Search

Contrastive loss has been increasingly used in learning representations from multiple modalities. In the limit, the nature of the contrastive loss encourages modalities to exactly match each other in the latent space. Yet it remains an open question how the modality alignment affects the downstream task performance. In this paper, based on an information-theoretic argument, we first prove that exact modality alignment is sub-optimal in general for downstream prediction tasks. Hence we advocate that the key of better performance lies in meaningful latent modality structures instead of perfect modality alignment. To this end, we propose three general approaches to construct latent modality structures. Specifically, we design 1) a deep feature separation loss for intra-modality regularization; 2) a Brownian-bridge loss for inter-modality regularization; and 3) a geometric consistency loss for both intra- and inter-modality regularization. Extensive experiments are conducted on two popular multi-modal representation learning frameworks: the CLIP-based two-tower model and the ALBEF-based fusion model. We test our model on a variety of tasks including zero/few-shot image classification, image-text retrieval, visual question answering, visual reasoning, and visual entailment. Our method achieves consistent improvements over existing methods, demonstrating the effectiveness and generalizability of our proposed approach on latent modality structure regularization., Comment: 14 pages, 8 figure, CVPR 2023 accepted

Aligning signals from different modalities is an important step in vision-language representation learning as it affects the performance of later stages such as cross-modality fusion. Since image and text typically reside in different regions of the feature space, directly aligning them at instance level is challenging especially when features are still evolving during training. In this paper, we propose to align at a higher and more stable level using cluster representation. Specifically, we treat image and text as two "views" of the same entity, and encode them into a joint vision-language coding space spanned by a dictionary of cluster centers (codebook). We contrast positive and negative samples via their cluster assignments while simultaneously optimizing the cluster centers. To further smooth out the learning process, we adopt a teacher-student distillation paradigm, where the momentum teacher of one view guides the student learning of the other. We evaluated our approach on common vision language benchmarks and obtain new SoTA on zero-shot cross modality retrieval while being competitive on various other transfer tasks., Comment: Accepted by CVPR 2022

The mixture of Expert (MoE) parallelism is a recent advancement that scales up the model size with constant computational cost. MoE selects different sets of parameters (i.e., experts) for each incoming token, resulting in a sparsely-activated model. Despite several successful applications of MoE, its training efficiency degrades significantly as the number of experts increases. The routing stage in MoE relies on the efficiency of the All2All communication collective, which suffers from network congestion and has poor scalability. To mitigate these issues, we introduce SMILE, which exploits heterogeneous network bandwidth and splits a single-step routing into bi-level routing. Our experimental results show that the proposed method obtains a 2.5x speedup over Switch Transformer in terms of pretraining throughput on the Colossal Clean Crawled Corpus without losing any convergence speed.

Can we combine heterogenous graph structure with text to learn high-quality semantic and behavioural representations? Graph neural networks (GNN)s encode numerical node attributes and graph structure to achieve impressive performance in a variety of supervised learning tasks. Current GNN approaches are challenged by textual features, which typically need to be encoded to a numerical vector before provided to the GNN that may incur some information loss. In this paper, we put forth an efficient and effective framework termed language model GNN (LM-GNN) to jointly train large-scale language models and graph neural networks. The effectiveness in our framework is achieved by applying stage-wise fine-tuning of the BERT model first with heterogenous graph information and then with a GNN model. Several system and design optimizations are proposed to enable scalable and efficient training. LM-GNN accommodates node and edge classification as well as link prediction tasks. We evaluate the LM-GNN framework in different datasets performance and showcase the effectiveness of the proposed approach. LM-GNN provides competitive results in an Amazon query-purchase-product application.

Existing general purpose frameworks for gigantic model training, i.e., dense models with billions of parameters, cannot scale efficiently on cloud environment with various networking conditions due to large communication overheads. In this paper, we propose MiCS, which Minimizes the Communication Scale to bring down communication overhead. Specifically, by decreasing the number of participants in a communication collective, MiCS can utilize heterogeneous network bandwidth, reduce network traffic over slower links, reduce the latency of communications for maintaining high network bandwidth utilization, and amortize expensive global gradient synchronization overhead. Our evaluation on AWS shows that the system throughput of MiCS is up to 2.89× that of the state-of-the-art large model training systems. MiCS achieves near-linear scaling efficiency, which is up to 1.27× that of DeepSpeed. MiCS allows us to train a proprietary model with 100 billion parameters on 512 GPUs with 99.4% weak-scaling efficiency, and it is able to saturate over 54.5% theoretical computation power of each GPU on a public cloud with less GPU memory and more restricted networks than DGX-A100 clusters.

Tiering is an essential technique for building large-scale information retrieval systems. While the selection of documents for high priority tiers critically impacts the efficiency of tiering, past work focuses on optimizing it with respect to a static set of queries in the history, and generalizes poorly to the future traffic. Instead, we formulate the optimal tiering as a stochastic optimization problem, and follow the methodology of regularized empirical risk minimization to maximize the \emph{generalization performance} of the system. We also show that the optimization problem can be cast as a stochastic submodular optimization problem with a submodular knapsack constraint, and we develop efficient optimization algorithms by leveraging this connection.

The cost of accessing main memory is increasing. Machine designers have tried to mitigate the consequences of the processor and memory technology trends underlying this increasing gap with a variety of techniques to reduce or tolerate memory latency. These techniques, unfortunately, are only occasionally successful for pointer-manipulating programs. Recent research has demonstrated the value of a complementary approach, in which pointer-based data structures are reorganized to improve cache locality.This paper studies a technique for using a generational garbage collector to reorganize data structures to produce a cache-conscious data layout, in which objects with high temporal affinity are placed next to each other, so that they are likely to reside in the same cache block. The paper explains how to collect, with low overhead, real-time profiling information about data access patterns in object-oriented languages, and describes a new copying algorithm that utilizes this information to produce a cache-conscious object layout.Preliminary results show that this technique reduces cache miss rates by 21--42%, and improves program performance by 14--37% over Cheney's algorithm. We also compare our layouts against those produced by the Wilson-Lam-Moher algorithm, which attempts to improve program locality at the page level. Our cache-conscious object layouts reduces cache miss rates by 20--41% and improves program performance by 18--31% over their algorithm, indicating that improving locality at the page level is not necessarily beneficial at the cache level.

Shared memory in a parallel computer provides programmers with the valuable abstraction of a shared address space--through which any part of a computation can access any datum. Although uniform access simplifies programming, it also hides communication, which can lead to inefficient programs. The check-in, check-out (CICO) performance model for cache-coherent, shared-memory parallel computers helps a programmer identify the communication underlying memory references and account for its cost. CICO consists of annotations that a programmer can use to elucidate communication and a model that attributes costs to these annotations. The annotations can also serve as directives to a memory system to improve program performance. Inserting CICO annotations requires reasoning about the dynamic cache behavior of a program, which is not always easy. This paper describes Cachier, a tool that automatically inserts CICO annotations into shared-memory programs. A novel feature of this tool is its use of both dynamic information, obtained from a program execution trace, as well as static information, obtained from program analysis. We measured several benchmarks annotated by Cachier by running them on a simulation of the Dir1SW cache coherence protocol, which supports these directives. The results show that programs annotated by Cachier perform significantly better than both programs without CICO annotations and programs that were annotated by hand.