Structured predictive modeling for Supply Chain - Lecture 5.1


00:01 Introduction
02:44 Survey of predictive needs
05:57 Models vs. Modeling
12:26 The story so far
15:50 A bit of theory and a bit of practice
17:41 Differentiable Programming, SGD 1/6
24:56 Differentiable Programming, autodiff 2/6
31:07 Differentiable Programming, functions 3/6
35:35 Differentiable Programming, meta-parameters 4/6
37:59 Differentiable Programming, parameters 5/6
40:55 Differentiable Programming, quirks 6/6
43:41 Walkthrough, retail demand forecasting
45:49 Walkthrough, parameter fitting 1/6
53:14 Walkthrough, parameter sharing 2/6
01:04:16 Walkthrough, loss masking 3/6
01:09:34 Walkthrough, covariable integration 4/6
01:14:09 Walkthrough, sparse decomposition 5/6
01:21:17 Walkthrough, free-scaling 6/6
01:25:14 Whiteboxing
01:33:22 Back to experimental optimization
01:39:53 Conclusion
01:44:40 5.1 Structured predictive modeling… - Questions?

Description

Differentiable Programming (DP) is a generative paradigm used to engineer a broad class of statistical models, which happen to be excellently suited for predictive supply chain challenges. DP supersedes almost all the “classic” forecasting literature based on parametric models. DP is also superior to “classic” machine learning algorithms - up to the late 2010s - in virtually every dimension that matters for a practical usage for supply chain purposes, including ease of adoption by practitioners.