Constrained Monotonic Neural Networks
summary
Presents a concrete method to enforce monotonicity in deep nets by combining convex and concave activations inside a simple monotonic layer, producing certified monotone outputs without heavyweight lattices or expensive verifiers. Implements as ~100 lines of Keras code and a lightweight layer that handles time via a separate highway and masks cyclic/linear components, tests mainly on synthetic benchmarks and standard datasets, and claims similar or better accuracy while using an order of magnitude fewer parameters.
AI-written summary · about
chapters (15)
- 0:00Constrained monotonic neural networks
- 1:28Monotonicity in deep learning
- 3:16NLP to business data
- 8:29Language vs business data
- 10:12Time as a feature
- 13:41Time-dependent probabilities
- 16:06Ethics of non-monotonic models
- 18:54Building monotonic networks
- 21:01Limitations of existing solutions
- 22:24Simple monotonic networks
- 23:27Industrial implementation
- 24:49Performance vs lattice networks
- 26:32Testing with synthetic data
- 29:01Start with synthetic examples
- 30:04Q&A
related talks
37mMonadic Deep Learning
42mTowards Verification of General Neural Networks: A Dual Approach
45mExplaining Landscape Connectivity of Low-cost Solutions for Multilayer Nets
50mLearning and Generalization in Over-parametrized Neural Networks, Going Beyond Kernels
31mAbstract Neural Networks
47mInterpreting Deep Neural Networks (DNNs)
11mProxy Convexity: A Unified Framework for the Analysis of Neural Networks Trained by Gradient Descent
33mLearn Neural Networks With Go - Not Math!