Transmission Neural Networks: From Virus Spread Models to Neural Networks
summary
A discrete-time virusspread model is rewritten by taking states S = −log(1−P) and becomes a linear summation through linkwise nonlinearities, leading to a new architecture called Transmission Neural Networks where activation levels live on edges and are trainable. Three closed-form activations emerge including a tunable log-sigmoid and variants, TransNNs with one hidden layer plus fixed bias are proved universal approximators, a spectral threshold (spectral radius of the Hadamard-modified matrix < 1) predicts die-out, and a continuous-time SIS model follows from a small-time scaling.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:07Network spreading basics
- 5:40Model transformation
- 8:12Transmission network model
- 12:15Tunable activation functions
- 14:28Neural network representation
- 17:56Epidemic threshold analysis
- 21:32Continuous time models
- 22:49Universal approximation
- 26:26Learning and training
- 29:14Future work and conclusion
- 33:05Q&A session
related talks
49mAn Algorithmic Theory of Brain Networks
15mSMT-Based Modeling and Verification of Spiking Neural Networks: A Case Study
55mNetworks of Spiking Neurons Learn to Learn and Remember
43mProbabilistic Models for Pathogen Evolution Within and Between Hosts
01h 59mSpiking Neural Networks
37mNonlocal Physics-Informed Neural Networks - A Unified Framework for Nonlocal Models
51mThe Brain as an Efficient, Robust Adaptive Learner
47mDynamic Neural Network Structures Through Stochastic Rewiring