Artificial Intelligence Perception, Communication, and Ethics Syllabus

What is histogram equalization and what problem does it address?

An image-enhancement technique that redistributes pixel intensities to produce a more uniform histogram, improving global contrast in images that are too dark, too bright, or low-contrast.

What quantity does the Sobel operator compute, and how is gradient magnitude obtained?

The Sobel operator estimates the image intensity gradient using two $3\times3$ kernels for horizontal ($G_x$) and vertical ($G_y$) derivatives; the edge magnitude is $G = \sqrt{G_x^{2} + G_y^{2}}$ and direction is $\theta = \arctan\!\left(\frac{G_y}{G_x}\right)$.

List the main stages of the Canny edge detector.

1) Gaussian smoothing, 2) gradient computation (magnitude and direction), 3) non-maximum suppression to thin edges, 4) double thresholding, and 5) edge tracking by hysteresis to link edges.

What does the Harris corner detector respond to, and what image property defines a corner?

It detects corners as points where intensity changes significantly in all directions. Using the structure tensor's eigenvalues $\lambda_1, \lambda_2$, a corner has both large; the response is $R = \det(M) - k(\text{trace}\,M)^{2}$.

What key invariance property does the SIFT feature descriptor provide?

SIFT (Scale-Invariant Feature Transform) produces keypoint descriptors that are invariant to scale and rotation and robust to changes in illumination and viewpoint, enabling reliable feature matching.

Distinguish image classification, object detection, and instance segmentation.

Classification assigns one label to the whole image; object detection outputs bounding boxes plus labels for multiple objects; instance segmentation outputs a pixel-level mask for each individual object instance.

What is Intersection over Union (IoU) and how is it used in object detection?

$\text{IoU} = \frac{\text{area of overlap}}{\text{area of union}}$ between a predicted and a ground-truth bounding box. A detection is usually counted as correct if IoU exceeds a threshold (commonly $0.5$).

What problem does Non-Maximum Suppression (NMS) solve in object detectors?

It removes redundant overlapping bounding boxes for the same object by keeping the highest-confidence box and suppressing others whose IoU with it exceeds a threshold.

Contrast two-stage detectors (e.g., Faster R-CNN) with one-stage detectors (e.g., YOLO).

Two-stage detectors first propose candidate regions then classify/refine them, giving high accuracy; one-stage detectors predict boxes and classes directly in a single pass over a grid, giving faster, often real-time, inference.

Distinguish semantic segmentation from instance segmentation.

Semantic segmentation labels every pixel with a class but does not separate individual objects of the same class; instance segmentation additionally distinguishes separate object instances (each car gets its own mask).

What is the role of the encoder-decoder (e.g., U-Net) structure in image segmentation?

The encoder progressively downsamples to capture context/semantic features; the decoder upsamples to recover spatial resolution, with skip connections passing fine-grained details from encoder to decoder for precise pixel-level masks.

What is the Otsu method used for in image segmentation?

Otsu's method automatically selects a global threshold for binarization by choosing the value that maximizes between-class variance (equivalently minimizes intra-class variance) of the pixel intensity histogram.

Distinguish proprioceptive from exteroceptive sensors on a robot, with examples.

Proprioceptive sensors measure the robot's internal state (e.g., encoders, IMU/gyroscope, motor current); exteroceptive sensors measure the external environment (e.g., cameras, LIDAR, sonar, GPS).

How does a LIDAR sensor measure distance?

By time-of-flight: it emits laser pulses and measures the round-trip time $t$; distance $d = \frac{c\,t}{2}$, where $c$ is the speed of light. Scanning yields a point cloud of the environment.

Define a robot's degrees of freedom (DOF).

The number of independent parameters (joint variables) needed to fully specify the configuration of the robot. A manipulator with $n$ independently actuated joints has $n$ degrees of freedom.

What is the difference between forward and inverse kinematics for a robot arm?

Forward kinematics computes the end-effector pose from given joint angles (unique solution); inverse kinematics computes the joint angles required to achieve a desired end-effector pose (may have multiple, one, or no solutions).

What problem does SLAM solve in robotics?

Simultaneous Localization and Mapping: a robot builds a map of an unknown environment while simultaneously estimating its own pose within that map, using sensor and motion data.

State the two-step cycle of a Kalman filter used in robot localization.

Prediction: propagate the state estimate and covariance forward using the motion model; Update (correction): incorporate a new measurement to refine the estimate, weighting by the Kalman gain $K = \frac{P H^{\top}}{H P H^{\top} + R}$.

How does a particle filter (Monte Carlo Localization) represent the robot's belief?

As a set of weighted samples (particles), each a hypothesis of the robot's pose. Particles are propagated by the motion model, reweighted by sensor likelihood, and resampled in proportion to their weights.

Compare grid-based (A*) motion planning with sampling-based (RRT) planning.

A* searches a discretized grid/graph for an optimal path using cost-plus-heuristic $f(n)=g(n)+h(n)$, complete and optimal but scales poorly in high dimensions; RRT randomly samples the continuous configuration space to grow a tree, efficient in high-DOF spaces but only probabilistically complete and not optimal (RRT* is asymptotically optimal).

What is the configuration space (C-space) in motion planning?

The space of all possible robot configurations. Planning finds a collision-free path in the free portion $C_{free}$; obstacles in the workspace are expanded into C-space obstacles, and the robot is treated as a point in C-space.

In A* search, what condition must the heuristic satisfy to guarantee an optimal path?

The heuristic must be admissible: it never overestimates the true cost to the goal, i.e., $h(n) \leq h^{*}(n)$ for all nodes $n$. (Consistency additionally guarantees efficiency.)

Distinguish open-loop from closed-loop (feedback) control in robotics.

Open-loop control sends commands without using sensor feedback, so it cannot correct disturbances; closed-loop control measures the output and uses the error between desired and actual state to adjust commands.

Write the PID control law and name what each term addresses.

$u(t) = K_p e(t) + K_i \int_{0}^{t} e(\tau)\,d\tau + K_d \frac{de(t)}{dt}$. Proportional reacts to present error, Integral eliminates steady-state error from accumulated past error, Derivative dampens by anticipating future error.

What is compliance (or force/impedance) control and why is it important in manipulation?

It regulates the interaction force between the manipulator and its environment rather than pure position, allowing the robot to comply with contact. This is essential for tasks like grasping, insertion, and assembly to avoid damage and handle uncertainty.

Distinguish weak (narrow) AI from strong (general) AI.

Weak/narrow AI performs specific tasks and only simulates intelligence within a limited domain; strong AI (AGI) would possess genuine, general, human-level understanding and consciousness across arbitrary tasks. Essentially all deployed systems today are narrow AI.

What is Searle's Chinese Room argument meant to demonstrate?

It argues that a system manipulating symbols by rules (as a program does) can produce correct outputs without any genuine understanding, challenging strong AI's claim that the right program constitutes a mind (syntax is not sufficient for semantics).

What does the Turing Test assess, and what is a key criticism of it?

It assesses whether a machine's conversational behavior is indistinguishable from a human's to an interrogator (behavioral test of intelligence). A key criticism is that it measures imitation of behavior, not genuine understanding or consciousness.

Define algorithmic bias in AI and give a common source.

Algorithmic bias is systematic, unfair discrimination in a model's outputs against certain groups. A common source is biased or unrepresentative training data that reflects historical or societal prejudices, which the model then learns and amplifies.

What is the value alignment problem in AI safety?

The challenge of ensuring an AI system's objectives and behavior remain aligned with human values and intentions, so that a capable system pursues what we actually want rather than a misspecified proxy goal.