Efficient second-order methods for model compression

Second-order information, in the form of Hessian- or Inverse-Hessian-vector products, is a fundamental tool for solving optimization problems. Recently, there has been a tremendous amount of work on utilizing this information for the current compute and memory-intensive deep neural networks, usually via coarse-grained approximations (such as diagonal, blockwise, or Kronecker-factorization). However, not much is known about the quality of these approximations. Our work addresses this question, and in particular, we propose a method called ‘WoodFisher’ that leverages the structure of the empirical Fisher information matrix, along with the Woodbury matrix identity, to compute a faithful and efficient estimate of the inverse Hessian. Our main application is to the task of compressing neural networks, where we build on the classical Optimal Brain Damage/Surgeon framework (LeCun et al., 1990; Hassibi and Stork, 1993). We demonstrate that WoodFisher significantly outperforms magnitude pruning (isotropic Hessian), as well as methods that maintain other diagonal estimates. Further, even when gradual pruning is considered, our method results in a gain in test accuracy over the state-of-the-art approaches, for standard image classification datasets such as CIFAR-10, ImageNet. We also propose a variant called ‘WoodTaylor’, which takes into account the first-order gradient term, and can lead to additional improvements. An important advantage of our methods is that they allow us to automatically set the layer-wise pruning thresholds, avoiding the need for any manual tuning or sensitivity analysis.

Theory of representation learning in cortical neural networks

Carlos Stein Naves de Brito

Our brain continuously self-organizes to construct and maintain an internal representation of the world based on the information arriving through sensory stimuli. Remarkably, cortical areas related to different sensory modalities appear to share the same functional unit, the neuron, and develop through the same learning mechanism, synaptic plasticity. It motivates the conjecture of a unifying theory to explain cortical representational learning across sensory modalities. In this thesis we present theories and computational models of learning and optimization in neural networks, postulating functional properties of synaptic plasticity that support the apparent universal learning capacity of cortical networks. In the past decades, a variety of theories and models have been proposed to describe receptive field formation in sensory areas. They include normative models such as sparse coding, and bottom-up models such as spike-timing dependent plasticity. We bring together candidate explanations by demonstrating that in fact a single principle is sufficient to explain receptive field development. First, we show that many representative models of sensory development are in fact implementing variations of a common principle: nonlinear Hebbian learning. Second, we reveal that nonlinear Hebbian learning is sufficient for receptive field formation through sensory inputs. A surprising result is that our findings are independent of specific details, and allow for robust predictions of the learned receptive fields. Thus nonlinear Hebbian learning and natural statistics can account for many aspects of receptive field formation across models and sensory modalities. The Hebbian learning theory substantiates that synaptic plasticity can be interpreted as an optimization procedure, implementing stochastic gradient descent. In stochastic gradient descent inputs arrive sequentially, as in sensory streams. However, individual data samples have very little information about the correct learning signal, and it becomes a fundamental problem to know how many samples are required for reliable synaptic changes. Through estimation theory, we develop a novel adaptive learning rate model, that adapts the magnitude of synaptic changes based on the statistics of the learning signal, enabling an optimal use of data samples. Our model has a simple implementation and demonstrates improved learning speed, making this a promising candidate for large artificial neural network applications. The model also makes predictions on how cortical plasticity may modulate synaptic plasticity for optimal learning. The optimal sampling size for reliable learning allows us to estimate optimal learning times for a given model. We apply this theory to derive analytical bounds on times for the optimization of synaptic connections. First, we show this optimization problem to have exponentially many saddle-nodes, which lead to small gradients and slow learning. Second, we show that the number of input synapses to a neuron modulates the magnitude of the initial gradient, determining the duration of learning. Our final result reveals that the learning duration increases supra-linearly with the number of synapses, suggesting an effective limit on synaptic connections and receptive field sizes in developing neural networks.

EPFL2016

PowerSGD: Practical Low-Rank Gradient Compression for Distributed Optimization

Martin Jaggi, Thijs Vogels

We study lossy gradient compression methods to alleviate the communication bottleneck in data-parallel distributed optimization. Despite the significant attention received, current compression schemes either do not scale well, or fail to achieve the target test accuracy. We propose a new low-rank gradient compressor based on power iteration that can i) compress gradients rapidly, ii) efficiently aggregate the compressed gradients using all-reduce, and iii) achieve test performance on par with SGD. The proposed algorithm is the only method evaluated that achieves consistent wall-clock speedups when benchmarked against regular SGD using highly optimized off-the-shelf tools for distributed communication. We demonstrate reduced training times for convolutional networks as well as LSTMs on common datasets. Our code is available at https://github.com/epfml/powersgd.

NEURAL INFORMATION PROCESSING SYSTEMS (NIPS)2019

Masked Training of Neural Networks with Partial Gradients

Martin Jaggi, Amirkeivan Mohtashami, Sebastian Urban Stich

State-of-the-art training algorithms for deep learning models are based on stochastic gradient descent (SGD). Recently, many variations have been explored: perturbing parameters for better accuracy (such as in Extra-gradient), limiting SGD updates to a subset of parameters for increased efficiency (such as meProp) or a combination of both (such as Dropout). However, the convergence of these methods is often not studied in theory. We propose a unified theoretical framework to study such SGD variants-encompassing the aforementioned algorithms and additionally a broad variety of methods used for communication efficient training or model compression. Our insights can be used as a guide to improve the efficiency of such methods and facilitate generalization to new applications. As an example, we tackle the task of jointly training networks, a version of which (limited to sub-networks) is used to create Slimmable Networks. By training a low-rank Transformer jointly with a standard one we obtain superior performance than when it is trained separately.

JMLR-JOURNAL MACHINE LEARNING RESEARCH2022

Theory of representation learning in cortical neural networks

Carlos Stein Naves de Brito

EPFL2016