Mixture-of-Experts (MoE) models have recently gained steam in achieving the state-of-the-art performance in a wide range of tasks in computer vision and natural language processing. They effectively expand the model capacity while incurring a minimal increase in computation cost during training. However, deploying such models for inference is difficult due to their large model size and complex communication pattern. In this work, we provide a characterization of two MoE workloads, namely Language Modeling (LM) and Machine Translation (MT) and identify their sources of inefficiencies at deployment. We propose three optimization techniques to mitigate sources of inefficiencies, namely (1) Dynamic gating, (2) Expert Buffering, and (3) Expert load balancing. We show that dynamic gating improves execution time by 1.25-4$\times$ for LM, 2-5$\times$ for MT Encoder and 1.09-1.5$\times$ for MT Decoder. It also reduces memory usage by up to 1.36$\times$ for LM and up to 1.1$\times$ for MT. We further propose Expert Buffering, a new caching mechanism that only keeps hot, active experts in GPU memory while buffering the rest in CPU memory. This reduces static memory allocation by 1.47$\times$. We finally propose a load balancing methodology that provides additional robustness to the workload. The code will be open-sourced upon acceptance.
Visual concept discovery has long been deemed important to improve interpretability of neural networks, because a bank of semantically meaningful concepts would provide us with a starting point for building machine learning models that exhibit intelligible reasoning process. Previous methods have disadvantages: either they rely on labelled support sets that incorporate human biases for objects that are "useful," or they fail to identify multiple concepts that occur within a single image. We reframe the concept discovery task as an unsupervised semantic segmentation problem, and present SegDiscover, a novel framework that discovers semantically meaningful visual concepts from imagery datasets with complex scenes without supervision. Our method contains three important pieces: generating concept primitives from raw images, discovering concepts by clustering in the latent space of a self-supervised pretrained encoder, and concept refinement via neural network smoothing. Experimental results provide evidence that our method can discover multiple concepts within a single image and outperforms state-of-the-art unsupervised methods on complex datasets such as Cityscapes and COCO-Stuff. Our method can be further used as a neural network explanation tool by comparing results obtained by different encoders.
Interpretability in machine learning (ML) is crucial for high stakes decisions and troubleshooting. In this work, we provide fundamental principles for interpretable ML, and dispel common misunderstandings that dilute the importance of this crucial topic. We also identify 10 technical challenge areas in interpretable machine learning and provide history and background on each problem. Some of these problems are classically important, and some are recent problems that have arisen in the last few years. These problems are: (1) Optimizing sparse logical models such as decision trees; (2) Optimization of scoring systems; (3) Placing constraints into generalized additive models to encourage sparsity and better interpretability; (4) Modern case-based reasoning, including neural networks and matching for causal inference; (5) Complete supervised disentanglement of neural networks; (6) Complete or even partial unsupervised disentanglement of neural networks; (7) Dimensionality reduction for data visualization; (8) Machine learning models that can incorporate physics and other generative or causal constraints; (9) Characterization of the "Rashomon set" of good models; and (10) Interpretable reinforcement learning. This survey is suitable as a starting point for statisticians and computer scientists interested in working in interpretable machine learning.
Dimension reduction (DR) techniques such as t-SNE, UMAP, and TriMAP have demonstrated impressive visualization performance on many real world datasets. One tension that has always faced these methods is the trade-off between preservation of global structure and preservation of local structure: these methods can either handle one or the other, but not both. In this work, our main goal is to understand what aspects of DR methods are important for preserving both local and global structure: it is difficult to design a better method without a true understanding of the choices we make in our algorithms and their empirical impact on the lower-dimensional embeddings they produce. Towards the goal of local structure preservation, we provide several useful design principles for DR loss functions based on our new understanding of the mechanisms behind successful DR methods. Towards the goal of global structure preservation, our analysis illuminates that the choice of which components to preserve is important. We leverage these insights to design a new algorithm for DR, called Pairwise Controlled Manifold Approximation Projection (PaCMAP), which preserves both local and global structure. Our work provides several unexpected insights into what design choices both to make and avoid when constructing DR algorithms.
Image segmentation with a volume constraint is an important prior for many real applications. In this work, we present a novel volume preserving image segmentation algorithm, which is based on the framework of entropic regularized optimal transport theory. The classical Total Variation (TV) regularizer and volume preserving are integrated into a regularized optimal transport model, and the volume and classification constraints can be regarded as two measures preserving constraints in the optimal transport problem. By studying the dual problem, we develop a simple and efficient dual algorithm for our model. Moreover, to be different from many variational based image segmentation algorithms, the proposed algorithm can be directly unrolled to a new Volume Preserving and TV regularized softmax (VPTV-softmax) layer for semantic segmentation in the popular Deep Convolution Neural Network (DCNN). The experiment results show that our proposed model is very competitive and can improve the performance of many semantic segmentation nets such as the popular U-net.
In this paper, we propose a normalized cut segmentation algorithm with spatial regularization priority and adaptive similarity matrix. We integrate the well-known expectation-maximum(EM) method in statistics and the regularization technique in partial differential equation (PDE) method into normalized cut (Ncut). The introduced EM technique makes our method can adaptively update the similarity matrix, which can help us to get a better classification criterion than the classical Ncut method. While the regularization priority can guarantee the proposed algorithm has a robust performance under noise. To unify the three totally different methods including EM, spatial regularization, and spectral graph clustering, we built a variational framework to combine them and get a general normalized cut segmentation algorithm. The well-defined theory of the proposed model is also given in the paper. Compared with some existing spectral clustering methods such as the traditional Ncut algorithm and the variational based Chan-Vese model, the numerical experiments show that our methods can achieve promising segmentation performance.
In this paper, the traditional model based variational method and learning based algorithms are naturally integrated to address mixed noise removal problem. To be different from single type noise (e.g. Gaussian) removal, it is a challenge problem to accurately discriminate noise types and levels for each pixel. We propose a variational method to iteratively estimate the noise parameters, and then the algorithm can automatically classify the noise according to the different statistical parameters. The proposed variational problem can be separated into regularization, synthesis, parameter estimation and noise classification four steps with the operator splitting scheme. Each step is related to an optimization subproblem. To enforce the regularization, the deep learning method is employed to learn the natural images priori. Compared with some model based regularizations, the CNN regularizer can significantly improve the quality of the restored images. Compared with some learning based methods, the synthesis step can produce better reconstructions by analyzing the recognized noise types and levels. In our method, the convolution neutral network (CNN) can be regarded as an operator which associated to a variational functional. From this viewpoint, the proposed method can be extended to many image reconstruction and inverse problems. Numerical experiments in the paper show that our method can achieve some state-of-the-art results for mixed noise removal.