Practical AI Towards stability and robustness
Jul 20, 2021
Roey Mechrez, CTO at BeyondMinds, shares his expertise on building robust AI systems and highlights the gap between academic research and practical deployment. He discusses why 90% of AI projects fail to create value, emphasizing the importance of stability and specificity in model performance. Roey dives into crucial techniques like data filtration, out-of-distribution detection, and the need for human oversight to enhance reliability. His insights reveal the critical role of tailored solutions and continuous model retraining to navigate the complexities of real-world AI implementation.
AI Snips
Chapters
Transcript
Episode notes
Technology-Driven Approach
- Many organizations struggle to implement AI effectively due to a lack of technical expertise and a technology-driven mindset.
- Building internal AI teams is challenging, and companies often face a buy-versus-build dilemma similar to the early CRM era.
Scaling Challenges
- A company struggled to scale its sales prediction model across 30 different product lines.
- Scaling required significant team expansion due to the need for constant retraining and hyperparameter tuning for each model.
The Giraffe Problem
- The "giraffe problem" illustrates how AI models can fail silently when encountering unfamiliar data.
- A model trained only on cats and dogs will misclassify a giraffe, highlighting the importance of understanding data distribution.