Complexity boosted adaptive training for better low resource ASR performance

During the entire training process of the ASR model, the intensity of data augmentation and the approach of calculating training loss are applied in a regulated manner based on preset parameters. For example, SpecAugment employs a predefined strength of augmentation to mask parts of the time-frequency domain spectrum. Similarly, in CTC-based multi-layer models, the loss is generally determined based on the output of the encoder's final layer during the training process. However, ignoring dynamic characteristics may suboptimally train models. To address the issue, we present a two-stage training method, known as complexity-boosted adaptive (CBA) training. It involves making dynamic adjustments to data augmentation strategies and CTC loss propagation based on the complexity of the training samples. In the first stage, we train the model with intermediate-CTC-based regularization and data augmentation without any adaptive policy. In the second stage, we propose a novel adaptive policy, called MinMax-IBF, which calculates the complexity of samples. We combine the MinMax-IBF policy to data augmentation and intermediate CTC loss regularization to continue training. The proposed CBA training approach shows considerable improvements, up to 13.4% and 14.1% relative reduction in WER on the LibriSpeech 100h test-clean and test-other dataset and also up to 6.3% relative reduction on AISHELL-1 test set, over the Conformer architecture in Wenet.

Sample adaptive data augmentation with progressive scheduling

Data augmentation is a widely adopted technique utilized to improve the robustness of automatic speech recognition (ASR). Employing a fixed data augmentation strategy for all training data is a common practice. However, it is important to note that there can be variations in factors such as background noise, speech rate, etc. among different samples within a single training batch. By using a fixed augmentation strategy, there is a risk that the model may reach a suboptimal state. In addition to the risks of employing a fixed augmentation strategy, the model's capabilities may differ across various training stages. To address these issues, this paper proposes the method of sample-adaptive data augmentation with progressive scheduling(PS-SapAug). The proposed method applies dynamic data augmentation in a two-stage training approach. It employs hybrid normalization to compute sample-specific augmentation parameters based on each sample's loss. Additionally, the probability of augmentation gradually increases throughout the training progression. Our method is evaluated on popular ASR benchmark datasets, including Aishell-1 and Librispeech-100h, achieving up to 8.13% WER reduction on LibriSpeech-100h test-clean, 6.23% on test-other, and 5.26% on AISHELL-1 test set, which demonstrate the efficacy of our approach enhancing performance and minimizing errors.

Lighten CARAFE: Dynamic Lightweight Upsampling with Guided Reassemble Kernels

As a fundamental operation in modern machine vision models, feature upsampling has been widely used and investigated in the literatures. An ideal upsampling operation should be lightweight, with low computational complexity. That is, it can not only improve the overall performance but also not affect the model complexity. Content-aware Reassembly of Features (CARAFE) is a well-designed learnable operation to achieve feature upsampling. Albeit encouraging performance achieved, this method requires generating large-scale kernels, which brings a mass of extra redundant parameters, and inherently has limited scalability. To this end, we propose a lightweight upsampling operation, termed Dynamic Lightweight Upsampling (DLU) in this paper. In particular, it first constructs a small-scale source kernel space, and then samples the large-scale kernels from the kernel space by introducing learnable guidance offsets, hence avoiding introducing a large collection of trainable parameters in upsampling. Experiments on several mainstream vision tasks show that our DLU achieves comparable and even better performance to the original CARAFE, but with much lower complexity, e.g., DLU requires 91% fewer parameters and at least 63% fewer FLOPs (Floating Point Operations) than CARAFE in the case of 16x upsampling, but outperforms the CARAFE by 0.3% mAP in object detection. Code is available at https://github.com/Fu0511/Dynamic-Lightweight-Upsampling.

LpQcM: Adaptable Lesion-Quantification-Consistent Modulation for Deep Learning Low-Count PET Image Denoising

Deep learning-based positron emission tomography (PET) image denoising offers the potential to reduce radiation exposure and scanning time by transforming low-count images into high-count equivalents. However, existing methods typically blur crucial details, leading to inaccurate lesion quantification. This paper proposes a lesion-perceived and quantification-consistent modulation (LpQcM) strategy for enhanced PET image denoising, via employing downstream lesion quantification analysis as auxiliary tools. The LpQcM is a plug-and-play design adaptable to a wide range of model architectures, modulating the sampling and optimization procedures of model training without adding any computational burden to the inference phase. Specifically, the LpQcM consists of two components, the lesion-perceived modulation (LpM) and the multiscale quantification-consistent modulation (QcM). The LpM enhances lesion contrast and visibility by allocating higher sampling weights and stricter loss criteria to lesion-present samples determined by an auxiliary segmentation network than lesion-absent ones. The QcM further emphasizes accuracy of quantification for both the mean and maximum standardized uptake value (SUVmean and SUVmax) across multiscale sub-regions throughout the entire image, thereby enhancing the overall image quality. Experiments conducted on large PET datasets from multiple centers and vendors, and varying noise levels demonstrated the LpQcM efficacy across various denoising frameworks. Compared to frameworks without LpQcM, the integration of LpQcM reduces the lesion SUVmean bias by 2.92% on average and increases the peak signal-to-noise ratio (PSNR) by 0.34 on average, for denoising images of extremely low-count levels below 10%.

Discrete Conditional Diffusion for Reranking in Recommendation

Reranking plays a crucial role in modern multi-stage recommender systems by rearranging the initial ranking list to model interplay between items. Considering the inherent challenges of reranking such as combinatorial searching space, some previous studies have adopted the evaluator-generator paradigm, with a generator producing feasible sequences and a evaluator selecting the best one based on estimated listwise utility. Inspired by the remarkable success of diffusion generative models, this paper explores the potential of diffusion models for generating high-quality sequences in reranking. However, we argue that it is nontrivial to take diffusion models as the generator in the context of recommendation. Firstly, diffusion models primarily operate in continuous data space, differing from the discrete data space of item permutations. Secondly, the recommendation task is different from conventional generation tasks as the purpose of recommender systems is to fulfill user interests. Lastly, real-life recommender systems require efficiency, posing challenges for the inference of diffusion models. To overcome these challenges, we propose a novel Discrete Conditional Diffusion Reranking (DCDR) framework for recommendation. DCDR extends traditional diffusion models by introducing a discrete forward process with tractable posteriors, which adds noise to item sequences through step-wise discrete operations (e.g., swapping). Additionally, DCDR incorporates a conditional reverse process that generates item sequences conditioned on expected user responses. Extensive offline experiments conducted on public datasets demonstrate that DCDR outperforms state-of-the-art reranking methods. Furthermore, DCDR has been deployed in a real-world video app with over 300 million daily active users, significantly enhancing online recommendation quality.

Measuring Item Global Residual Value for Fair Recommendation

In the era of information explosion, numerous items emerge every day, especially in feed scenarios. Due to the limited system display slots and user browsing attention, various recommendation systems are designed not only to satisfy users' personalized information needs but also to allocate items' exposure. However, recent recommendation studies mainly focus on modeling user preferences to present satisfying results and maximize user interactions, while paying little attention to developing item-side fair exposure mechanisms for rational information delivery. This may lead to serious resource allocation problems on the item side, such as the Snowball Effect. Furthermore, unfair exposure mechanisms may hurt recommendation performance. In this paper, we call for a shift of attention from modeling user preferences to developing fair exposure mechanisms for items. We first conduct empirical analyses of feed scenarios to explore exposure problems between items with distinct uploaded times. This points out that unfair exposure caused by the time factor may be the major cause of the Snowball Effect. Then, we propose to explicitly model item-level customized timeliness distribution, Global Residual Value (GRV), for fair resource allocation. This GRV module is introduced into recommendations with the designed Timeliness-aware Fair Recommendation Framework (TaFR). Extensive experiments on two datasets demonstrate that TaFR achieves consistent improvements with various backbone recommendation models. By modeling item-side customized Global Residual Value, we achieve a fairer distribution of resources and, at the same time, improve recommendation performance.

Alleviating Matthew Effect of Offline Reinforcement Learning in Interactive Recommendation

Offline reinforcement learning (RL), a technology that offline learns a policy from logged data without the need to interact with online environments, has become a favorable choice in decision-making processes like interactive recommendation. Offline RL faces the value overestimation problem. To address it, existing methods employ conservatism, e.g., by constraining the learned policy to be close to behavior policies or punishing the rarely visited state-action pairs. However, when applying such offline RL to recommendation, it will cause a severe Matthew effect, i.e., the rich get richer and the poor get poorer, by promoting popular items or categories while suppressing the less popular ones. It is a notorious issue that needs to be addressed in practical recommender systems. In this paper, we aim to alleviate the Matthew effect in offline RL-based recommendation. Through theoretical analyses, we find that the conservatism of existing methods fails in pursuing users' long-term satisfaction. It inspires us to add a penalty term to relax the pessimism on states with high entropy of the logging policy and indirectly penalizes actions leading to less diverse states. This leads to the main technical contribution of the work: Debiased model-based Offline RL (DORL) method. Experiments show that DORL not only captures user interests well but also alleviates the Matthew effect. The implementation is available via https://github.com/chongminggao/DORL-codes.

Tree based Progressive Regression Model for Watch-Time Prediction in Short-video Recommendation

An accurate prediction of watch time has been of vital importance to enhance user engagement in video recommender systems. To achieve this, there are four properties that a watch time prediction framework should satisfy: first, despite its continuous value, watch time is also an ordinal variable and the relative ordering between its values reflects the differences in user preferences. Therefore the ordinal relations should be reflected in watch time predictions. Second, the conditional dependence between the video-watching behaviors should be captured in the model. For instance, one has to watch half of the video before he/she finishes watching the whole video. Third, modeling watch time with a point estimation ignores the fact that models might give results with high uncertainty and this could cause bad cases in recommender systems. Therefore the framework should be aware of prediction uncertainty. Forth, the real-life recommender systems suffer from severe bias amplifications thus an estimation without bias amplification is expected. Therefore we propose TPM for watch time prediction. Specifically, the ordinal ranks of watch time are introduced into TPM and the problem is decomposed into a series of conditional dependent classification tasks which are organized into a tree structure. The expectation of watch time can be generated by traversing the tree and the variance of watch time predictions is explicitly introduced into the objective function as a measurement for uncertainty. Moreover, we illustrate that backdoor adjustment can be seamlessly incorporated into TPM, which alleviates bias amplifications. Extensive offline evaluations have been conducted in public datasets and TPM have been deployed in a real-world video app Kuaishou with over 300 million DAUs. The results indicate that TPM outperforms state-of-the-art approaches and indeed improves video consumption significantly.

Adaptive Contextual Biasing for Transducer Based Streaming Speech Recognition

By incorporating additional contextual information, deep biasing methods have emerged as a promising solution for speech recognition of personalized words. However, for real-world voice assistants, always biasing on such personalized words with high prediction scores can significantly degrade the performance of recognizing common words. To address this issue, we propose an adaptive contextual biasing method based on Context-Aware Transformer Transducer (CATT) that utilizes the biased encoder and predictor embeddings to perform streaming prediction of contextual phrase occurrences. Such prediction is then used to dynamically switch the bias list on and off, enabling the model to adapt to both personalized and common scenarios. Experiments on Librispeech and internal voice assistant datasets show that our approach can achieve up to 6.7% and 20.7% relative reduction in WER and CER compared to the baseline respectively, mitigating up to 96.7% and 84.9% of the relative WER and CER increase for common cases. Furthermore, our approach has a minimal performance impact in personalized scenarios while maintaining a streaming inference pipeline with negligible RTF increase.

FedFTN: Personalized Federated Learning with Deep Feature Transformation Network for Multi-institutional Low-count PET Denoising

Low-count PET is an efficient way to reduce radiation exposure and acquisition time, but the reconstructed images often suffer from low signal-to-noise ratio (SNR), thus affecting diagnosis and other downstream tasks. Recent advances in deep learning have shown great potential in improving low-count PET image quality, but acquiring a large, centralized, and diverse dataset from multiple institutions for training a robust model is difficult due to privacy and security concerns of patient data. Moreover, low-count PET data at different institutions may have different data distribution, thus requiring personalized models. While previous federated learning (FL) algorithms enable multi-institution collaborative training without the need of aggregating local data, addressing the large domain shift in the application of multi-institutional low-count PET denoising remains a challenge and is still highly under-explored. In this work, we propose FedFTN, a personalized federated learning strategy that addresses these challenges. FedFTN uses a local deep feature transformation network (FTN) to modulate the feature outputs of a globally shared denoising network, enabling personalized low-count PET denoising for each institution. During the federated learning process, only the denoising network's weights are communicated and aggregated, while the FTN remains at the local institutions for feature transformation. We evaluated our method using a large-scale dataset of multi-institutional low-count PET imaging data from three medical centers located across three continents, and showed that FedFTN provides high-quality low-count PET images, outperforming previous baseline FL reconstruction methods across all low-count levels at all three institutions.