Intelligence plays dice: Stochasticity is essential for machine learning
arXiv:2008.07496
Abstract
Many fields view stochasticity as a way to gain computational efficiency, while often having to trade off accuracy. In this perspective article, we argue that stochasticity plays a fundamentally different role in machine learning (ML) and is likely a critical ingredient of intelligent systems. As we review the ML literature, we notice that stochasticity features in many ML methods, affording them robustness, generalizability, and calibration. We also note that randomness seems to be prominent in biological intelligence, from the spiking patterns of individual neurons to the complex behavior of animals. We conclude with a discussion of how we believe stochasticity might shape the future of ML.
References in corpus (17)
- Adam: A Method for Stochastic Optimization
- mixup: Beyond Empirical Risk Minimization
- Towards A Rigorous Science of Interpretable Machine Learning
- The Effectiveness of Data Augmentation in Image Classification using Deep Learning
- Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
- On Calibration of Modern Neural Networks
- The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks
- Understanding deep learning requires rethinking generalization
- SmoothGrad: removing noise by adding noise
- Group Equivariant Convolutional Networks
- SNAS: Stochastic Neural Architecture Search
- Random Search and Reproducibility for Neural Architecture Search
- Stochastic Activation Pruning for Robust Adversarial Defense
- Structured Bayesian Pruning via Log-Normal Multiplicative Noise
- Confidence Calibration for Convolutional Neural Networks Using Structured Dropout
- An Alternative View: When Does SGD Escape Local Minima?
- Adversarial Robustness Guarantees for Random Deep Neural Networks