Transformers have achieved state-of-the-art performance across various tasks, but suffer from a notable quadratic complexity in sequence length due to the attention mechanism. In this work, we propose MonarchAttention -- a novel approach to sub-quadratic attention approximation via Monarch matrices, an expressive class of structured matrices. Based on the variational form of softmax, we describe an efficient optimization-based algorithm to compute an approximate projection of softmax attention onto the class of Monarch matrices with $\Theta(N\sqrt{N} d)$ computational complexity and $\Theta(Nd)$ memory/IO complexity. Unlike previous approaches, MonarchAttention is both (1) transferable, yielding minimal performance loss with no additional training, even when replacing every attention layer of the transformer, and (2) hardware-efficient, utilizing the highest-throughput tensor core units on modern GPUs. With optimized kernels, MonarchAttention achieves substantial speed-ups in wall-time over FlashAttention-2: $1.4\times$ for shorter sequences $(N=256)$, $4.5\times$ for medium-length sequences $(N=4K)$, and $8.2\times$ for longer sequences $(N=16K)$. We demonstrate the quality of MonarchAttention on diverse tasks and architectures in vision and language problems, showing that it flexibly and accurately approximates softmax attention in a variety of contexts.

This paper investigates efficient deep neural networks (DNNs) to replace dense unstructured weight matrices with structured ones that possess desired properties. The challenge arises because the optimal weight matrix structure in popular neural network models is obscure in most cases and may vary from layer to layer even in the same network. Prior structured matrices proposed for efficient DNNs were mostly hand-crafted without a generalized framework to systematically learn them. To address this issue, we propose a generalized and differentiable framework to learn efficient structures of weight matrices by gradient descent. We first define a new class of structured matrices that covers a wide range of structured matrices in the literature by adjusting the structural parameters. Then, the frequency-domain differentiable parameterization scheme based on the Gaussian-Dirichlet kernel is adopted to learn the structural parameters by proximal gradient descent. On the image and language tasks, our method learns efficient DNNs with structured matrices, achieving lower complexity and/or higher performance than prior approaches that employ low-rank, block-sparse, or block-low-rank matrices.

Large-scale foundation models have demonstrated exceptional performance in language and vision tasks. However, the numerous dense matrix-vector operations involved in these large networks pose significant computational challenges during inference. To address these challenges, we introduce the Block-Level Adaptive STructured (BLAST) matrix, designed to learn and leverage efficient structures prevalent in the weight matrices of linear layers within deep learning models. Compared to existing structured matrices, the BLAST matrix offers substantial flexibility, as it can represent various types of structures that are either learned from data or computed from pre-existing weight matrices. We demonstrate the efficiency of using the BLAST matrix for compressing both language and vision tasks, showing that (i) for medium-sized models such as ViT and GPT-2, training with BLAST weights boosts performance while reducing complexity by 70\% and 40\%, respectively; and (ii) for large foundation models such as Llama-7B and DiT-XL, the BLAST matrix achieves a 2x compression while exhibiting the lowest performance degradation among all tested structured matrices. Our code is available at https://github.com/changwoolee/BLAST.

In this paper, we propose an iterative source error correction (ISEC) decoding scheme for deep-learning-based joint source-channel coding (Deep JSCC). Given a noisy codeword received through the channel, we use a Deep JSCC encoder and decoder pair to update the codeword iteratively to find a (modified) maximum a-posteriori (MAP) solution. For efficient MAP decoding, we utilize a neural network-based denoiser to approximate the gradient of the log-prior density of the codeword space. Albeit the non-convexity of the optimization problem, our proposed scheme improves various distortion and perceptual quality metrics from the conventional one-shot (non-iterative) Deep JSCC decoding baseline. Furthermore, the proposed scheme produces more reliable source reconstruction results compared to the baseline when the channel noise characteristics do not match the ones used during training.

The combination of pre-trained models and task-specific fine-tuning schemes, such as BERT, has achieved great success in various natural language processing (NLP) tasks. However, the large memory and computation costs of such models make it challenging to deploy them in edge devices. Moreover, in real-world applications like chatbots, multiple NLP tasks need to be processed together to achieve higher response credibility. Running multiple NLP tasks with specialized models for each task increases the latency and memory cost latency linearly with the number of tasks. Though there have been recent works on parameter-shared tuning that aim to reduce the total parameter size by partially sharing weights among multiple tasks, computation remains intensive and redundant despite different tasks using the same input. In this work, we identify that a significant portion of activations and weights can be reused among different tasks, to reduce cost and latency for efficient multi-task NLP. Specifically, we propose TaskFusion, an efficient transfer learning software-hardware co-design that exploits delta sparsity in both weights and activations to boost data sharing among tasks. For training, TaskFusion uses ℓ1 regularization on delta activation to learn inter-task data redundancies. A novel hardware-aware sub-task inference algorithm is proposed to exploit the dual delta sparsity. We then designed a dedicated heterogeneous architecture to accelerate multi-task inference with an optimized scheduling to increase hardware utilization and reduce off-chip memory access. Extensive experiments demonstrate that TaskFusion can reduce the number of floating point operations (FLOPs) by over 73% in multi-task NLP with negligible accuracy loss, while adding a new task at the cost of only < 2% parameter size increase. With the proposed architecture and optimized scheduling, Task-Fusion can achieve 1.48--2.43× performance and 1.62--3.77× energy efficiency than those using state-of-the-art single-task accelerators for multi-task NLP applications.

Massive machine type communication (mMTC) has attracted new coding schemes optimized for reliable short message transmission. In this paper, a novel deep learning-based near-orthogonal superposition (NOS) coding scheme is proposed to transmit short messages in multiple-input multiple-output (MIMO) channels for mMTC applications. In the proposed MIMO-NOS scheme, a neural network-based encoder is optimized via end-to-end learning with a corresponding neural network-based detector/decoder in a superposition-based auto-encoder framework including a MIMO channel. The proposed MIMO-NOS encoder spreads the information bits to multiple near-orthogonal high dimensional vectors to be combined (superimposed) into a single vector and reshaped for the space-time transmission. For the receiver, we propose a novel looped$K$-best tree-search algorithm with cyclic redundancy check (CRC) assistance to enhance the error correcting ability in the block-fading MIMO channel. For a comprehensive understanding of the proposed MIMO-NOS scheme, we further quantify the gain from individual components/modules in the framework, and analyze the decoding complexity measured by the floating point operations (FLOPs). Simulation results show the proposed MIMO-NOS scheme outperforms maximum likelihood (ML) MIMO detection combined with a polar code with CRC-assisted list decoding by 1 – 2 dB in various MIMO systems for short (32 – 64 bit) message transmission.

In this paper, we propose a confidence-calibration method for predicting the winner of a famous multiplayer online battle arena (MOBA) game, League of Legends. In MOBA games, the dataset may contain a large amount of input-dependent noise; not all of such noise is observable. Hence, it is desirable to attempt a confidence-calibrated prediction. Unfortunately, most existing confidence calibration methods are pertaining to image and document classification tasks where consideration on uncertainty is not crucial. In this paper, we propose a novel calibration method that takes data uncertainty into consideration. The proposed method achieves an outstanding expected calibration error (ECE) (0.57%) mainly owing to data uncertainty consideration, compared to a conventional temperature scaling method of which ECE value is 1.11%.

In this paper, a new learning method to quantify data uncertainty without suffering from performance degradation in Single Image Super Resolution (SISR) is proposed. Our work is motivated by the fact that the idea of loss design for capturing uncertainty and that for solving SISR are contradictory. As to capturing data uncertainty, we often model the output of a network as a Euclidian distance divided by a predictive variance, negative log-likelihood (NLL) for the Gaussian distribution, so that images with high variance have less impact on training. On the other hand, in the SISR domain, recent works give more weights to the loss of challenging images to improve the performance by using attention models. Nonetheless, the conflict should be handled to make neural networks capable of predicting the uncertainty of a super-resolved image, without suffering from performance degradation. Therefore, we propose a method called Gradient Rescaling Attention Model (GRAM) that combines both attempts effectively. Since variance may reflect the difficulty of an image, we rescale the gradient of NLL by the degree of variance. Hence, the neural network can focus on the challenging images, similarly to attention models. We conduct performance evaluation using standard SISR benchmarks in terms of peak signal-noise ratio (PSNR) and structural similarity (SSIM). The experimental results show that the proposed gradient rescaling method generates negligible performance degradation compared to SISR outputs with the Euclidian loss, whereas NLL without attention degrades the SR quality.