<\/p>\n\n\n\n
Task 3 (Text Generation):
a) What is tokenization?
Basically, tokenization deals with the segmenting process of a whole text into minimal units such
as words, subwords, and\/or even characters for a machine to work on (Islam et al., 2021).
b) Why do we need to tokenize the input text?
Text tokenization is essential because neural networks can’t process raw text directly. It converts
text into numerical sequences that machines can understand and process (Islam et al., 2021).
c) How can we tokenize the input text?
Word-based tokenization: Splitting by spaces or punctuation.
Subword tokenization: Methods like BPE (Byte Pair Encoding) or WordPiece.
Character-based tokenization: Splitting text into individual characters.
Sentence tokenization: Dividing text into sentences (Islam et al., 2021).
Task 4: NLP and Sequence Models
a) What is Natural Language Processing (NLP)?
Natural Language Processing alludes to a technique of using computational methods to make
computers generate and comprehend human languages (Min et al., 2023).
b) Sequence Models in Deep Learning:
Sequence models process data where order matters, like text or time series. They maintain
context across inputs, essential for understanding language structure and meaning.
c) Sequence Model Types:
One-to-One: Single input to single output (e.g., standard classification tasks).
One-to-Many: Single input generating multiple outputs (e.g., image captioning).
3
Many-to-One: Multiple inputs producing a single output (e.g., sentiment analysis)..
Many-to-Many: Multiple inputs generating multiple outputs (e.g., language translation) (Smith
et al., 2022).
Task 6: Image Generation Challenges
a) Diffusion Model Challenges:
High Computational Cost: Training and inference require significant computational resources
due to the iterative denoising process.
Slow Generation Speed: Diffusion models are slower compared to GANs because of their
sequential denoising nature (Huang et al., 2024).
b) Conditional GAN Operation:
Conditional GANs incorporate additional input (e.g., labels or conditions) alongside random
noise to control generation. The discriminator evaluates both the generated image and the
conditioning information, enabling targeted content generation.
Task 7: Advanced GenAI Concepts
a) Parallelization:
Parallelization is the process of splitting computations across multiple processors or GPUs to
perform tasks simultaneously, increasing efficiency and speed (Luitse and Denkena 2021).
b) End-to-End Deep Learning:
The system processes raw input directly to the desired output without manual feature engineering
and learns all transformations automatically (Qi et al., 2023).
c) Video Generation Challenges:
Temporal Consistency: Ensuring realistic motion and seamless transitions across frames while
maintaining coherence.
4
High Computational Requirements: Generating multiple frames while preserving spatialtemporal relationships demands significant resources.
5
References
Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S. and Cao,
L. (2024). Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525.
Islam, T., Hossain, M. and Arefin, M.F. (2021, December). Comparative analysis of different
text summarization techniques using enhanced tokenization. In 2021 3rd International
Conference on Sustainable Technologies for Industry 4.0 (STI) (pp. 1-6). IEEE.
Luitse, D. and Denkena, W. (2021). The great transformer: Examining the role of large language
models in the political economy of AI. Big Data & Society, 8(2), p.20539517211047734.
Min, B., Ross, H., Sulem, E., Veyseh, A.P.B., Nguyen, T.H., Sainz, O., Agirre, E., Heintz, I. and
Roth, D. (2023). Recent advances in natural language processing via large pre-trained language
models: A survey. ACM Computing Surveys, 56(2), pp.1-40.
Smith, J.T., Warrington, A. and Linderman, S.W. (2022). Simplified state space layers for
sequence modeling. arXiv preprint arXiv:2208.04933.
Qi, M., Shi, Y., Qi, Y., Ma, C., Yuan, R., Wu, D. and Shen, Z.J. (2023). A practical end-to-end
inventory management model with deep learning. Management Science, 69(2), pp.759-773.<\/p>\n\n\n\n
<\/p>\n\n\n\n
\n- Use IoT devices to monitor player movements in sports such as:
o Soccer: Determination of player positioning and their levels of energy.
o Basketball: Collect data on the heights of jumps and player ability to move
quickly during the game.
o Rugby: Measure the efficiency of sprints and tackling methods.
Impact:<\/li>\n\n\n\n - Enhances team results by establishing player strengths and weaknesses.<\/li>\n\n\n\n
- Offering coaches with comprehensive game analytics (Yuan et al., 2022).
Future Research:<\/li>\n\n\n\n - Integration with drones for synchronized aerial and ground-level motion analysis.<\/li>\n\n\n\n
- Player performance assessments using artificial intelligence for creation of individual
training programmes.
Example 2: Fall Prevention for Elderly Care
Application:<\/li>\n\n\n\n - IoT wearables detect the gait patterns of elderly people and monitor them.<\/li>\n\n\n\n
- Signal caregivers in cases of detection of irregular movements or when a fall may occur.
Impact:<\/li>\n\n\n\n - Enhances safety and independence for the elderly.<\/li>\n\n\n\n
- Reduces the cost of health care by preventing fall injuries (Kulurkar et al., 2023).
Future Research:<\/li>\n\n\n\n - Advanced AI algorithms to predict fall risks based on historical motion data.<\/li>\n\n\n\n
- Smart home system integrations for automatic responses in case of emergencies.
3
Example 3: Motion Monitoring in Fitness Apps
Application:<\/li>\n\n\n\n - Wearable fitness devices can track workout efficiency in various activities including the
following:
o Yoga: Evaluate the correctness and stability of the pose.
o Weightlifting: Keep track of the posture and range of motion.
o Cycling: Cadence and energy efficiency analysis.
Impact:<\/li>\n\n\n\n - Personalizes fitness programs with individual performance data.<\/li>\n\n\n\n
- Enhances engagement through feedback provided in real time and indicating progress.
Future Research:<\/li>\n\n\n\n - Integration of virtual reality into fitness for better immersion.<\/li>\n\n\n\n
- New, wearable designs that offer unhindered motion tracking for several activities (Cai,
2022).
Task 2 \u2013 Individual Research: Tools for Motion Analysis
Summary of Findings<\/li>\n<\/ul>\n\n\n\n\n- Other Tools for Motion Analysis in Sports and Health Contexts
a. IMU (Inertial Measurement Unit) Sensors<\/li>\n<\/ol>\n\n\n\n\n- What It Is: Small devices for measuring acceleration, angular speed, and orientation.<\/li>\n\n\n\n
- Applications:
o Sports: Monitor biomechanics when it comes to exercising, for instance, jogging,
swimming or cycling.
o Healthcare: Controlling the mobility of the concerned patient during the
rehabilitation exercise (Samatas and Pachidis, 2022).<\/li>\n\n\n\n - Example Tools:
4
o Perception Neuron: An inexpensive wearable system for full body tracking.
o Xsens: High-precision sensors for biomechanical investigations.
b. Marker-Based Motion Capture Systems<\/li>\n\n\n\n - What It Is: Systems that rely on camera to detect reflectance markers of the body
(Kanko et al., 2021).<\/li>\n\n\n\n - Applications:
o Sports: Assess some intricate patterns in either gymnastics or dance.
o Healthcare: Applied in the clinical evaluation of walking pattern and in the
design of artificial limbs.<\/li>\n\n\n\n - Example Tools:
o Vicon: The high-accuracy systems used in biomechanics are motion capture
systems.
o Qualisys: The latest technological equipment for three-dimensional motion
tracking in special laboratories.
c. Markerless Motion Capture<\/li>\n\n\n\n - What It Is: AI-based systems that doesn’t take use of physical characteristics (Kanko et
al., 2021).<\/li>\n\n\n\n - Applications:
o Sports: Evaluate current athletic performance during a certain sport.
o Healthcare: Management of natural movements during rehabilitation of patient
that had a stroke.<\/li>\n\n\n\n - Example Tools:
o OpenPose: Open-source tool that estimates pose.
o Kinovea: Free software analyzing motion from videos
5
d. Pressure and Force Platforms<\/li>\n\n\n\n - What It Is: Tools that measure ground reaction forces during movements.<\/li>\n\n\n\n
- Applications:
o Sports: Evaluation of jumps and sprint starts.
o Healthcare: Assess balance and fall risks in elderly individuals (Wang et al.,
2022).<\/li>\n\n\n\n - Example Tools:
o AMTI: Force plates for biomechanical research.
o Bertec: Specific systems intended for gait and balance analysis.
e. Smart Wearables<\/li>\n\n\n\n - What It Is: The wearable devices popular among consumers as fitness and health
tracking gadgets.<\/li>\n\n\n\n - Applications:
o Sports: Monitor metrics such as, heart rate, steps, and general body motions.
o Healthcare: Keep track of physical activity as well as quality of sleep.<\/li>\n\n\n\n - Example Tools:
o Garmin and Fitbit: Popular fitness trackers.
o WHOOP: A performance and recovery monitoring wearable.<\/li>\n<\/ul>\n\n\n\n\n- Research Trends and Directions
a. Real-Time Feedback Systems<\/li>\n<\/ol>\n\n\n\n\n- Tools are increasingly focused on delivering instant feedback to improve performance or
recovery.<\/li>\n\n\n\n - Examples: Smart glasses for information purposes, and vibrating feedback devices in
wearable technology.
6
b. AI Integration<\/li>\n\n\n\n - AI improves motion data interpretation by enabling predictions and further analysis.<\/li>\n\n\n\n
- Examples: Machine learning models for estimating the probability of an injury from the
movement profile.
c. Remote Monitoring<\/li>\n\n\n\n - Due to IoT, patient or athlete tracking systems are possible through cloud platforms.
d. Immersive Visualization<\/li>\n\n\n\n - Utilization of augmented and virtual reality for the visualization of biomechanical data.
7
References
Cai, H. (2022). Application of intelligent real-time image processing in fitness motion detection
under internet of things. The Journal of Supercomputing, 78(6), pp.7788-7804.
Kanko, R.M., Laende, E.K., Davis, E.M., Selbie, W.S. and Deluzio, K.J. (2021). Concurrent
assessment of gait kinematics using marker-based and markerless motion capture. Journal of
biomechanics, 127, p.110665.
Kulurkar, P., kumar Dixit, C., Bharathi, V.C., Monikavishnuvarthini, A., Dhakne, A. and Preethi,
P. (2023). AI based elderly fall prediction system using wearable sensors: A smart home-care
technology with IOT. Measurement: Sensors, 25, p.100614.
Samatas, G.G. and Pachidis, T.P. (2022). Inertial measurement units (imus) in mobile robots
over the last five years: A review. Designs, 6(1), p.17.
Yuan, Y., Lu, Z., Yang, Z., Jian, M., Wu, L., Li, Z. and Liu, X. (2022). Key frame extraction
based on global motion statistics for team-sport videos. Multimedia Systems, 28(2), pp.387-401.
Wang, J., Zhu, Y., Wu, Z., Zhang, Y., Lin, J., Chen, T., Liu, H., Wang, F. and Sun, L. (2022).
Wearable multichannel pulse condition monitoring system based on flexible pressure sensor
arrays. Microsystems & Nanoengineering, 8(1), p.16.<\/li>\n<\/ul>\n\n\n\n1
Week 5 Exercises
(By Name)
Course
University
Professor
State
City
Date
2
Exercises and Answers
Exercise 1: Application Area of Computer Vision
Favorite Application: Tesla’s Autonomous Driving
Tesla employs computer vision in its Autopilot and Full Self-Driving (FSD) systems to enable
real-time navigation with minimal human intervention. Key components include neural networks
trained on camera feeds, ultrasonic sensors, and radar.<\/p>\n\n\n\n
\n- Examples:
o Tesla\u2019s Smart Summon feature autonomously drives cars to owners in parking
lots.
o AI Day presentations demonstrate the vision-first approach, avoiding LiDAR and
focusing on camera-based perception.
o Dojo Supercomputer trains its neural networks using millions of driving videos.<\/li>\n\n\n\n - Videos and research on Tesla\u2019s capabilities highlight advancements in lane detection,
object prediction, and traffic analysis.
Exercise 2: Deep Fake Videos
Observations:<\/li>\n<\/ul>\n\n\n\n\n- Best Video: The Barack Obama deep fake video created by Jordan Peele. It convincingly
showcased how AI can manipulate speech and facial expressions.<\/li>\n\n\n\n - How They Did It: This video was created using GANs (Generative Adversarial
Networks) and facial mapping technology. Specific tools like DeepFaceLab or customtrained models were used to replicate Obama\u2019s speech patterns and expressions.<\/li>\n\n\n\n - Time and Hardware: Likely required weeks of training on high-end GPUs, given its
quality and precision.
3<\/li>\n\n\n\n - Realism: While highly realistic, close inspection reveals slight lip-sync inconsistencies
and unnatural transitions in facial movements.
Exercise 3: Wacky Object Variations
Category: Potted Plants<\/li>\n<\/ol>\n\n\n\n\n- Found images include:
o A plant shaped like a cat.
o A pot designed as a robot holding a small cactus.
o A plant growing from a shoe instead of a traditional pot.
Exercise 4: Rise of Deep Learning
Key Points:<\/li>\n<\/ul>\n\n\n\n\n- Trends in GPU Price Performance:<\/li>\n<\/ol>\n\n\n\n
\n- GPUs have improved dramatically in computational power while becoming more
affordable, enabling broader adoption of deep learning.<\/li>\n<\/ul>\n\n\n\n\n- NVIDIA and AI Dominance:<\/li>\n<\/ol>\n\n\n\n
\n- NVIDIA\u2019s GPUs revolutionized AI training, surpassing CPU-based methods,
with partnerships accelerating AI integration in diverse industries.
Exercise 5: Object Detection<\/li>\n\n\n\n - Tested images include the unusual plants from Exercise 3.<\/li>\n\n\n\n
- Observations:
o Accurate detection of plant-like shapes.
o Unique pots were classified as unrelated objects like furniture.
4
Exercise 6: Image Classification
Observations:<\/li>\n<\/ul>\n\n\n\n\n- Single Object: A clean headshot was identified with high confidence and minimal
ambiguity.<\/li>\n\n\n\n - Cluttered Scene: A coffee cup amidst a desk setup confused the model, leading to
multiple low-confidence classifications.
Exercise 7: One-Stage vs Two-Stage Models
Models:<\/li>\n<\/ol>\n\n\n\n\n- YOLO (One-Stage):
o Speed: ~45 FPS
o Accuracy: ~55 mAP on COCO dataset.<\/li>\n\n\n\n - Faster-RCNN (Two-Stage):
o Speed: ~5 FPS
o Accuracy: ~60-65 mAP on COCO dataset.
Comparison: YOLO prioritizes speed for real-time applications, while Faster-RCNN offers
higher accuracy at the cost of slower inference.
Bonus Insights:<\/li>\n\n\n\n - DETR achieves mAP scores around 65-70 on COCO but has slower inference times (~10
FPS).<\/li>\n<\/ul>\n\n\n\nExercises and Answers
Exercise 1: Application Area of Computer Vision
Favorite Application: Tesla’s Autonomous Driving
Tesla employs computer vision in its Autopilot and Full Self-Driving (FSD) systems to enable
real-time navigation with minimal human intervention. Key components include neural networks
trained on camera feeds, ultrasonic sensors, and radar.<\/p>\n\n\n\n
\n- Examples:
o Tesla\u2019s Smart Summon feature autonomously drives cars to owners in parking
lots.
o AI Day presentations demonstrate the vision-first approach, avoiding LiDAR and
focusing on camera-based perception.
o Dojo Supercomputer trains its neural networks using millions of driving videos.<\/li>\n\n\n\n - Videos and research on Tesla\u2019s capabilities highlight advancements in lane detection,
object prediction, and traffic analysis.
Exercise 2: Deep Fake Videos
Observations:<\/li>\n<\/ul>\n\n\n\n\n- Best Video: The Barack Obama deep fake video created by Jordan Peele. It convincingly
showcased how AI can manipulate speech and facial expressions.<\/li>\n\n\n\n - How They Did It: This video was created using GANs (Generative Adversarial
Networks) and facial mapping technology. Specific tools like DeepFaceLab or customtrained models were used to replicate Obama\u2019s speech patterns and expressions.<\/li>\n\n\n\n - Time and Hardware: Likely required weeks of training on high-end GPUs, given its
quality and precision.
3<\/li>\n\n\n\n - Realism: While highly realistic, close inspection reveals slight lip-sync inconsistencies
and unnatural transitions in facial movements.
Exercise 3: Wacky Object Variations
Category: Potted Plants<\/li>\n<\/ol>\n\n\n\n\n- Found images include:
o A plant shaped like a cat.
o A pot designed as a robot holding a small cactus.
o A plant growing from a shoe instead of a traditional pot.
Exercise 4: Rise of Deep Learning
Key Points:<\/li>\n<\/ul>\n\n\n\n\n- Trends in GPU Price Performance:<\/li>\n<\/ol>\n\n\n\n
\n- GPUs have improved dramatically in computational power while becoming more
affordable, enabling broader adoption of deep learning.<\/li>\n<\/ul>\n\n\n\n\n- NVIDIA and AI Dominance:<\/li>\n<\/ol>\n\n\n\n
\n- NVIDIA\u2019s GPUs revolutionized AI training, surpassing CPU-based methods,
with partnerships accelerating AI integration in diverse industries.
Exercise 5: Object Detection<\/li>\n\n\n\n - Tested images include the unusual plants from Exercise 3.<\/li>\n\n\n\n
- Observations:
o Accurate detection of plant-like shapes.
o Unique pots were classified as unrelated objects like furniture.
4
Exercise 6: Image Classification
Observations:<\/li>\n<\/ul>\n\n\n\n\n- Single Object: A clean headshot was identified with high confidence and minimal
ambiguity.<\/li>\n\n\n\n - Cluttered Scene: A coffee cup amidst a desk setup confused the model, leading to
multiple low-confidence classifications.
Exercise 7: One-Stage vs Two-Stage Models
Models:<\/li>\n<\/ol>\n\n\n\n\n- YOLO (One-Stage):
o Speed: ~45 FPS
o Accuracy: ~55 mAP on COCO dataset.<\/li>\n\n\n\n - Faster-RCNN (Two-Stage):
o Speed: ~5 FPS
o Accuracy: ~60-65 mAP on COCO dataset.
Comparison: YOLO prioritizes speed for real-time applications, while Faster-RCNN offers
higher accuracy at the cost of slower inference.
Bonus Insights:<\/li>\n\n\n\n - DETR achieves mAP scores around 65-70 on COCO but has slower inference times (~10
FPS).<\/li>\n<\/ul>\n\n\n\nExercise 1: Application Area of Computer Vision
Favorite Application: Tesla’s Autonomous Driving
Tesla employs computer vision in its Autopilot and Full Self-Driving (FSD) systems to enable
real-time navigation with minimal human intervention. Key components include neural networks
trained on camera feeds, ultrasonic sensors, and radar.<\/p>\n\n\n\n
\n- Examples:
o Tesla\u2019s Smart Summon feature autonomously drives cars to owners in parking
lots.
o AI Day presentations demonstrate the vision-first approach, avoiding LiDAR and
focusing on camera-based perception.
o Dojo Supercomputer trains its neural networks using millions of driving videos.<\/li>\n\n\n\n - Videos and research on Tesla\u2019s capabilities highlight advancements in lane detection,
object prediction, and traffic analysis.
Exercise 2: Deep Fake Videos
Observations:<\/li>\n<\/ul>\n\n\n\n\n- Best Video: The Barack Obama deep fake video created by Jordan Peele. It convincingly
showcased how AI can manipulate speech and facial expressions.<\/li>\n\n\n\n - How They Did It: This video was created using GANs (Generative Adversarial
Networks) and facial mapping technology. Specific tools like DeepFaceLab or customtrained models were used to replicate Obama\u2019s speech patterns and expressions.<\/li>\n\n\n\n - Time and Hardware: Likely required weeks of training on high-end GPUs, given its
quality and precision.
3<\/li>\n\n\n\n - Realism: While highly realistic, close inspection reveals slight lip-sync inconsistencies
and unnatural transitions in facial movements.
Exercise 3: Wacky Object Variations
Category: Potted Plants<\/li>\n<\/ol>\n\n\n\n\n- Found images include:
o A plant shaped like a cat.
o A pot designed as a robot holding a small cactus.
o A plant growing from a shoe instead of a traditional pot.
Exercise 4: Rise of Deep Learning
Key Points:<\/li>\n<\/ul>\n\n\n\n\n- Trends in GPU Price Performance:<\/li>\n<\/ol>\n\n\n\n
\n- GPUs have improved dramatically in computational power while becoming more
affordable, enabling broader adoption of deep learning.<\/li>\n<\/ul>\n\n\n\n\n- NVIDIA and AI Dominance:<\/li>\n<\/ol>\n\n\n\n
\n- NVIDIA\u2019s GPUs revolutionized AI training, surpassing CPU-based methods,
with partnerships accelerating AI integration in diverse industries.
Exercise 5: Object Detection<\/li>\n\n\n\n - Tested images include the unusual plants from Exercise 3.<\/li>\n\n\n\n
- Observations:
o Accurate detection of plant-like shapes.
o Unique pots were classified as unrelated objects like furniture.
4
Exercise 6: Image Classification
Observations:<\/li>\n<\/ul>\n\n\n\n\n- Single Object: A clean headshot was identified with high confidence and minimal
ambiguity.<\/li>\n\n\n\n - Cluttered Scene: A coffee cup amidst a desk setup confused the model, leading to
multiple low-confidence classifications.
Exercise 7: One-Stage vs Two-Stage Models
Models:<\/li>\n<\/ol>\n\n\n\n\n- YOLO (One-Stage):
o Speed: ~45 FPS
o Accuracy: ~55 mAP on COCO dataset.<\/li>\n\n\n\n - Faster-RCNN (Two-Stage):
o Speed: ~5 FPS
o Accuracy: ~60-65 mAP on COCO dataset.
Comparison: YOLO prioritizes speed for real-time applications, while Faster-RCNN offers
higher accuracy at the cost of slower inference.
Bonus Insights:<\/li>\n\n\n\n - DETR achieves mAP scores around 65-70 on COCO but has slower inference times (~10
FPS).<\/li>\n<\/ul>\n\n\n\nExercises and Answers
Exercise 1: Application Area of Computer Vision
Favorite Application: Tesla’s Autonomous Driving
Tesla employs computer vision in its Autopilot and Full Self-Driving (FSD) systems to enable
real-time navigation with minimal human intervention. Key components include neural networks
trained on camera feeds, ultrasonic sensors, and radar.<\/p>\n\n\n\n
\n- Examples:
o Tesla\u2019s Smart Summon feature autonomously drives cars to owners in parking
lots.
o AI Day presentations demonstrate the vision-first approach, avoiding LiDAR and
focusing on camera-based perception.
o Dojo Supercomputer trains its neural networks using millions of driving videos.<\/li>\n\n\n\n - Videos and research on Tesla\u2019s capabilities highlight advancements in lane detection,
object prediction, and traffic analysis.
Exercise 2: Deep Fake Videos
Observations:<\/li>\n<\/ul>\n\n\n\n\n- Best Video: The Barack Obama deep fake video created by Jordan Peele. It convincingly
showcased how AI can manipulate speech and facial expressions.<\/li>\n\n\n\n - How They Did It: This video was created using GANs (Generative Adversarial
Networks) and facial mapping technology. Specific tools like DeepFaceLab or customtrained models were used to replicate Obama\u2019s speech patterns and expressions.<\/li>\n\n\n\n - Time and Hardware: Likely required weeks of training on high-end GPUs, given its
quality and precision.
3<\/li>\n\n\n\n - Realism: While highly realistic, close inspection reveals slight lip-sync inconsistencies
and unnatural transitions in facial movements.
Exercise 3: Wacky Object Variations
Category: Potted Plants<\/li>\n<\/ol>\n\n\n\n\n- Found images include:
o A plant shaped like a cat.
o A pot designed as a robot holding a small cactus.
o A plant growing from a shoe instead of a traditional pot.
Exercise 4: Rise of Deep Learning
Key Points:<\/li>\n<\/ul>\n\n\n\n\n- Trends in GPU Price Performance:<\/li>\n<\/ol>\n\n\n\n
\n- GPUs have improved dramatically in computational power while becoming more
affordable, enabling broader adoption of deep learning.<\/li>\n<\/ul>\n\n\n\n\n- NVIDIA and AI Dominance:<\/li>\n<\/ol>\n\n\n\n
\n- NVIDIA\u2019s GPUs revolutionized AI training, surpassing CPU-based methods,
with partnerships accelerating AI integration in diverse industries.
Exercise 5: Object Detection<\/li>\n\n\n\n - Tested images include the unusual plants from Exercise 3.<\/li>\n\n\n\n
- Observations:
o Accurate detection of plant-like shapes.
o Unique pots were classified as unrelated objects like furniture.
4
Exercise 6: Image Classification
Observations:<\/li>\n<\/ul>\n\n\n\n\n- Single Object: A clean headshot was identified with high confidence and minimal
ambiguity.<\/li>\n\n\n\n - Cluttered Scene: A coffee cup amidst a desk setup confused the model, leading to
multiple low-confidence classifications.
Exercise 7: One-Stage vs Two-Stage Models
Models:<\/li>\n<\/ol>\n\n\n\n