Analyzing Monotonic Linear Interpolation in Neural Network Loss Landscapes
arXiv:2104.11044
Abstract
Linear interpolation between initial neural network parameters and converged parameters after training with stochastic gradient descent (SGD) typically leads to a monotonic decrease in the training objective. This Monotonic Linear Interpolation (MLI) property, first observed by Goodfellow et al. (2014) persists in spite of the non-convex objectives and highly non-linear training dynamics of neural networks. Extending this work, we evaluate several hypotheses for this property that, to our knowledge, have not yet been explored. Using tools from differential geometry, we draw connections between the interpolated paths in function space and the monotonicity of the network - providing sufficient conditions for the MLI property under mean squared error. While the MLI property holds under various settings (e.g. network architectures and learning problems), we show in practice that networks violating the MLI property can be produced systematically, by encouraging the weights to move far from initialization. The MLI property raises important questions about the loss landscape geometry of neural networks and highlights the need to further study their global properties.
arXiv comment15 pages in main paper, 4 pages of references, 24 pages in appendix. 29 figures in total
References in corpus (25)
- Adam: A Method for Stochastic Optimization
- Very Deep Convolutional Networks for Large-Scale Image Recognition
- Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
- Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour
- Neural Tangent Kernel: Convergence and Generalization in Neural Networks
- The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks
- Exponential expressivity in deep neural networks through transient chaos
- No More Pesky Learning Rates
- A mathematical theory of semantic development in deep neural networks
- Qualitatively characterizing neural network optimization problems
- Essentially No Barriers in Neural Network Energy Landscape
- Optimization for deep learning: theory and algorithms
- Fixup Initialization: Residual Learning Without Normalization
- Gradient Descent Happens in a Tiny Subspace
- Measuring the Intrinsic Dimension of Objective Landscapes
- The large learning rate phase of deep learning: the catapult mechanism
- Understanding Short-Horizon Bias in Stochastic Meta-Optimization
- The Break-Even Point on Optimization Trajectories of Deep Neural Networks
- Deep learning versus kernel learning: an empirical study of loss landscape geometry and the time evolution of the Neural Tangent Kernel
- Emergent properties of the local geometry of neural loss landscapes
- When Does Preconditioning Help or Hurt Generalization?
- Revisiting "Qualitatively Characterizing Neural Network Optimization Problems"
- Disentangling Adaptive Gradient Methods from Learning Rates
- On Connected Sublevel Sets in Deep Learning
- Landscape Connectivity and Dropout Stability of SGD Solutions for Over-parameterized Neural Networks