Sourav Bhattacharya

Shucong Zhang, Titouan Parcollet, Rogier van Dalen et al. · cambridge

Self-supervised learning (SSL) models usually require weeks of pre-training with dozens of high-end GPUs. These models typically have a multi-headed self-attention (MHSA) context encoder. However, MHSA takes quadratic time and space in the input length, contributing to the high pre-training cost. Linear-complexity alternatives to MHSA have been proposed. For instance, in supervised training, the SummaryMixing model is the first to outperform MHSA across multiple speech processing tasks. However, these cheaper alternatives have not been explored for SSL yet. This paper studies a linear-complexity context encoder for SSL for the first time. With better or equivalent performance for the downstream tasks of the MP3S benchmark, SummaryMixing reduces the pre-training time and peak VRAM of wav2vec 2.0 model by 18% and by 23%, respectively, leading to the pre-training of a 155M wav2vec 2.0 model finished within one week with 4 Tesla A100 GPUs. Code is available at https://github.com/SamsungLabs/SummaryMixing.

Titouan Parcollet, Rogier van Dalen, Shucong Zhang et al. · cambridge

Modern speech processing systems rely on self-attention. Unfortunately, token mixing with self-attention takes quadratic time in the length of the speech utterance, slowing down inference and training and increasing memory consumption. Cheaper alternatives to self-attention for ASR have been developed, but they fail to consistently reach the same level of accuracy. This paper, therefore, proposes a novel linear-time alternative to self-attention. It summarises an utterance with the mean over vectors for all time steps. This single summary is then combined with time-specific information. We call this method "SummaryMixing". Introducing SummaryMixing in state-of-the-art ASR models makes it feasible to preserve or exceed previous speech recognition performance while making training and inference up to 28% faster and reducing memory use by half.

Fuwen Tan, Royson Lee, Łukasz Dudziak et al.

Large language models (LLMs) have revolutionized language processing, delivering outstanding results across multiple applications. However, deploying LLMs on edge devices poses several challenges with respect to memory, energy, and compute costs, limiting their widespread use in devices such as mobile phones. A promising solution is to reduce the number of bits used to represent weights and activations. While existing works have found partial success at quantizing LLMs to lower bitwidths, e.g. 4-bit weights, quantizing activations beyond 16 bits often leads to large computational overheads due to poor on-device quantization support, or a considerable accuracy drop. Yet, 8-bit activations are very attractive for on-device deployment as they would enable LLMs to fully exploit mobile-friendly hardware, e.g. Neural Processing Units (NPUs). In this work, we make a first attempt to facilitate the on-device deployment of LLMs using integer-only quantization. We first investigate the limitations of existing quantization methods for on-device deployment, with a special focus on activation quantization. We then address these limitations by introducing a simple post-training quantization method, named MobileQuant, that extends previous weight equivalent transformation works by jointly optimizing the weight transformation and activation range parameters in an end-to-end manner. MobileQuant demonstrates superior capabilities over existing methods by 1) achieving near-lossless quantization on a wide range of LLM benchmarks, 2) reducing latency and energy consumption by 20\%-50\% compared to current on-device quantization strategies, 3) requiring limited compute budget, 4) being compatible with mobile-friendly compute units, e.g. NPU.

Shucong Zhang, Titouan Parcollet, Rogier van Dalen et al. · cambridge

Self-attention relies on positional embeddings to encode input order. Relative Position (RelPos) embeddings are widely used in Automatic Speech Recognition (ASR). However, RelPos has quadratic time complexity to input length and is often incompatible with fast GPU implementations of attention. In contrast, Rotary Positional Embedding (RoPE) rotates each input vector based on its absolute position, taking linear time to sequence length, implicitly encoding relative distances through self-attention dot products. Thus, it is usually compatible with efficient attention. However, its use in ASR remains underexplored. This work evaluates RoPE across diverse ASR tasks with training data ranging from 100 to 50,000 hours, covering various speech types (read, spontaneous, clean, noisy) and different accents in both streaming and non-streaming settings. ASR error rates are similar or better than RelPos, while training time is reduced by up to 21%. Code is available via the SpeechBrain toolkit.

Young D. Kwon, Rui Li, Sijia Li et al.

State-of-the-art text-to-image diffusion models (DMs) achieve remarkable quality, yet their massive parameter scale (8-11B) poses significant challenges for inferences on resource-constrained devices. In this paper, we present HierarchicalPrune, a novel compression framework grounded in a key observation: DM blocks exhibit distinct functional hierarchies, where early blocks establish semantic structures while later blocks handle texture refinements. HierarchicalPrune synergistically combines three techniques: (1) Hierarchical Position Pruning, which identifies and removes less essential later blocks based on position hierarchy; (2) Positional Weight Preservation, which systematically protects early model portions that are essential for semantic structural integrity; and (3) Sensitivity-Guided Distillation, which adjusts knowledge-transfer intensity based on our discovery of block-wise sensitivity variations. As a result, our framework brings billion-scale diffusion models into a range more suitable for on-device inference, while preserving the quality of the output images. Specifically, combined with INT4 weight quantisation, HierarchicalPrune achieves 77.5-80.4% memory footprint reduction (e.g., from 15.8 GB to 3.2 GB) and 27.9-38.0% latency reduction, measured on server and consumer grade GPUs, with the minimum drop of 2.6% in GenEval score and 7% in HPSv2 score compared to the original model. Finally, our comprehensive user study with 85 participants demonstrates that HierarchicalPrune maintains perceptual quality comparable to the original model while significantly outperforming prior works.

Mehdi Noroozi, Alberto Gil Ramos, Luca Morreale et al.

Current sampling mechanisms for conditional diffusion models rely mainly on Classifier Free Guidance (CFG) to generate high-quality images. However, CFG requires several denoising passes in each time step, e.g., up to three passes in image editing tasks, resulting in excessive computational costs. This paper introduces a novel conditioning technique to ease the computational burden of the well-established guidance techniques, thereby significantly improving the inference time of diffusion models. We present Explicit Conditioning (EC) of the noise distribution on the input modalities to achieve this. Intuitively, we model the noise to guide the conditional diffusion model during the diffusion process. We present evaluations on image editing tasks and demonstrate that EC outperforms CFG in generating diverse high-quality images with significantly reduced computations.

Philipp Becker, Abhinav Mehrotra, Ruchika Chavhan et al.

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.

Yuan Tseng, Titouan Parcollet, Rogier van Dalen et al. · cambridge

Recent work suggests that large language models (LLMs) can improve performance of speech tasks compared to existing systems. To support their claims, results on LibriSpeech and Common Voice are often quoted. However, this work finds that a substantial amount of the LibriSpeech and Common Voice evaluation sets appear in public LLM pretraining corpora. This calls into question the reliability of findings drawn from these two datasets. To measure contamination impact, LLMs trained with/without contamination are compared. A contaminated LLM is more likely to generate test sentences it has seen during training. Then, speech recognisers based on LLMs are compared. They show only subtle error rate differences if the LLM is contaminated, but assign significantly higher probabilities to transcriptions seen during LLM training. Results show that LLM outputs can be biased by tiny amounts of data contamination, highlighting the importance of evaluating LLM-based speech systems with held-out data.

Ruchika Chavhan, Abhinav Mehrotra, Malcolm Chadwick et al.

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.

Rosco Hunter, Łukasz Dudziak, Mohamed S. Abdelfattah et al.

Text-to-image diffusion models have demonstrated unprecedented capabilities for flexible and realistic image synthesis. Nevertheless, these models rely on a time-consuming sampling procedure, which has motivated attempts to reduce their latency. When improving efficiency, researchers often use the original diffusion model to train an additional network designed specifically for fast image generation. In contrast, our approach seeks to reduce latency directly, without any retraining, fine-tuning, or knowledge distillation. In particular, we find the repeated calculation of attention maps to be costly yet redundant, and instead suggest reusing them during sampling. Our specific reuse strategies are based on ODE theory, which implies that the later a map is reused, the smaller the distortion in the final image. We empirically compare our reuse strategies with few-step sampling procedures of comparable latency, finding that reuse generates images that are closer to those produced by the original high-latency diffusion model.

Luca Morreale, Alberto Gil C. P. Ramos, Malcolm Chadwick et al.

State-of-the-art (SOTA) generative models have demonstrated impressive capabilities in image synthesis or text generation, often with a large capacity model. However, these large models cannot be deployed on smartphones due to the limited availability of on-board memory and computations. Quantization methods lower the precision of the model parameters, allowing for efficient computations, \eg, in \INT{8}. Although aggressive quantization addresses efficiency and memory constraints, preserving the quality of the model remains a challenge. To retain quality in previous aggressive quantization, we propose a new fractional bits quantization (\short) approach. The novelty is a simple yet effective idea: we progressively reduce the model's precision from 32 to 4 bits per parameter, and exploit the fractional bits during optimization to maintain high generation quality. We show that the \short{} yields improved quality on a variety of diffusion models, including SD3.5-Medium, Sana, \pixart, and FLUX.1-schnell, while achieving $4-7\%$ lower FiD than standard QAT. Finally, we deploy and run Sana on a Samsung S25U, which runs on the Qualcomm SM8750-AB Snapdragon 8 Elite Hexagon Tensor Processor (HTP).

Young D. Kwon, Abhinav Mehrotra, Malcolm Chadwick et al.

High-resolution (4K) image-to-image synthesis has become increasingly important for mobile applications. Existing diffusion models for image editing face significant challenges, in terms of memory and image quality, when deployed on resource-constrained devices. In this paper, we present MobilePicasso, a novel system that enables efficient image editing at high resolutions, while minimising computational cost and memory usage. MobilePicasso comprises three stages: (i) performing image editing at a standard resolution with hallucination-aware loss, (ii) applying latent projection to overcome going to the pixel space, and (iii) upscaling the edited image latent to a higher resolution with adaptive context-preserving tiling. Our user study with 46 participants reveals that MobilePicasso not only improves image quality by 18-48% but reduces hallucinations by 14-51% over existing methods. MobilePicasso demonstrates significantly lower latency, e.g., up to 55.8$\times$ speed-up, yet with a small increase in runtime memory, e.g., a mere 9% increase over prior work. Surprisingly, the on-device runtime of MobilePicasso is observed to be faster than a server-based high-resolution image editing model running on an A100 GPU.

Rogier C. van Dalen, Shucong Zhang, Titouan Parcollet et al. · cambridge

Speech recognisers usually perform optimally only in a specific environment and need to be adapted to work well in another. For adaptation to a new speaker, there is often too little data for fine-tuning to be robust, and that data is usually unlabelled. This paper proposes a combination of approaches to make adaptation to a single minute of data robust. First, instead of estimating the adaptation parameters with cross-entropy on a single error-prone hypothesis or "pseudo-label", this paper proposes a novel loss function, the conditional entropy over complete hypotheses. Using multiple hypotheses makes adaptation more robust to errors in the initial recognition. Second, a "speaker code" characterises a speaker in a vector short enough that it requires little data to estimate. On a far-field noise-augmented version of Common Voice, the proposed scheme yields a 20% relative improvement in word error rate on one minute of adaptation data, increasing on 10 minutes to 29%.

Adrian Bulat, Jean Kossaifi, Sourav Bhattacharya et al.

We propose defensive tensorization, an adversarial defence technique that leverages a latent high-order factorization of the network. The layers of a network are first expressed as factorized tensor layers. Tensor dropout is then applied in the latent subspace, therefore resulting in dense reconstructed weights, without the sparsity or perturbations typically induced by the randomization.Our approach can be readily integrated with any arbitrary neural architecture and combined with techniques like adversarial training. We empirically demonstrate the effectiveness of our approach on standard image classification benchmarks. We validate the versatility of our approach across domains and low-precision architectures by considering an audio classification task and binary networks. In all cases, we demonstrate improved performance compared to prior works.

Ravichander Vipperla, Sangjun Park, Kihyun Choo et al.

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).

Abhinav Mehrotra, Łukasz Dudziak, Jinsu Yeo et al.

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.

Royson Lee, Stylianos I. Venieris, Łukasz Dudziak et al.

In recent years, convolutional networks have demonstrated unprecedented performance in the image restoration task of super-resolution (SR). SR entails the upscaling of a single low-resolution image in order to meet application-specific image quality demands and plays a key role in mobile devices. To comply with privacy regulations and reduce the overhead of cloud computing, executing SR models locally on-device constitutes a key alternative approach. Nevertheless, the excessive compute and memory requirements of SR workloads pose a challenge in mapping SR networks on resource-constrained mobile platforms. This work presents MobiSR, a novel framework for performing efficient super-resolution on-device. Given a target mobile platform, the proposed framework considers popular model compression techniques and traverses the design space to reach the highest performing trade-off between image quality and processing speed. At run time, a novel scheduler dispatches incoming image patches to the appropriate model-engine pair based on the patch's estimated upscaling difficulty in order to meet the required image quality with minimum processing latency. Quantitative evaluation shows that the proposed framework yields on-device SR designs that achieve an average speedup of 2.13x over highly-optimized parallel difficulty-unaware mappings and 4.79x over highly-optimized single compute engine implementations.

Petar Veličković, Laurynas Karazija, Nicholas D. Lane et al.

We analyse multimodal time-series data corresponding to weight, sleep and steps measurements. We focus on predicting whether a user will successfully achieve his/her weight objective. For this, we design several deep long short-term memory (LSTM) architectures, including a novel cross-modal LSTM (X-LSTM), and demonstrate their superiority over baseline approaches. The X-LSTM improves parameter efficiency by processing each modality separately and allowing for information flow between them by way of recurrent cross-connections. We present a general hyperparameter optimisation technique for X-LSTMs, which allows us to significantly improve on the LSTM and a prior state-of-the-art cross-modal approach, using a comparable number of parameters. Finally, we visualise the model's predictions, revealing implications about latent variables in this task.

Sourav Bhattacharya, Otto Huhta, N. Asokan

Unsafe websites consist of malicious as well as inappropriate sites, such as those hosting questionable or offensive content. Website reputation systems are intended to help ordinary users steer away from these unsafe sites. However, the process of assigning safety ratings for websites typically involves humans. Consequently it is time consuming, costly and not scalable. This has resulted in two major problems: (i) a significant proportion of the web space remains unrated and (ii) there is an unacceptable time lag before new websites are rated. In this paper, we show that by leveraging structural and content-based properties of websites, it is possible to reliably and efficiently predict their safety ratings, thereby mitigating both problems. We demonstrate the effectiveness of our approach using four datasets of up to 90,000 websites. We use ratings from Web of Trust (WOT), a popular crowdsourced web reputation system, as ground truth. We propose a novel ensemble classification technique that makes opportunistic use of available structural and content properties of webpages to predict their eventual ratings in two dimensions used by WOT: trustworthiness and child safety. Ours is the first classification system to predict such subjective ratings and the same approach works equally well in identifying malicious websites. Across all datasets, our classification performs well with average F$_1$-score in the 74--90\% range.

Sourav Bhattacharya, Petteri Nurmi, Nils Hammerla et al.

We propose a sparse-coding framework for activity recognition in ubiquitous and mobile computing that alleviates two fundamental problems of current supervised learning approaches. (i) It automatically derives a compact, sparse and meaningful feature representation of sensor data that does not rely on prior expert knowledge and generalizes extremely well across domain boundaries. (ii) It exploits unlabeled sample data for bootstrapping effective activity recognizers, i.e., substantially reduces the amount of ground truth annotation required for model estimation. Such unlabeled data is trivial to obtain, e.g., through contemporary smartphones carried by users as they go about their everyday activities. Based on the self-taught learning paradigm we automatically derive an over-complete set of basis vectors from unlabeled data that captures inherent patterns present within activity data. Through projecting raw sensor data onto the feature space defined by such over-complete sets of basis vectors effective feature extraction is pursued. Given these learned feature representations, classification backends are then trained using small amounts of labeled training data. We study the new approach in detail using two datasets which differ in terms of the recognition tasks and sensor modalities. Primarily we focus on transportation mode analysis task, a popular task in mobile-phone based sensing. The sparse-coding framework significantly outperforms the state-of-the-art in supervised learning approaches. Furthermore, we demonstrate the great practical potential of the new approach by successfully evaluating its generalization capabilities across both domain and sensor modalities by considering the popular Opportunity dataset. Our feature learning approach outperforms state-of-the-art approaches to analyzing activities in daily living.

Hien Thi Thu Truong, Eemil Lagerspetz, Petteri Nurmi et al.

There is little information from independent sources in the public domain about mobile malware infection rates. The only previous independent estimate (0.0009%) [12], was based on indirect measurements obtained from domain name resolution traces. In this paper, we present the first independent study of malware infection rates and associated risk factors using data collected directly from over 55,000 Android devices. We find that the malware infection rates in Android devices estimated using two malware datasets (0.28% and 0.26%), though small, are significantly higher than the previous independent estimate. Using our datasets, we investigate how indicators extracted inexpensively from the devices correlate with malware infection. Based on the hypothesis that some application stores have a greater density of malicious applications and that advertising within applications and cross-promotional deals may act as infection vectors, we investigate whether the set of applications used on a device can serve as an indicator for infection of that device. Our analysis indicates that this alone is not an accurate indicator for pinpointing infection. However, it is a very inexpensive but surprisingly useful way for significantly narrowing down the pool of devices on which expensive monitoring and analysis mechanisms must be deployed. Using our two malware datasets we show that this indicator performs 4.8 and 4.6 times (respectively) better at identifying infected devices than the baseline of random checks. Such indicators can be used, for example, in the search for new or previously undetected malware. It is therefore a technique that can complement standard malware scanning by anti-malware tools. Our analysis also demonstrates a marginally significant difference in battery use between infected and clean devices.