COE 592Lecture 00Glossary

Glossary

Every term in Course overview, defined once and used the same way in every part. Each entry links to the slides where the idea appears.

Terms
57
Letters
20

A

Activations

The intermediate outputs each layer produces for one input; during inference they must be held in memory alongside the weights, and during training they are also stored for the backward pass.

Adversarial machine learning

The study of attacks that manipulate a model's inputs or training to make it fail, and of defenses that make models robust to them.

Automated visual inspection

Using cameras and vision models to find defects, such as cracks or missing parts on rail car undercarriages, without manual inspection.

B

Bit width

The number of bits used to store one weight or activation; weight size in bytes is parameter count times bit width divided by eight.

Blackboard

KFUPM's learning management system, where the COE 592 syllabus is posted.

C

Confidence score

The number a detector attaches to each box, such as UAV 0.93, expressing how certain it is of the class; boxes below a chosen threshold are discarded.

Confidence threshold

The cut-off below which a detector discards its boxes; raising it removes low-confidence false positives but also drops correct boxes that scored low.

Convolutional neural network

A neural network that learns filters applied across its input by convolution, widely used for images and adapted in PNet-IDS to network traffic features.

D

Deployment

Implementing a model on the target hardware and software stack so it runs in the real application.

Distribution shift

A change between the data a model was trained on and the data it meets in deployment, such as rain-degraded images or unseen network traffic; also called domain shift.

E

EcoWeedNet

A lightweight, computationally and energy-efficient weed detection model proposed for low-carbon consumer electronics in precision agriculture.

Edge AI

Running machine learning inference on or near the device that captures the data, instead of sending the data to a distant cloud server.

Efficiency metrics

Measures of the cost of deep learning computing, such as model size, memory, compute, latency, throughput and energy, used alongside accuracy.

Embedded system

A computer built into a larger device to perform a dedicated function under tight limits on memory, compute, power and cost, such as the processor on a drone, robot or IoT sensor.

F

False positive

A detection of something that is not there or is the wrong class, such as YOLOv5-m labelling a bird as a UAV with confidence 0.29.

Flash memory

The non-volatile storage on a microcontroller that holds program code and, in a deployed model, the weights; its size caps the model size that can be stored.

FLOPs

Floating point operations needed for one inference, a hardware-independent measure of compute cost; a GFLOP is a billion of them.

G

GradCAM++

A gradient-based visualization that produces a heatmap of the image regions that most influenced a network's prediction.

I

Inference

Running a trained neural network on new inputs to produce predictions.

Intelligent transportation systems

Applying sensing, communication and analytics to road and rail transport to improve safety and efficiency, for example detecting conflicts between turning vehicles and pedestrians.

Internet of Things

Networks of connected physical devices and sensors that exchange data, typically with limited compute and energy.

Intrusion detection system

A system that monitors network traffic or host activity and flags malicious behaviour or attacks.

K

Knowledge distillation

Training a small student network to reproduce the outputs of a larger teacher network, so the student keeps most of the teacher's accuracy at a fraction of its size; PNet-IDS uses it to hold up under distribution shift.

L

Latency

The time from one input reaching a model to its output being ready; for a moving vehicle or a passing wagon the scene changes while it elapses, so it is a safety budget as well as a performance metric.

Left-turn pedestrian conflict

The intersection safety pattern in which a vehicle turning left across oncoming traffic enters a crosswalk a pedestrian is already using, drawn on slide 7 as a red vehicle arrow crossing a blue pedestrian arrow.

LiDAR

Light detection and ranging: a sensor that measures distance with laser pulses and produces a 3D point cloud of the scene.

Lightweight model

A neural network designed with few parameters and low compute so it can run on constrained hardware with acceptable accuracy.

Line rate

The full data rate of a network link; a detector runs at line rate when it decides on every packet or flow as fast as the link delivers them, without falling behind.

M

mAP

Mean average precision: the standard object detection accuracy score, averaging precision over recall levels and classes; mAP@0.5 counts a box correct at 50% overlap with the truth, and mAP50-95 averages over overlap thresholds from 50% to 95%.

Microcontroller

A single chip that combines a small processor, flash memory and SRAM for a dedicated task, typically with kilobytes to a few megabytes of memory and no operating system; the smallest deployment target in the course.

Model compression

Techniques that shrink a trained model's size and compute, such as pruning and quantization, so it can be deployed on constrained devices.

Multi-sensor platform

A vehicle or UAV carrying several sensors, such as visual and thermal cameras, LiDAR and GPS, whose data are combined for inspection.

N

O

Object detection

The task of locating objects in an image with bounding boxes and assigning each box a class label and a confidence score.

Optimization tradeoff

The balance between the benefits of an optimization technique, such as smaller size or lower latency, and its costs, such as lost accuracy or extra engineering.

Orthomosaic

A top-down image of an area assembled from many overlapping aerial photographs, each corrected for perspective and placed by its GPS position, such as the track corridor strips on slide 3.

P

Parameter count

The number of learnable weights and biases in a model, usually quoted in millions; it sets the memory needed to store the model, such as 2.6M for YOLO11n.

PNet-IDS

A lightweight and generalizable convolutional neural network for intrusion detection in the Internet of Things, published in IEEE Access in 2025.

Point cloud

A set of 3D points sampled from surfaces in a scene, produced by LiDAR or photogrammetry and often coloured by elevation.

Pre-course survey

The form students fill in at the start of the course so the instructor learns their background, skills and expectations.

Precision

The fraction of a detector's reported boxes that are correct: true positives divided by true positives plus false positives.

Precision agriculture

Farming that uses sensors, robots and data analysis to act on individual plants or small areas, for example automated weed detection.

Pruning

Removing weights, connections or whole channels that contribute little to a network's output, then usually retraining, so the model has fewer parameters and operations.

Q

Quantization

Storing and computing a network's weights and activations in fewer bits, for example 8-bit integers instead of 32-bit floats, which divides model size and speeds up arithmetic at some accuracy cost.

R

Recall

The fraction of the real objects a detector finds: true positives divided by true positives plus false negatives.

Resource-constrained device

Hardware whose memory, storage, compute and energy budget are small enough that a deep learning model must be optimized before it can run on it.

Rolling stock

The wheeled vehicles that run on a railway, such as wagons and carriages, whose undercarriages are photographed for inspection on slide 8.

S

SDAIA-KFUPM JRC for AI

The SDAIA-KFUPM Joint Research Center for Artificial Intelligence, where the instructor is a research scholar.

SERV Lab

The Secure, Efficient, Robust Vision lab led by Dr. Abdul Jabbar Siddiqui at KFUPM.

SRAM

Static random-access memory, the small, fast working memory of a microcontroller that must hold the activations during inference; it is usually far smaller than flash.

T

Thermal camera

A camera that images infrared radiation, so it shows temperature differences rather than visible colour; labelled IR camera on the slide.

Throughput

The number of inferences a system completes per second; it can rise through batching while the latency of each individual input stays the same or grows.

TinyML

Machine learning that runs on microcontrollers and other devices with kilobytes of memory and milliwatts of power; the field this course's deployment work belongs to.

Training

Adjusting a neural network's weights from data, which needs far more memory and compute than inference.

U

UAV

Unmanned aerial vehicle: an aircraft without an on-board pilot, used here as a sensor platform for inspection and monitoring.

Y

YOLO

You Only Look Once: a family of single-stage object detectors that predict boxes and class scores in one network pass; the n suffix (YOLO11n, YOLO12n) marks the smallest nano variant and m the medium one.

YOLO-RAW

The lab's detector shown correctly separating birds from UAVs in rain-degraded images where YOLOv5-m confused them.