Diffusion Transformers (DiTs) have emerged as a leading architecture for text-to-image synthesis, producing high-quality and photorealistic images. However, the quadratic scaling properties of the attention in DiTs hinder image generation with higher resolution or on devices with limited resources. This work introduces an efficient diffusion transformer (EDiT) to alleviate these efficiency bottlenecks in conventional DiTs and Multimodal DiTs (MM-DiTs). First, we present a novel linear compressed attention method that uses a multi-layer convolutional network to modulate queries with local information while keys and values are aggregated spatially. Second, we formulate a hybrid attention scheme for multimodal inputs that combines linear attention for image-to-image interactions and standard scaled dot-product attention for interactions involving prompts. Merging these two approaches leads to an expressive, linear-time Multimodal Efficient Diffusion Transformer (MM-EDiT). We demonstrate the effectiveness of the EDiT and MM-EDiT architectures by integrating them into PixArt-Sigma (conventional DiT) and Stable Diffusion 3.5-Medium (MM-DiT), achieving up to 2.2x speedup with comparable image quality after distillation.

Text-to-image synthesis has witnessed remarkable advancements in recent years. Many attempts have been made to adopt text-to-image models to support multiple tasks. However, existing approaches typically require resource-intensive re-training or additional parameters to accommodate for the new tasks, which makes the model inefficient for on-device deployment. We propose *Multi-Task Upcycling* (MTU), a simple yet effective recipe that extends the capabilities of a pre-trained text-to-image diffusion model to support a variety of image-to-image generation tasks. MTU replaces Feed-Forward Network (FFN) layers in the diffusion model with smaller FFNs, referred to as *experts*, and combines them with a dynamic routing mechanism. To the best of our knowledge, MTU is the first multi-task diffusion modeling approach that seamlessly blends multi-tasking with on-device compatibility, by mitigating the issue of parameter inflation. We show that the performance of MTU is on par with the single-task fine-tuned diffusion models across several tasks including *image editing, super-resolution*, and *inpainting*, while maintaining similar latency and computational load (GFLOPs) as the single-task fine-tuned models.

International audience

Increasing demand for on-device Automatic Speech Recognition (ASR) systems has resulted in renewed interests in developing automatic model compression techniques. Past research have shown that AutoML-based Low Rank Factorization (LRF) technique, when applied to an end-to-end Encoder-Attention-Decoder style ASR model, can achieve a speedup of up to 3.7x, outperforming laborious manual rank-selection approaches. However, we show that current AutoML-based search techniques only work up to a certain compression level, beyond which they fail to produce compressed models with acceptable word error rates (WER). In this work, we propose an iterative AutoML-based LRF approach that achieves over 5x compression without degrading the WER, thereby advancing the state-of-the-art in ASR compression.

LPCNet is an efficient vocoder that combines linear prediction and deep neural network modules to keep the computational complexity low. In this work, we present two techniques to further reduce it's complexity, aiming for a low-cost LPCNet vocoder-based neural Text-to-Speech (TTS) System. These techniques are: 1) Sample-bunching, which allows LPCNet to generate more than one audio sample per inference; and 2) Bit-bunching, which reduces the computations in the final layer of LPCNet. With the proposed bunching techniques, LPCNet, in conjunction with a Deep Convolutional TTS (DCTTS) acoustic model, shows a 2.19x improvement over the baseline run-time when running on a mobile device, with a less than 0.1 decrease in TTS mean opinion score (MOS).

Deep learning has enabled personal and IoT devices to rethink microphones as a multi-purpose sensor for understanding conversation and the surrounding environment. This resulted in a proliferation of Voice Controllable Systems (VCS) around us. The increasing popularity of such systems is also prone to attracting miscreants, who often want to take advantage of the VCS without the knowledge of the user. Consequently, understanding the robustness of VCS, especially under adversarial attacks, has become an important research topic. Although there exists some previous work on audio adversarial attacks, their scopes are limited to embedding the attacks onto pre-recorded music clips, which when played through speakers cause VCS to misbehave. As an attack-audio needs to be played, the occurrence of this type of attacks can be suspected by a human listener. In this paper, we focus on audio-based Denial-of-Service (DoS) attack, which is unexplored in the literature. Contrary to previous work, we show that adversarial audio attacks in real-time and overthe-air are possible, while a user interacts with VCS. We show that the attacks are effective regardless of the user's command and interaction timings. In this paper, we present a first-of-itskind imperceptible and always-on universal audio perturbation technique that enables such DoS attack to be successful. We thoroughly evaluate the performance of the attacking scheme across (i) two learning tasks, (ii) two model architectures and (iii) three datasets. We demonstrate that the attack can introduce as high as 78% error rate in audio recognition tasks.

We introduce a framework for the reconstruction of the amplitude, phase, and polarization of an optical vector-field using measurements acquired by an imaging device characterized by an integral transform with an unknown spatially variant kernel. By incorporating effective regularization terms, this new approach is able to recover an optical vector-field with respect to an arbitrary representation system, which may be different from the one used for device calibration. In particular, it enables the recovery of an optical vector-field with respect to a Fourier basis, which is shown to yield indicative features of increased scattering associated with tissue abnormalities. We demonstrate the effectiveness of our approach using synthetic holographic images and biological tissue samples in an experimental setting, where the measurements of an optical vector-field are acquired by a multicore fiber endoscope, and observe that indeed the recovered Fourier coefficients are useful in distinguishing healthy tissues from tumors in early stages of oesophageal cancer.

We consider the characterization of the asymptotic behavior of the average minimum mean-squared error (MMSE) in scalar and vector fading coherent channels, where the receiver knows the exact fading channel state but the transmitter knows only the fading channel distribution, driven by a range of inputs. In particular, we construct expansions of the quantities that are asymptotic in the signal-to-noise ratio (snr) for coherent channels subject to Rayleigh, Ricean or Nakagami fading and driven by discrete inputs and continuous inputs. The construction of these expansions leverages the fact that the average MMSE can be seen as an ψ-transform with a kernel of monotonic argument: this offers the means to use a powerful asymptotic expansion of integrals technique-the Mellin transform method-that leads immediately to the expansions of the average MMSE and- via the I-MMSE relationship-to expansions of the average mutual information, in terms of the so called canonical MMSE of a standard additive white Gaussian noise (AWGN) channel. We conclude with applications of the results to the optimization of the constrained capacity of a bank of parallel independent coherent fading channels driven by arbitrary discrete inputs.

We unveil asymptotic characterizations of the average minimum mean-squared error (MMSE) and the average mutual information in scalar fading coherent channels, where the receiver knows the exact fading channel state but the transmitter knows only the fading channel distribution, driven by a range of inputs both in the regimes of low-SNR - and at the heart of the novelty of the contribution - high-SNR. We also unveil connections to and generalizations of the MMSE dimension. By capitalizing on the characterizations, we conclude with applications of the results to the optimization of the constrained capacity of a bank of parallel independent coherent fading channels.