DAMO-YOLO : A Report on Real-Time Object Detection Design

In this report, we present a fast and accurate object detection method dubbed DAMO-YOLO, which achieves higher performance than the state-of-the-art YOLO series. DAMO-YOLO is extended from YOLO with some new technologies, including Neural Architecture Search (NAS), efficient Reparameterized Generalized-FPN (RepGFPN), a lightweight head with AlignedOTA label assignment, and distillation enhancement. In particular, we use MAE-NAS, a method guided by the principle of maximum entropy, to search our detection backbone under the constraints of low latency and high performance, producing ResNet-like / CSP-like structures with spatial pyramid pooling and focus modules. In the design of necks and heads, we follow the rule of "large neck, small head". We import Generalized-FPN with accelerated queen-fusion to build the detector neck and upgrade its CSPNet with efficient layer aggregation networks (ELAN) and reparameterization. Then we investigate how detector head size affects detection performance and find that a heavy neck with only one task projection layer would yield better results. In addition, AlignedOTA is proposed to solve the misalignment problem in label assignment. And a distillation schema is introduced to improve performance to a higher level. Based on these new techs, we build a suite of models at various scales to meet the needs of different scenarios, i.e., DAMO-YOLO-Tiny/Small/Medium. They can achieve 43.0/46.8/50.0 mAPs on COCO with the latency of 2.78/3.83/5.62 ms on T4 GPUs respectively. The code is available at https://github.com/tinyvision/damo-yolo.

Mask the Correct Tokens: An Embarrassingly Simple Approach for Error Correction

Text error correction aims to correct the errors in text sequences such as those typed by humans or generated by speech recognition models. Previous error correction methods usually take the source (incorrect) sentence as encoder input and generate the target (correct) sentence through the decoder. Since the error rate of the incorrect sentence is usually low (e.g., 10\%), the correction model can only learn to correct on limited error tokens but trivially copy on most tokens (correct tokens), which harms the effective training of error correction. In this paper, we argue that the correct tokens should be better utilized to facilitate effective training and then propose a simple yet effective masking strategy to achieve this goal. Specifically, we randomly mask out a part of the correct tokens in the source sentence and let the model learn to not only correct the original error tokens but also predict the masked tokens based on their context information. Our method enjoys several advantages: 1) it alleviates trivial copy; 2) it leverages effective training signals from correct tokens; 3) it is a plug-and-play module and can be applied to different models and tasks. Experiments on spelling error correction and speech recognition error correction on Mandarin datasets and grammar error correction on English datasets with both autoregressive and non-autoregressive generation models show that our method improves the correction accuracy consistently.

Disentangled Feature Learning for Real-Time Neural Speech Coding

Recently end-to-end neural audio/speech coding has shown its great potential to outperform traditional signal analysis based audio codecs. This is mostly achieved by following the VQ-VAE paradigm where blind features are learned, vector-quantized and coded. In this paper, instead of blind end-to-end learning, we propose to learn disentangled features for real-time neural speech coding. Specifically, more global-like speaker identity and local content features are learned with disentanglement to represent speech. Such a compact feature decomposition not only achieves better coding efficiency by exploiting bit allocation among different features but also provides the flexibility to do audio editing in embedding space, such as voice conversion in real-time communications. Both subjective and objective results demonstrate its coding efficiency and we find that the learned disentangled features show comparable performance on any-to-any voice conversion with modern self-supervised speech representation learning models with far less parameters and low latency, showing the potential of our neural coding framework.

Differentially Private Optimizers Can Learn Adversarially Robust Models

Machine learning models have shone in a variety of domains and attracted increasing attention from both the security and the privacy communities. One important yet worrying question is: will training models under the differential privacy (DP) constraint unfavorably impact on the adversarial robustness? While previous works have postulated that privacy comes at the cost of worse robustness, we give the first theoretical analysis to show that DP models can indeed be robust and accurate, even sometimes more robust than their naturally-trained non-private counterparts. We observe three key factors that influence the privacy-robustness-accuracy tradeoff: (1) hyperparameters for DP optimizers are critical; (2) pre-training on public data significantly mitigates the accuracy and robustness drop; (3) choice of DP optimizers makes a difference. With these factors set properly, we achieve 90\% natural accuracy, 72\% robust accuracy ($+9\%$ than the non-private model) under $l_2(0.5)$ attack, and 69\% robust accuracy ($+16\%$ than the non-private model) with pre-trained SimCLRv2 model under $l_\infty(4/255)$ attack on CIFAR10 with $\epsilon=2$. In fact, we show both theoretically and empirically that DP models are Pareto optimal on the accuracy-robustness tradeoff. Empirically, the robustness of DP models is consistently observed on MNIST, Fashion MNIST and CelebA datasets, with ResNet and Vision Transformer. We believe our encouraging results are a significant step towards training models that are private as well as robust.

FV2ES: A Fully End2End Multimodal System for Fast Yet Effective Video Emotion Recognition Inference

In the latest social networks, more and more people prefer to express their emotions in videos through text, speech, and rich facial expressions. Multimodal video emotion analysis techniques can help understand users' inner world automatically based on human expressions and gestures in images, tones in voices, and recognized natural language. However, in the existing research, the acoustic modality has long been in a marginal position as compared to visual and textual modalities. That is, it tends to be more difficult to improve the contribution of the acoustic modality for the whole multimodal emotion recognition task. Besides, although better performance can be obtained by introducing common deep learning methods, the complex structures of these training models always result in low inference efficiency, especially when exposed to high-resolution and long-length videos. Moreover, the lack of a fully end-to-end multimodal video emotion recognition system hinders its application. In this paper, we designed a fully multimodal video-to-emotion system (named FV2ES) for fast yet effective recognition inference, whose benefits are threefold: (1) The adoption of the hierarchical attention method upon the sound spectra breaks through the limited contribution of the acoustic modality and outperforms the existing models' performance on both IEMOCAP and CMU-MOSEI datasets; (2) the introduction of the idea of multi-scale for visual extraction while single-branch for inference brings higher efficiency and maintains the prediction accuracy at the same time; (3) the further integration of data pre-processing into the aligned multimodal learning model allows the significant reduction of computational costs and storage space.

KuaiRand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos

Recommender systems deployed in real-world applications can have inherent exposure bias, which leads to the biased logged data plaguing the researchers. A fundamental way to address this thorny problem is to collect users' interactions on randomly expose items, i.e., the missing-at-random data. A few works have asked certain users to rate or select randomly recommended items, e.g., Yahoo!, Coat, and OpenBandit. However, these datasets are either too small in size or lack key information, such as unique user ID or the features of users/items. In this work, we present KuaiRand, an unbiased sequential recommendation dataset containing millions of intervened interactions on randomly exposed videos, collected from the video-sharing mobile App, Kuaishou. Different from existing datasets, KuaiRand records 12 kinds of user feedback signals (e.g., click, like, and view time) on randomly exposed videos inserted in the recommendation feeds in two weeks. To facilitate model learning, we further collect rich features of users and items as well as users' behavior history. By releasing this dataset, we enable the research of advanced debiasing large-scale recommendation scenarios for the first time. Also, with its distinctive features, KuaiRand can support various other research directions such as interactive recommendation, long sequential behavior modeling, and multi-task learning. The dataset and its news will be available at https://kuairand.com.

Real-time Short Video Recommendation on Mobile Devices

Short video applications have attracted billions of users in recent years, fulfilling their various needs with diverse content. Users usually watch short videos on many topics on mobile devices in a short period of time, and give explicit or implicit feedback very quickly to the short videos they watch. The recommender system needs to perceive users' preferences in real-time in order to satisfy their changing interests. Traditionally, recommender systems deployed at server side return a ranked list of videos for each request from client. Thus it cannot adjust the recommendation results according to the user's real-time feedback before the next request. Due to client-server transmitting latency, it is also unable to make immediate use of users' real-time feedback. However, as users continue to watch videos and feedback, the changing context leads the ranking of the server-side recommendation system inaccurate. In this paper, we propose to deploy a short video recommendation framework on mobile devices to solve these problems. Specifically, we design and deploy a tiny on-device ranking model to enable real-time re-ranking of server-side recommendation results. We improve its prediction accuracy by exploiting users' real-time feedback of watched videos and client-specific real-time features. With more accurate predictions, we further consider interactions among candidate videos, and propose a context-aware re-ranking method based on adaptive beam search. The framework has been deployed on Kuaishou, a billion-user scale short video application, and improved effective view, like and follow by 1.28\%, 8.22\% and 13.6\% respectively.

Higher-order accurate two-sample network inference and network hashing

Two-sample hypothesis testing for comparing two networks is an important yet difficult problem. Major challenges include: potentially different sizes and sparsity levels; non-repeated observations of adjacency matrices; computational scalability; and theoretical investigations, especially on finite-sample accuracy and minimax optimality. In this article, we propose the first provably higher-order accurate two-sample inference method by comparing network moments. Our method extends the classical two-sample t-test to the network setting. We make weak modeling assumptions and can effectively handle networks of different sizes and sparsity levels. We establish strong finite-sample theoretical guarantees, including rate-optimality properties. Our method is easy to implement and computes fast. We also devise a novel nonparametric framework of offline hashing and fast querying particularly effective for maintaining and querying very large network databases. We demonstrate the effectiveness of our method by comprehensive simulations. We apply our method to two real-world data sets and discover interesting novel structures.

Predictive Neural Speech Coding

Neural audio/speech coding has shown its capability to deliver a high quality at much lower bitrates than traditional methods recently. However, existing neural audio/speech codecs employ either acoustic features or learned blind features with a convolutional neural network for encoding, by which there are still temporal redundancies inside encoded features. This paper introduces latent-domain predictive coding into the VQ-VAE framework to fully remove such redundancies and proposes the TF-Codec for low-latency neural speech coding in an end-to-end way. Specifically, the extracted features are encoded conditioned on a prediction from past quantized latent frames so that temporal correlations are further removed. What's more, we introduce a learnable compression on the time-frequency input to adaptively adjust the attention paid on main frequencies and details at different bitrates. A differentiable vector quantization scheme based on distance-to-soft mapping and Gumbel-Softmax is proposed to better model the latent distributions with rate constraint. Subjective results on multilingual speech datasets show that with a latency of 40ms, the proposed TF-Codec at 1kbps can achieve a much better quality than Opus 9kbps and TF-Codec at 3kbps outperforms both EVS 9.6kbps and Opus 12kbps. Numerous studies are conducted to show the effectiveness of these techniques.

Cross-Scale Vector Quantization for Scalable Neural Speech Coding

Bitrate scalability is a desirable feature for audio coding in real-time communications. Existing neural audio codecs usually enforce a specific bitrate during training, so different models need to be trained for each target bitrate, which increases the memory footprint at the sender and the receiver side and transcoding is often needed to support multiple receivers. In this paper, we introduce a cross-scale scalable vector quantization scheme (CSVQ), in which multi-scale features are encoded progressively with stepwise feature fusion and refinement. In this way, a coarse-level signal is reconstructed if only a portion of the bitstream is received, and progressively improves the quality as more bits are available. The proposed CSVQ scheme can be flexibly applied to any neural audio coding network with a mirrored auto-encoder structure to achieve bitrate scalability. Subjective results show that the proposed scheme outperforms the classical residual VQ (RVQ) with scalability. Moreover, the proposed CSVQ at 3 kbps outperforms Opus at 9 kbps and Lyra at 3kbps and it could provide a graceful quality boost with bitrate increase.