Radial Networks: Dynamic Layer Routing for High-Performance Large Language Models

Large language models (LLMs) often struggle with strict memory, latency, and power demands. To meet these demands, various forms of dynamic sparsity have been proposed that reduce compute on an input-by-input basis. These methods improve over static methods by exploiting the variance across individual inputs, which has steadily grown with the exponential increase in training data. Yet, the increasing depth within modern models, currently with hundreds of layers, has opened opportunities for dynamic layer sparsity, which skips the computation for entire layers. In this work, we explore the practicality of layer sparsity by profiling residual connections and establish the relationship between model depth and layer sparsity. For example, the residual blocks in the OPT-66B model have a median contribution of 5% to its output. We then take advantage of this dynamic sparsity and propose Radial Networks, which perform token-level routing between layers guided by a trained router module. These networks can be used in a post-training distillation from sequential networks or trained from scratch to co-learn the router and layer weights. They enable scaling to larger model sizes by decoupling the number of layers from the dynamic depth of the network, and their design allows for layer reuse. By varying the compute token by token, they reduce the overall resources needed for generating entire sequences. Overall, this leads to larger capacity networks with significantly lower compute and serving costs for large language models.

Exploring the Limits of Semantic Image Compression at Micro-bits per Pixel

Traditional methods, such as JPEG, perform image compression by operating on structural information, such as pixel values or frequency content. These methods are effective to bitrates around one bit per pixel (bpp) and higher at standard image sizes. In contrast, text-based semantic compression directly stores concepts and their relationships using natural language, which has evolved with humans to efficiently represent these salient concepts. These methods can operate at extremely low bitrates by disregarding structural information like location, size, and orientation. In this work, we use GPT-4V and DALL-E3 from OpenAI to explore the quality-compression frontier for image compression and identify the limitations of current technology. We push semantic compression as low as 100 $\mu$bpp (up to $10,000\times$ smaller than JPEG) by introducing an iterative reflection process to improve the decoded image. We further hypothesize this 100 $\mu$bpp level represents a soft limit on semantic compression at standard image resolutions.

The Right Losses for the Right Gains: Improving the Semantic Consistency of Deep Text-to-Image Generation with Distribution-Sensitive Losses

One of the major challenges in training deep neural networks for text-to-image generation is the significant linguistic discrepancy between ground-truth captions of each image in most popular datasets. The large difference in the choice of words in such captions results in synthesizing images that are semantically dissimilar to each other and to their ground-truth counterparts. Moreover, existing models either fail to generate the fine-grained details of the image or require a huge number of parameters that renders them inefficient for text-to-image synthesis. To fill this gap in the literature, we propose using the contrastive learning approach with a novel combination of two loss functions: fake-to-fake loss to increase the semantic consistency between generated images of the same caption, and fake-to-real loss to reduce the gap between the distributions of real images and fake ones. We test this approach on two baseline models: SSAGAN and AttnGAN (with style blocks to enhance the fine-grained details of the images.) Results show that our approach improves the qualitative results on AttnGAN with style blocks on the CUB dataset. Additionally, on the challenging COCO dataset, our approach achieves competitive results against the state-of-the-art Lafite model, outperforms the FID score of SSAGAN model by 44.

ArtELingo: A Million Emotion Annotations of WikiArt with Emphasis on Diversity over Language and Culture

This paper introduces ArtELingo, a new benchmark and dataset, designed to encourage work on diversity across languages and cultures. Following ArtEmis, a collection of 80k artworks from WikiArt with 0.45M emotion labels and English-only captions, ArtELingo adds another 0.79M annotations in Arabic and Chinese, plus 4.8K in Spanish to evaluate "cultural-transfer" performance. More than 51K artworks have 5 annotations or more in 3 languages. This diversity makes it possible to study similarities and differences across languages and cultures. Further, we investigate captioning tasks, and find diversity improves the performance of baseline models. ArtELingo is publicly available at https://www.artelingo.org/ with standard splits and baseline models. We hope our work will help ease future research on multilinguality and culturally-aware AI.

Logic Shrinkage: Learned FPGA Netlist Sparsity for Efficient Neural Network Inference

FPGA-specific DNN architectures using the native LUTs as independently trainable inference operators have been shown to achieve favorable area-accuracy and energy-accuracy tradeoffs. The first work in this area, LUTNet, exhibited state-of-the-art performance for standard DNN benchmarks. In this paper, we propose the learned optimization of such LUT-based topologies, resulting in higher-efficiency designs than via the direct use of off-the-shelf, hand-designed networks. Existing implementations of this class of architecture require the manual specification of the number of inputs per LUT, K. Choosing appropriate K a priori is challenging, and doing so at even high granularity, e.g. per layer, is a time-consuming and error-prone process that leaves FPGAs' spatial flexibility underexploited. Furthermore, prior works see LUT inputs connected randomly, which does not guarantee a good choice of network topology. To address these issues, we propose logic shrinkage, a fine-grained netlist pruning methodology enabling K to be automatically learned for every LUT in a neural network targeted for FPGA inference. By removing LUT inputs determined to be of low importance, our method increases the efficiency of the resultant accelerators. Our GPU-friendly solution to LUT input removal is capable of processing large topologies during their training with negligible slowdown. With logic shrinkage, we better the area and energy efficiency of the best-performing LUTNet implementation of the CNV network classifying CIFAR-10 by 1.54x and 1.31x, respectively, while matching its accuracy. This implementation also reaches 2.71x the area efficiency of an equally accurate, heavily pruned BNN. On ImageNet with the Bi-Real Net architecture, employment of logic shrinkage results in a post-synthesis area reduction of 2.67x vs LUTNet, allowing for implementation that was previously impossible on today's largest FPGAs.

Journey Towards Tiny Perceptual Super-Resolution

Recent works in single-image perceptual super-resolution (SR) have demonstrated unprecedented performance in generating realistic textures by means of deep convolutional networks. However, these convolutional models are excessively large and expensive, hindering their effective deployment to end devices. In this work, we propose a neural architecture search (NAS) approach that integrates NAS and generative adversarial networks (GANs) with recent advances in perceptual SR and pushes the efficiency of small perceptual SR models to facilitate on-device execution. Specifically, we search over the architectures of both the generator and the discriminator sequentially, highlighting the unique challenges and key observations of searching for an SR-optimized discriminator and comparing them with existing discriminator architectures in the literature. Our tiny perceptual SR (TPSR) models outperform SRGAN and EnhanceNet on both full-reference perceptual metric (LPIPS) and distortion metric (PSNR) while being up to 26.4$\times$ more memory efficient and 33.6$\times$ more compute efficient respectively.