Not a magic : What to expect from Machine Learning projects
summary
Organisations need a concrete plan: focus on the three pillars data model deployment, expect heavy work on pipelines, monitoring and versioning, and budget for read/write and GPU costs. Typical surprises: modeling often takes ~5% of project time, test splits of 20–30% reveal overfitting, and GDPR/ownership issues block access. Use existing APIs (fit/train in TensorFlow), prefer transfer learning over custom cost functions, run A/B tests with a clear metric, and treat reports as a can of worms to be maintained.
AI-written summary · about
chapters (27)
- 0:00Intro
- 3:47Machine Learning Project
- 5:35Data mining
- 6:42Data protection
- 7:27Data cleaning
- 8:07Data domain
- 8:45The nature of Data
- 9:37Data – sharing the progress
- 10:25Data reporting
- 13:05Learning data
- 15:51Learning algorithms
- 19:32Training: Loss and Cost funtion
- 20:40Training: Objective function and loss
- 21:32Training: summary
- 23:25Feature engineering
- 27:26Evaluate Metric: Supervised learning
- 29:12Evaluate model: high variance
- 31:00Evaluate: compare with what
- 32:29Evaluate model: same distirbution
- 34:11Deep learning
- 37:53Data Pipeline - reality
- 38:04Monitoring
- 39:08Reports
- 39:48Model deployment
- 41:37A/B testing
- 47:02Proof of Concept (PC)
- 47:27Modelling
related talks
27mKeys to Building Machine Learning Systems
15m5 Lessons Learned from Implementing 40+ Machine Learning Projects
30mChallenges of machine learning: The new ML lifecycle
32mAn Only One Step Ahead Guide for Machine Learning Projects
41mChallenges in incorporating ML models
27m10 Things That Can Go Wrong with ML Projects
32mWhy ML Projects Fail and How to Ensure Their Success
19mSustaining ML in the enterprise