Bridging Gaps: Classic Approaches Beyond Deep Learning Constraints
summary
Practical cases show when classic methods beat heavy models: document classification uses contour detection plus pixel/histogram comparison instead of TensorFlow.js and CNNs, avoiding months of work; warehouse picking uses a simple moving-weighted formula (last week/15 days seasonality) achieving 97% reduction in forklift usage; unloading-time prediction relies on customer type, item count/weight/volume and GPS heuristics rather than deep nets; multi-object tracking mixes YOLOv8 with frame-difference and pixel tracking to reach 20× real-time throughput and cut GPU needs.
AI-written summary · about
chapters (12)
- 0:00AI limitations in production
- 1:31Speaker background
- 2:57When AI is overhyped and unnecessary
- 4:22Resource constraints in AI
- 6:09Need for AI?
- 7:14Document classification
- 10:06Warehouse sales forecasting
- 12:35Warehouse forecasting solution
- 15:44Predicting unloading time
- 18:13Object tracking optimization
- 20:26GPU optimization for tracking
- 22:34Key takeaways
related talks
46mML Use Cases at Think Big Analytics
33mOvercoming the Barriers to Deep Learning in Production
58mDeep Learning in Real Life
37mRecent Solutions Using Deep Learning Models/ The How and Why
03mInterview
41mDeep Learning to Solve Challenging Problems
01h 01mA practical guide to deep learning
41mPractical Aspects of Applying Modern Computer Vision Technologies