Intermediate everything in AI & ML, page 29
Region Proposal Network (RPN): The 'Where to Look' Engine
An RPN is an 'attention' mechanism for object detection, telling the model where to look. It replaces slow, external proposal methods in systems like Faster R-CNN for autonomous driving.
AlexNet: The CNN That Sparked the Deep Learning Boom
AlexNet is the blueprint that proved deep CNNs could master image recognition, kicking off the modern AI boom. Its architecture is foundational for modern computer vision. The footgun is thinking it was just bigger; its novelty was combining new techniques.
Dropout: Forcing a Network to Generalize
Dropout prevents overfitting by randomly zeroing out a fraction of neurons during training. This forces the network to learn more robust features instead of relying on specific neurons. It's a standard regularizer for large, dense layers.
Data Augmentation: Getting More Images for Free
Data augmentation creates "fake" training data by modifying existing images—flipping, rotating, or color-shifting them. This fights overfitting when your dataset is small, forcing the model to generalize.
Backpropagation: How Neural Networks Learn from Mistakes
Backpropagation is how a network learns from its mistakes. It works backward from the output error, calculating how much each weight contributed and adjusting it. This is the core training loop for most deep learning models.
Cross-Entropy Loss: How Wrong Is Your Model's Guess?
Cross-entropy loss measures the penalty when a model's predicted probabilities diverge from the true labels. It's the standard loss for classification tasks, like telling a cat from a dog.
ReLU: The 'On/Off' Switch for Neural Networks
ReLU acts as a simple on/off switch for neurons: positive inputs pass through, negative ones become zero. It's the default activation in deep learning, especially for vision tasks, as it's fast and helps gradients flow. The footgun: neurons can "die".
Structure from Motion (SfM): Rebuilding 3D from 2D Photos
Structure from Motion (SfM) reconstructs a 3D scene from 2D images taken from different viewpoints, much like your brain perceives depth. It's used for creating 3D maps from drone footage. The main footgun is its failure on textureless or reflective surfaces.
Triangulation: Finding 3D Points from 2D Images
Like your two eyes judging distance, triangulation finds a point's 3D location by seeing it from two different 2D camera views. It's key for 3D reconstruction and robotics.
The Essential Matrix: Geometry for Stereo Vision
The Essential Matrix is a geometric bridge between two photos of the same scene, encoding the camera's relative motion. It's used in 3D reconstruction to find corresponding points; it tells you where a point from one image must lie in the other.
The Fundamental Matrix: Constraining 3D Search to a Line
The Fundamental Matrix reduces a 2D search for a point in a second image to a 1D search along a line. It's the geometric glue linking two uncalibrated camera views of the same scene, crucial for stereo vision and 3D reconstruction.
Homography: Mapping Flat Surfaces Between Images
A homography is a 3x3 matrix that maps points between two images of a flat surface. It's used for panorama stitching and perspective correction. The footgun: it only works for planes or pure camera rotation, failing on scenes with depth.
SIFT: Finding Image Features That Scale
SIFT finds key points in an image that are stable across changes in scale. It's used for object recognition or stitching panoramas by matching these features between images. The main footgun is assuming it's invariant to all changes, not just scale.
Histogram of Oriented Gradients (HOG)
HOG describes an object's shape by summarizing the directions of local intensity changes (gradients). It's a classic computer vision method for tasks like pedestrian detection where form matters more than color.
FAST: High-Speed Corner Detection for Real-Time Vision
FAST finds corners by checking if a pixel is significantly brighter or darker than a ring of its neighbors. This simple, high-speed test makes it perfect for real-time video processing.
Difference of Gaussians: Finding Edges by Subtracting Blurs
Difference of Gaussians (DoG) finds features by subtracting a heavily blurred image from a lightly blurred one, isolating details at a specific scale. It's used for fast blob and edge detection.
Scale-Space: Analyzing Images at Multiple Scales
Scale-space representation is like viewing an image from far away and up close simultaneously. It creates a stack of progressively blurred images to reveal structures at different sizes, which is key for robust feature detection.
Morphological Transformations: Shaping Images with Kernels
Think of it as sanding or stamping an image with a small shape (kernel) to modify object boundaries. It's used for noise removal or separating touching objects. The footgun: transformations alter object size, so you often need a second operation to compensate.
The Sobel Operator: Fast, Cheap Edge Detection
The Sobel operator finds image edges by measuring how fast pixel brightness changes horizontally and vertically. It's a fast, cheap first pass for edge detection in computer vision. The footgun is treating it as precise; it's a crude approximation.
Median Filter: Smoothing Images Without Blurring Edges
A median filter cleans image noise by replacing each pixel with the median value of its neighbors, ignoring outliers. It's used to remove "salt-and-pepper" noise before object detection.
We are hiring for this. Every open role lists the topics its interview covers, so you can prepare for the real thing rather than guessing.
See open roles