“Practical Strategies for Successful Implementation and Deployment of AI-based Solutions,” a Presentation from Globus Medical

For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2024/09/practical-strategies-for-successful-implementation-and-deployment-of-ai-based-solutions-a-presentation-from-globus-medical/

Ritesh Agarwal, Computer Vision Lead at Globus Medical, presents the “Practical Strategies for Successful Implementation and Deployment of AI-based Solutions” tutorial at the May 2024 Embedded Vision Summit.

AI models that produce accurate results on test data are a necessary component of successful applications, but by themselves they are insufficient. In this talk, Agarwal delves into often-overlooked, but critical, elements required to create and sustain a robust application solution based on AI models.

Agarwal explores the need for deep understanding of model performance and highlights techniques for identification of specific hyperparameters that can be tuned to optimize real-world accuracy. He also examines the importance of pre- and post-processing and the advantages of using multiple models in concert to improve accuracy. Finally, at the organization level, he provides recommendations for improving product development via synergy between data engineers, application domain experts, machine learning engineers, data scientists, quality assurance engineers, DevOps specialists and software engineers.

1. Practical Strategies for
Successful Implementation
and Deployment of AI-Based
Solutions
Ritesh Agarwal, Computer Vision Lead
Globus Medical

2. Defining Business
Requirements
•End-user experience
•Resource and time constraints
•Competitive advantage
•Risk management
•Cost-effectiveness
General Practice
Where to start?
Defining Technical
Requirements
•Problem type
•Data size and complexity
•Input data
•Resource constraints
•Domain knowledge
•Experimentation
•Scalability
© Globus Medical 2

3. Application Requirements
Traffic management
© Globus Medical
Understanding Business
Requirements
User needs
Resource and time constraints
Traffic level
Vehicles kind –car, bus, bikes
Integration with drone cameras
Integration with existing
infrastructure
Data security and privacy
Understanding Technical
Requirements
Image processing –small object
detection
Drone camera specifications
Real time video streams
Data transmission and connectivity
Environmental adaptability-
lighting, weather, and terrain
Accuracy and performance
Data security and privacy
Scalability and flexibility
3

4. Designing the Algorithm
Create a roadmap
© Globus Medical
Conceptualizing the design and identifying the problems
Visual reference
4

5. •Pre-processing:
•Motion flow determination
•Image clarity
•Weather considerations
•Weather detection
•Adaptation strategies –
reduced visibility
•Object detection
•Selecting the right
architecture
Designing the Algorithm
Create a roadmap
© Globus Medical
•Post-processing:
•Refinement of detected
objects
•Traffic flow analysis
•Density estimation
•Data refinement –
statistical analysis
•Optimization
•Deployment strategies –
failure mechanisms
•Feedback loop
•Final assessment
After the problem is identified and the requirements are gathered, the design phase begins
5

6. Model capabilities
•Small object detection
•Computation speed
•Flexibility
YOLO: Fast processing, moderate accuracy.
SSD: Balance between speed and accuracy.
EfficientDet: High accuracy with efficient resource utilization.
Mask-RCNN: High accuracy, Compute Intensive
After careful consideration we ended up using RetinaNet–Focal loss, anchors and feature pyramid
networks.
Identifying Architecture
Based on model characteristics
© Globus Medical 6

7. •Open-source datasets
•Simulation environment/synthetic
•Collaborative partnerships
•Crowdsourcing
Data Collection Strategies
Application specific
© Globus Medical
Several strategies can be employed to collect data, some of which are:
Vis-Drone dataset
7

8. •Identifying evaluation metrics
•Understanding hyperparameters
•Pitfall identification –fitting issues
•Model comparison
•Visualization techniques
Training and Evaluating Models
Understanding model hyperparameters and their impact on performance
© Globus Medical
Setting up a baseline model for comparison is the next stage, whether training from
scratch or utilizing transfer learning
8

9. © Globus Medical
Missing carsExtra cars
9

10. •Integration of multiple models andtheir outputs
•Selection of optimal parameters
•Adaptivecontrol strategies –weather/lighting
conditions
•Feedbackmechanism –to better the algorithm
Creating the Algorithm
Enhancing object detection with advanced approaches
© Globus Medical
Creating the final algorithm is the culmination of the entire process of developing a
solution
Ensemble model
10

11. Deployment strategy considerations
•Data security and compliance
•Latency and performance
•Infrastructure costs
•Scalability and flexibility
•IT support and maintenance
Deployment
Strategies for successful implementation
© Globus Medical
11

12. A synergy between
•Data engineers
•Domain experts
•Machine learning engineers
•Deep learning engineers
•Data scientists
•DevOps engineers
•Software engineers
•Quality assurance engineers
The AI Loop
Navigating the continuous cycle of development, deployment, and iteration
© Globus Medical
12

13. Final Product
© Globus Medical 13

14. Consider the following when developing an AI application
•Holistic approach
•Data-driven decision making
•Iterative development
•Collaborative effort
•Real-world considerations
•Bias mitigation
•Ethical considerations
Conclusion
© Globus Medical 14

15. Resource
Algorithms Design Techniques –GeeksforGeeks
A Step By Step Guide To AI Model Development -DataScienceCentral.com
Open Datasets For AI/ML | AI Training Datasets –Shaip
On hyperparameter optimization of machine learning algorithms: Theory and practice –
ScienceDirect
Drones with Artificial Intelligence will soon become a powerful tool —a new perspective | by Ritesh
Agarwal | Medium
© Globus Medical
15

16. © Globus Medical
Happy AI!
16