Natural Language Processing (NLP), and the specific task of assigning predefined categories to text is known as text classification.

The environment is dynamic because it changes while the agent is deliberating. It is continuous in space and time. It is partially observable because sensors are imperfect and cannot see everything (e.g., a car hidden behind a truck). It is also multi-agent and stochastic.

The rational agent approach defines AI as the study and construction of agents that do the 'right thing.' The 'right thing' is the action that maximizes a performance measure to achieve the best expected outcome, which doesn't necessarily have to be human-like.

The goal is to predict a continuous numerical value (the stock price). This is a regression task. More specifically, since the data is a sequence ordered by time, it is a time-series forecasting problem, which is a specialized form of regression.

The Chinese Room argument is a thought experiment designed to challenge the central thesis of 'Strong AI'. Searle concedes that a machine could pass the Turing Test (a behavioral test) and that AI can be a useful tool (Weak AI). His core objection is that symbol manipulation (syntax), no matter how complex, is insufficient to produce genuine understanding or intentionality (semantics). The man in the room follows rules to manipulate Chinese symbols perfectly but understands nothing of Chinese, analogously arguing that a computer running a program does the same.

In the classic 8-puzzle, every move is reversible, making the state space graph undirected. The 'zap' action, however, is irreversible; you cannot 'un-zap' a tile. This makes the graph directed. The branching factor, which is constant in the classic puzzle (depending on the blank's position), now becomes variable. If the blank is in the center, a tile can be moved from 4 positions OR one of the 4 adjacent tiles can be zapped, changing the branching factor. The state space remains finite because there are a finite number of tiles that can be zapped. Heuristics like A* can still be used, but the heuristic function would need to be carefully designed to remain admissible.

This scenario is a classic example of the accuracy-fairness trade-off. While the model has high overall accuracy, it is not fair. Specifically, it violates the principle of 'Equality of Opportunity,' which suggests that individuals from different groups with similar qualifications should have similar outcomes. A much higher False Rejection Rate for one group indicates that qualified individuals from that group are being unfairly denied loans at a disproportionate rate. Optimizing for overall accuracy alone has masked this significant fairness issue.

The core difference is the execution model. TensorFlow 1.x used a static computation graph ('define-then-run'). You would first define the entire graph of operations, compile it, and then run data through it. This model is efficient for static architectures but very difficult for models where the control flow (e.g., loops, conditionals) depends on the runtime values of tensors. PyTorch and TensorFlow 2.x use eager execution ('define-by-run'), where the graph is built dynamically as operations are executed. This makes implementing models with dynamic structures, like the one described, as natural as writing standard Python code with if statements and for loops.

Data-centric AI prioritizes improving data quality over model architecture changes. The problem is a clear case of data drift or a gap in the training distribution (lack of foggy images). While collecting more data is a data-centric approach, it's not targeted or efficient given the resource constraints. Switching models or tuning hyperparameters are model-centric approaches. The most direct, efficient, and data-centric solution is to synthetically create the data that is missing. Targeted data augmentation (adding synthetic fog) directly addresses the model's specific weakness without requiring new data collection or expensive model re-architecting.

Local search algorithms, like Hill Climbing, operate by iteratively moving to a neighboring state that improves the current state's value. Their fundamental weakness is their 'greedy' nature. They will get stuck on local maxima because no neighboring state offers an improvement, even if the global maximum is elsewhere. Similarly, on a plateau (a flat area of the search space), the algorithm has no gradient to follow and may wander aimlessly or terminate prematurely. While other options present challenges, the presence of local maxima and plateaus is the most direct cause of failure for this specific class of algorithms.

AGI's defining characteristic is not just being good at many tasks, but the ability to generalize, reason, and transfer knowledge across disparate domains, much like humans do. Current ANI systems, even those using transfer learning, are mostly effective in closely related tasks (e.g., transferring knowledge from classifying cats to classifying dogs). The leap to AGI requires a form of abstraction and common-sense reasoning that allows knowledge to be applied to fundamentally different contexts. This challenge of cross-domain generalization is far more profound than simply scaling hardware (Option A), mastering more narrow tasks (Option B), or perfecting a single complex skill like NLP (Option D).

This scenario (reminiscent of a real Tesla accident) highlights two key issues. First, the model lacks 'common sense' — a human driver understands that large, solid objects can exist in the sky's visual space and must be avoided. The AI, trained on typical road scenes, lacked this implicit knowledge. Second, the specific visual configuration of a white truck against a bright sky was likely an 'out-of-distribution' or edge case not well-represented in the training data, demonstrating the 'brittleness' of the model. The other options are less applicable: The Frame Problem is about what doesn't change after an action, Symbol Grounding is about connecting symbols to the real world (related, but less specific), and the Qualification Problem is about listing all preconditions for an action.

This is a classic dilemma in Responsible AI. 'Interpretability' (or Explainability) is the principle that stakeholders should be able to understand why the AI made a particular decision. The simple model provides this. 'Beneficence' is the ethical principle of acting for the benefit of others; in this context, it translates to maximizing the model's utility and accuracy to provide the best possible patient care. The more accurate black box model better serves the principle of beneficence by potentially preventing more readmissions. The hospital must therefore trade off the ability to explain every decision against the potential for better patient outcomes.

Functionalism decouples mental properties (like 'belief' or 'pain') from the specific physical material that implements them. The theory argues that what matters is the function or the 'causal role' the state plays in the system. This is a cornerstone belief for many in AI, as it provides a philosophical basis for the idea that intelligence is not unique to biological brains. If the complex web of causal relationships that constitutes human intelligence can be replicated in a different substrate, like a computer, then according to functionalism, that system would genuinely be intelligent and have mental states.

The Markov Property, or memorylessness, is the core assumption of an MDP. It states that the current state contains all the information necessary to decide the future. The history of how the agent arrived at is irrelevant for predicting the next state and reward . This is a powerful simplifying assumption, as it means the agent doesn't need to consider its entire history when choosing an action; it only needs to consider its present state. This makes learning value functions and policies computationally tractable.

The Traveling Salesperson Problem (TSP) and its variants are classic NP-hard problems. The state in a search formulation could be defined as '(current city, {set of visited cities})'. The number of possible paths (permutations) to visit n cities is related to . This factorial growth leads to a combinatorial explosion, where the size of the state space becomes astronomically large even for a moderate number of cities (e.g., 20! is a massive number). Uninformed search algorithms like BFS, which explore the state space level by level, are completely impractical because they would need to explore an unmanageable number of states.

This is a sophisticated failure mode beyond simple overfitting. The model is not just memorizing noise; it's learning a statistically powerful but medically irrelevant feature (the scanner artifact) that is correlated with the label (tumor presence) in the biased dataset. This is known as 'shortcut learning.' The model found an easy way to get the right answer for the wrong reason. When deployed in a setting with different X-ray machines, it would fail completely, as the spurious correlation no longer holds. This highlights the danger of dataset bias and the importance of model interpretability (e.g., using saliency maps) to ensure the model is 'looking' at the right things.

The 'Model-Centric' approach, common in academic research, holds the data fixed and iterates on the model architecture. The 'Data-Centric' approach holds the model architecture fixed and focuses on systematically improving the data (e.g., fixing labels, adding specific examples, augmenting data). The key insight, championed by figures like Andrew Ng, is that for many practical applications, once a reasonable model architecture is in place, the ceiling on performance is determined not by the model but by the quality of the data. Systematically engineering the data is often a more direct and effective path to better performance than endlessly tweaking a complex model.

The Frame Problem highlights the issue of efficiently representing the consequences of actions. When an agent acts, most things in the world do not change. For example, if a robot picks up a cup, the color of the walls, the position of the sun, and the president's name all remain the same. A naive representation would require an enormous number of 'frame axioms' to state all these non-changes. The challenge is to find a representational framework where non-effects are implicitly handled, and only the direct effects of an action need to be specified. This is less of a problem for modern statistical AI but was a major conceptual hurdle for logic-based AI systems.

Automation involves performing a predefined task without human intervention. The industrial robot, backup script, and simple chatbot are all examples of automation; they follow a fixed set of rules or a script. Autonomy, however, implies the ability to make decisions and adapt behavior in a dynamic, unforeseen environment to achieve a goal. The Mars rover, faced with an unexpected obstacle (the crater) and without human guidance, must perceive its environment, reason about its options, and decide on a new course of action to continue its mission. This adaptive, goal-oriented decision-making in a novel situation is the hallmark of autonomy.

This is a classic case of Historical Bias. The bias exists in the data itself because it reflects past societal or organizational prejudices, even if those prejudices are no longer explicit. The AI is simply codifying a historical pattern where the tech industry has been male-dominated. The model correctly learns the statistical correlation present in the data ('Jared' -> good hire), but in doing so, it perpetuates and amplifies a harmful historical bias. Selection bias relates to how data is sampled, interaction bias comes from user feedback loops, and latent bias relates to inherent stereotypes in language or representations, but historical bias is the most direct description of a model learning from a world that was, and is, biased.

This question addresses the core concepts of scaling up AI training. In Data Parallelism (the most common method), you have multiple copies of the same model on different GPUs/machines. You split your large dataset into mini-batches and send one mini-batch to each device. Each model replica computes its gradients independently, and then the gradients are aggregated (e.g., averaged) to update the master model. In Model Parallelism, the model itself is too large to fit into the memory of a single device. Therefore, you split the model's layers or components across different devices. Data flows through these devices sequentially to complete a forward and backward pass. Data parallelism speeds up training time, while model parallelism makes it possible to train models that would otherwise be too large.

While the other options were contributing factors (especially the critique of perceptrons), the most significant cause of the first AI winter was the disillusionment that followed initial hype. Early AI, like Shakey the robot or SHRDLU, performed impressively in highly constrained, simulated 'microworlds'. However, researchers discovered that these techniques did not scale to the complexity, ambiguity, and vastness of the real world. The combinatorial explosion of possibilities and the immense amount of implicit, common-sense knowledge required proved to be insurmountable hurdles for the logic-based and search-based methods of the era, leading funding agencies to pull back.

The Turing Test is a behavioral test. It doesn't measure internal states like understanding, consciousness, or true intelligence. Its sole criterion is whether a machine can behave (specifically, converse) in a way that is indistinguishable from a human. Critics argue that this is a test of successful simulation or deception, not genuine intelligence. A system could potentially pass by using a vast database of conversational tricks and patterns (like early chatbots such as ELIZA attempted) without any real understanding of the conversation's meaning. This highlights the gap between behaviorally mimicking intelligence and actually possessing it.