Yike Zhang

Yike Zhang, Eduardo Davalos, Jack Noble

Accurate 6D pose estimation has gained more attention over the years for robotics-assisted tasks that require precise interaction with physical objects. This paper presents an interactive 3D-to-2D visualization and annotation tool to support the 6D pose estimation research community. To the best of our knowledge, the proposed work is the first tool that allows users to visualize and manipulate 3D objects interactively on a 2D real-world scene, along with a comprehensive user study. This system supports robust 6D camera pose annotation by providing both visual cues and spatial relationships to determine object position and orientation in various environments. The annotation feature in Vision6D is particularly helpful in scenarios where the transformation matrix between the camera and world objects is unknown, as it enables accurate annotation of these objects' poses using only the camera intrinsic matrix. This capability serves as a foundational step in developing and training advanced pose estimation models across various domains. We evaluate Vision6D's effectiveness by utilizing widely-used open-source pose estimation datasets Linemod and HANDAL through comparisons between the default ground-truth camera poses with manual annotations. A user study was performed to show that Vision6D generates accurate pose annotations via visual cues in an intuitive 3D user interface. This approach aims to bridge the gap between 2D scene projections and 3D scenes, offering an effective way for researchers and developers to solve 6D pose annotation related problems. The software is open-source and publicly available at https://github.com/InteractiveGL/vision6D.

Zhenyu Li, Yike Zhang, Tengyu Pan et al. · tsinghua

Empowering LLMs with the ability to precisely understand long contexts is crucial for many downstream applications. However, handling long contexts with conventional transformer architecture requires substantial training and inference resources. Existing context condensing methods cannot accurately understand the full context, as there is a considerable amount of information loss in the condensing process. To address these issues, we present FocusLLM, a framework designed to extend the fixed context length of any decoder-only LLM, allowing the model to focus on relevant information from very long sequences. FocusLLM first divides long text input into chunks based on the model's original context length. It then employs the dynamic condensing process to distill crucial information from each chunk. Ultimately, through the novel parallel decoding mechanism, FocusLLM can integrate the extracted information into its local context. FocusLLM stands out for great training efficiency and versatility: trained with an 8K input length and with much less training cost than previous methods, FocusLLM exhibits superior performance across downstream tasks and maintains strong language modeling ability when handling extensive long texts, even up to 400K tokens. Our code is available at https://github.com/leezythu/FocusLLM.

Kun Wei, Yike Zhang, Sining Sun et al.

Leveraging context information is an intuitive idea to improve performance on conversational automatic speech recognition(ASR). Previous works usually adopt recognized hypotheses of historical utterances as preceding context, which may bias the current recognized hypothesis due to the inevitable historicalrecognition errors. To avoid this problem, we propose an audio-textual cross-modal representation extractor to learn contextual representations directly from preceding speech. Specifically, it consists of two modal-related encoders, extracting high-level latent features from speech and the corresponding text, and a cross-modal encoder, which aims to learn the correlation between speech and text. We randomly mask some input tokens and input sequences of each modality. Then a token-missing or modal-missing prediction with a modal-level CTC loss on the cross-modal encoder is performed. Thus, the model captures not only the bi-directional context dependencies in a specific modality but also relationships between different modalities. Then, during the training of the conversational ASR system, the extractor will be frozen to extract the textual representation of preceding speech, while such representation is used as context fed to the ASR decoder through attention mechanism. The effectiveness of the proposed approach is validated on several Mandarin conversation corpora and the highest character error rate (CER) reduction up to 16% is achieved on the MagicData dataset.

Zhanheng Yang, Sining Sun, Xiong Wang et al.

It is difficult for an E2E ASR system to recognize words such as entities appearing infrequently in the training data. A widely used method to mitigate this issue is feeding contextual information into the acoustic model. Previous works have proven that a compact and accurate contextual list can boost the performance significantly. In this paper, we propose an efficient approach to obtain a high quality contextual list for a unified streaming/non-streaming based E2E model. Specifically, we make use of the phone-level streaming output to first filter the predefined contextual word list then fuse it into non-casual encoder and decoder to generate the final recognition results. Our approach improve the accuracy of the contextual ASR system and speed up the inference process. Experiments on two datasets demonstrates over 20% CER reduction comparing to the baseline system. Meanwhile, the RTF of our system can be stabilized within 0.15 when the size of the contextual word list grows over 6,000.

Yike Zhang, Jack Noble

AI-assisted surgeries have drawn the attention of the medical image research community due to their real-world impact on improving surgery success rates. For image-guided surgeries, such as Cochlear Implants (CIs), accurate object segmentation can provide useful information for surgeons before an operation. Recently published image segmentation methods that leverage machine learning usually rely on a large number of manually predefined ground truth labels. However, it is a laborious and time-consuming task to prepare the dataset. This paper presents a novel technique using a self-supervised 3D-UNet that produces a dense deformation field between an atlas and a target image that can be used for atlas-based segmentation of the ossicles. Our results show that our method outperforms traditional image segmentation methods and generates a more accurate boundary around the ossicles based on Dice similarity coefficient and point-to-point error comparison. The mean Dice coefficient is improved by 8.51% with our proposed method.

Ange Lou, Yamin Li, Xing Yao et al.

The accurate reconstruction of surgical scenes from surgical videos is critical for various applications, including intraoperative navigation and image-guided robotic surgery automation. However, previous approaches, mainly relying on depth estimation, have limited effectiveness in reconstructing surgical scenes with moving surgical tools. To address this limitation and provide accurate 3D position prediction for surgical tools in all frames, we propose a novel approach called SAMSNeRF that combines Segment Anything Model (SAM) and Neural Radiance Field (NeRF) techniques. Our approach generates accurate segmentation masks of surgical tools using SAM, which guides the refinement of the dynamic surgical scene reconstruction by NeRF. Our experimental results on public endoscopy surgical videos demonstrate that our approach successfully reconstructs high-fidelity dynamic surgical scenes and accurately reflects the spatial information of surgical tools. Our proposed approach can significantly enhance surgical navigation and automation by providing surgeons with accurate 3D position information of surgical tools during surgery.The source code will be released soon.

Ange Lou, Yamin Li, Yike Zhang et al.

The Segment Anything Model 2 (SAM 2) is the latest generation foundation model for image and video segmentation. Trained on the expansive Segment Anything Video (SA-V) dataset, which comprises 35.5 million masks across 50.9K videos, SAM 2 advances its predecessor's capabilities by supporting zero-shot segmentation through various prompts (e.g., points, boxes, and masks). Its robust zero-shot performance and efficient memory usage make SAM 2 particularly appealing for surgical tool segmentation in videos, especially given the scarcity of labeled data and the diversity of surgical procedures. In this study, we evaluate the zero-shot video segmentation performance of the SAM 2 model across different types of surgeries, including endoscopy and microscopy. We also assess its performance on videos featuring single and multiple tools of varying lengths to demonstrate SAM 2's applicability and effectiveness in the surgical domain. We found that: 1) SAM 2 demonstrates a strong capability for segmenting various surgical videos; 2) When new tools enter the scene, additional prompts are necessary to maintain segmentation accuracy; and 3) Specific challenges inherent to surgical videos can impact the robustness of SAM 2.

Yike Zhang, Eduardo Davalos, Dingjie Su et al.

Cochlear Implant (CI) procedures require the insertion of an electrode array into the cochlea within the inner ear. To achieve this, mastoidectomy, a surgical procedure involving the removal of part of the mastoid region of the temporal bone using a high-speed drill provides safe access to the cochlea through the middle and inner ear. In this paper, we propose a novel Mamba-based method to synthesize the mastoidectomy volume using only preoperative Computed Tomography (CT) scans, where the mastoid remains intact. Our approach introduces a self-supervised learning framework designed to predict the mastoidectomy shape and reconstruct a 3D post-mastoidectomy surface directly from preoperative CT scans. This reconstruction aligns with intraoperative microscope views, enabling various downstream surgical applications. For training, we leverage postoperative CT scans to bypass manual data cleaning and labeling, even when the region removed during mastoidectomy is affected by challenges such as metal artifacts, low signal-to-noise ratio, or electrode wiring. Our method achieves a mean Dice score of 0.70 in estimating mastoidectomy regions, demonstrating its effectiveness for accurate and efficient surgical preoperative planning.

Yike Zhang, Jack Noble

Cochlear Implant (CI) procedures involve performing an invasive mastoidectomy to insert an electrode array into the cochlea. In this paper, we introduce a novel pipeline that is capable of generating synthetic multi-view videos from a single CI microscope image. In our approach, we use a patient's pre-operative CT scan to predict the post-mastoidectomy surface using a method designed for this purpose. We manually align the surface with a selected microscope frame to obtain an accurate initial pose of the reconstructed CT mesh relative to the microscope. We then perform UV projection to transfer the colors from the frame to surface textures. Novel views of the textured surface can be used to generate a large dataset of synthetic frames with ground truth poses. We evaluated the quality of synthetic views rendered using Pytorch3D and PyVista. We found both rendering engines lead to similarly high-quality synthetic novel-view frames compared to ground truth with a structural similarity index for both methods averaging about 0.86. A large dataset of novel views with known poses is critical for ongoing training of a method to automatically estimate microscope pose for 2D to 3D registration with the pre-operative CT to facilitate augmented reality surgery. This dataset will empower various downstream tasks, such as integrating Augmented Reality (AR) in the OR, tracking surgical tools, and supporting other video analysis studies.

Zhiyuan He, Xufang Luo, Yike Zhang et al.

We propose VL Norm (Variance-reduced Length-dependent Normalization), a simple yet effective loss aggregation method tailored to the characteristic of dynamic generation lengths in Reinforcement Learning with Verifiable Rewards (RLVR). Recently, RLVR has demonstrated strong potential in improving the reasoning capabilities of large language models (LLMs), but a major challenge lies in the large variability of response lengths during training, which leads to high gradient variance and unstable optimization. Although previous methods such as GRPO, DAPO, and Dr. GRPO introduce different loss normalization terms to address this issue, they either produce biased estimates or still suffer from high gradient variance. By analyzing the effect of varying lengths on policy loss both theoretically and empirically, we reformulate the problem as finding a minimum-variance unbiased estimator. Our proposed VL Norm not only provides an unbiased estimate of the true policy loss but also minimizes gradient variance in theory. Besides, VL Norm is easy to implement with less than 10 lines of code change. Extensive experiments show that it consistently achieves superior results across different model sizes, maximum lengths, and tasks. When integrated into the state-of-the-art RL algorithm DAPO, it achieves up to 2.67x faster convergence on the CountDown task. Our code is public at https://github.com/zerolllin/Delta-L-Normalization.

Yike Zhang, Zhiyuan He, Huiqiang Jiang et al. · microsoft-research

Large language models (LLMs) enable long-context tasks but face efficiency challenges due to the growing key-value (KV) cache. We propose LeanK, a learning-based method that prunes unimportant key (K) cache channels by leveraging static channel sparsity. With a novel two-stage training process, LeanK learns channel-wise static mask that could satisfy specific sparsity ratio and hardware alignment requirement. LeanK reduces GPU memory and accelerates decoding without sacrificing accuracy. Experiments demonstrate up to 70% K cache and 16%-18% V cache memory reduction. Custom decoding kernel enables 1.3x speedup for attention computation. We also provide insights into model channels and attention heads during long-context inference by analyzing the learned importance distribution. Our code is available at https://aka.ms/LeanK.

Eduardo Davalos, Yike Zhang, Namrata Srivastava et al.

With advancements in AI, new gaze estimation methods are exceeding state-of-the-art (SOTA) benchmarks, but their real-world application reveals a gap with commercial eye-tracking solutions. Factors like model size, inference time, and privacy often go unaddressed. Meanwhile, webcam-based eye-tracking methods lack sufficient accuracy, in particular due to head movement. To tackle these issues, we introduce We bEyeTrack, a framework that integrates lightweight SOTA gaze estimation models directly in the browser. It incorporates model-based head pose estimation and on-device few-shot learning with as few as nine calibration samples (k < 9). WebEyeTrack adapts to new users, achieving SOTA performance with an error margin of 2.32 cm on GazeCapture and real-time inference speeds of 2.4 milliseconds on an iPhone 14. Our open-source code is available at https://github.com/RedForestAi/WebEyeTrack.

Heting Gao, Hang Shao, Xiong Wang et al.

The film Her features Samantha, a sophisticated AI audio agent who is capable of understanding both linguistic and paralinguistic information in human speech and delivering real-time responses that are natural, informative and sensitive to emotional subtleties. Moving one step toward more sophisticated audio agent from recent advancement in end-to-end (E2E) speech systems, we propose LUCY, a E2E speech model that (1) senses and responds to user's emotion, (2) deliver responses in a succinct and natural style, and (3) use external tool to answer real-time inquiries. Experiment results show that LUCY is better at emotion control than peer models, generating emotional responses based on linguistic emotional instructions and responding to paralinguistic emotional cues. Lucy is also able to generate responses in a more natural style, as judged by external language models, without sacrificing much performance on general question answering. Finally, LUCY can leverage function calls to answer questions that are out of its knowledge scope.

Joyce Fonteles, Eduardo Davalos, Ashwin T. S. et al.

Investigating children's embodied learning in mixed-reality environments, where they collaboratively simulate scientific processes, requires analyzing complex multimodal data to interpret their learning and coordination behaviors. Learning scientists have developed Interaction Analysis (IA) methodologies for analyzing such data, but this requires researchers to watch hours of videos to extract and interpret students' learning patterns. Our study aims to simplify researchers' tasks, using Machine Learning and Multimodal Learning Analytics to support the IA processes. Our study combines machine learning algorithms and multimodal analyses to support and streamline researcher efforts in developing a comprehensive understanding of students' scientific engagement through their movements, gaze, and affective responses in a simulated scenario. To facilitate an effective researcher-AI partnership, we present an initial case study to determine the feasibility of visually representing students' states, actions, gaze, affect, and movement on a timeline. Our case study focuses on a specific science scenario where students learn about photosynthesis. The timeline allows us to investigate the alignment of critical learning moments identified by multimodal and interaction analysis, and uncover insights into students' temporal learning progressions.

Yike Zhang, Eduardo Davalos, Dingjie Su et al.

For those experiencing severe-to-profound sensorineural hearing loss, the cochlear implant (CI) is the preferred treatment. Augmented reality (AR) aided surgery can potentially improve CI procedures and hearing outcomes. Typically, AR solutions for image-guided surgery rely on optical tracking systems to register pre-operative planning information to the display so that hidden anatomy or other important information can be overlayed and co-registered with the view of the surgical scene. In this paper, our goal is to develop a method that permits direct 2D-to-3D registration of the microscope video to the pre-operative Computed Tomography (CT) scan without the need for external tracking equipment. Our proposed solution involves using surface mapping of a portion of the incus in surgical recordings and determining the pose of this structure relative to the surgical microscope by performing pose estimation via the perspective-n-point (PnP) algorithm. This registration can then be applied to pre-operative segmentations of other anatomy-of-interest, as well as the planned electrode insertion trajectory to co-register this information for the AR display. Our results demonstrate the accuracy with an average rotation error of less than 25 degrees and a translation error of less than 2 mm, 3 mm, and 0.55% for the x, y, and z axes, respectively. Our proposed method has the potential to be applicable and generalized to other surgical procedures while only needing a monocular microscope during intra-operation.

Yutao Sun, Zhenyu Li, Yike Zhang et al.

Transformer-based architectures have become the prevailing backbone of large language models. However, the quadratic time and memory complexity of self-attention remains a fundamental obstacle to efficient long-context modeling. To address this limitation, recent research has introduced two principal categories of efficient attention mechanisms. Linear attention methods achieve linear complexity through kernel approximations, recurrent formulations, or fastweight dynamics, thereby enabling scalable inference with reduced computational overhead. Sparse attention techniques, in contrast, limit attention computation to selected subsets of tokens based on fixed patterns, block-wise routing, or clustering strategies, enhancing efficiency while preserving contextual coverage. This survey provides a systematic and comprehensive overview of these developments, integrating both algorithmic innovations and hardware-level considerations. In addition, we analyze the incorporation of efficient attention into largescale pre-trained language models, including both architectures built entirely on efficient attention and hybrid designs that combine local and global components. By aligning theoretical foundations with practical deployment strategies, this work aims to serve as a foundational reference for advancing the design of scalable and efficient language models.

Eduardo Davalos, Yike Zhang, Namrata Srivastava et al.

Reading assessments are essential for enhancing students' comprehension, yet many EdTech applications focus mainly on outcome-based metrics, providing limited insights into student behavior and cognition. This study investigates the use of multimodal data sources -- including eye-tracking data, learning outcomes, assessment content, and teaching standards -- to derive meaningful reading insights. We employ unsupervised learning techniques to identify distinct reading behavior patterns, and then a large language model (LLM) synthesizes the derived information into actionable reports for educators, streamlining the interpretation process. LLM experts and human educators evaluate these reports for clarity, accuracy, relevance, and pedagogical usefulness. Our findings indicate that LLMs can effectively function as educational analysts, turning diverse data into teacher-friendly insights that are well-received by educators. While promising for automating insight generation, human oversight remains crucial to ensure reliability and fairness. This research advances human-centered AI in education, connecting data-driven analytics with practical classroom applications.

Yike Zhang, Eduardo Davalos, Jack Noble

Recent deep learning-based image completion methods, including both inpainting and outpainting, have demonstrated promising results in restoring corrupted images by effectively filling various missing regions. Among these, Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models (DDPMs) have been employed as key generative image completion approaches, excelling in the field of generating high-quality restorations with reduced artifacts and improved fine details. In previous work, we developed a method aimed at synthesizing views from novel microscope positions for mastoidectomy surgeries; however, that approach did not have the ability to restore the surrounding surgical scene environment. In this paper, we propose an efficient method to complete the surgical scene of the synthetic postmastoidectomy dataset. Our approach leverages self-supervised learning on real surgical datasets to train a Single-Step Denoising Diffusion-GAN (SSDD-GAN), combining the advantages of diffusion models with the adversarial optimization of GANs for improved Structural Similarity results of 6%. The trained model is then directly applied to the synthetic postmastoidectomy dataset using a zero-shot approach, enabling the generation of realistic and complete surgical scenes without the need for explicit ground-truth labels from the synthetic postmastoidectomy dataset. This method addresses key limitations in previous work, offering a novel pathway for full surgical microscopy scene completion and enhancing the usability of the synthetic postmastoidectomy dataset in surgical preoperative planning and intraoperative navigation.

Eduardo Davalos, Jorge Alberto Salas, Yike Zhang et al.

Understanding reader behaviors such as skimming, deep reading, and scanning is essential for improving educational instruction. While prior eye-tracking studies have trained models to recognize reading behaviors, they often rely on instructed reading tasks, which can alter natural behaviors and limit the applicability of these findings to in-the-wild settings. Additionally, there is a lack of clear definitions for reading behavior archetypes in the literature. We conducted a classroom study to address these issues by collecting instructed and in-the-wild reading data. We developed a mixed-method framework, including a human-driven theoretical model, statistical analyses, and an AI classifier, to differentiate reading behaviors based on their velocity, density, and sequentiality. Our lightweight 2D CNN achieved an F1 score of 0.8 for behavior recognition, providing a robust approach for understanding in-the-wild reading. This work advances our ability to provide detailed behavioral insights to educators, supporting more targeted and effective assessment and instruction.

Eduardo Davalos, Yike Zhang

The rapid integration of conversational AI systems into educational settings has intensified ethical concerns about academic integrity, fairness, and students' cognitive development. Institutional responses have largely centered on AI detection tools and restrictive policies, yet such approaches have proven unreliable and ethically contentious. This paper reframes AI misuse in education not primarily as a detection problem, but as a measurement problem rooted in the loss of visibility into the learning process. When AI enters the assessment loop, educators often retain access to final outputs but lose valuable insight into how those outputs were produced. Drawing on research in cognitive offloading, learning analytics, and multimodal timeline reconstruction, we propose the Learning Visibility Framework, grounded in three principles: clear specification and modeling of acceptable AI use, recognition of learning processes as assessable evidence alongside outcomes, and the establishment of transparent timelines of student activity. Rather than promoting surveillance, the framework emphasizes transparency and shared evidence as foundations for ethical AI integration in classroom settings. By shifting focus from adversarial detection toward process visibility, this work offers a principled pathway for aligning AI use with educational values while preserving trust and transparency between students and educators

Yike Zhang, Eduardo Davalos, Dingjie Su et al.

Cochlear Implant (CI) surgery treats severe hearing loss by inserting an electrode array into the cochlea to stimulate the auditory nerve. An important step in this procedure is mastoidectomy, which removes part of the mastoid region of the temporal bone to provide surgical access. Accurate mastoidectomy shape prediction from preoperative imaging improves pre-surgical planning, reduces risks, and enhances surgical outcomes. Despite its importance, there are limited deep-learning-based studies regarding this topic due to the challenges of acquiring ground-truth labels. We address this gap by investigating self-supervised and weakly-supervised learning models to predict the mastoidectomy region without human annotations. We propose a hybrid self-supervised and weakly-supervised learning framework to predict the mastoidectomy region directly from preoperative CT scans, where the mastoid remains intact. Our hybrid method achieves a mean Dice score of 0.72 when predicting the complex and boundary-less mastoidectomy shape, surpassing state-of-the-art approaches and demonstrating strong performance. The method provides groundwork for constructing 3D postmastoidectomy surfaces directly from the corresponding preoperative CT scans. To our knowledge, this is the first work that integrating self-supervised and weakly-supervised learning for mastoidectomy shape prediction, offering a robust and efficient solution for CI surgical planning while leveraging 3D T-distribution loss in weakly-supervised medical imaging.

Eduardo Davalos, Yike Zhang, Shruti Jain et al.

Eye-tracking offers rich insights into student cognition and engagement, but remains underutilized in classroom-facing educational technology due to challenges in data interpretation and accessibility. In this paper, we present the iterative design and evaluation of a gaze-based learning analytics dashboard for English Language Arts (ELA), developed through five studies involving teachers and students. Guided by user-centered design and data storytelling principles, we explored how gaze data can support reflection, formative assessment, and instructional decision-making. Our findings demonstrate that gaze analytics can be approachable and pedagogically valuable when supported by familiar visualizations, layered explanations, and narrative scaffolds. We further show how a conversational agent, powered by a large language model (LLM), can lower cognitive barriers to interpreting gaze data by enabling natural language interactions with multimodal learning analytics. We conclude with design implications for future EdTech systems that aim to integrate novel data modalities in classroom contexts.

Zhiyuan He, Yike Zhang, Chengruidong Zhang et al. · microsoft-research

Large Language Models (LLMs) incur quadratic attention complexity with input length, creating a major time bottleneck in the prefilling stage. Existing acceleration methods largely exploit attention score sparsity by estimating blocks with high attention scores and applying dynamic sparse attention. In this work, we identify another untapped form of sparsity in the prefilling stage, namely decoding-time contribution sparsity, where many attention blocks exhibit nontrivial attention scores during prefilling yet contribute negligibly to subsequent decoding, as indicated by gradient-based analysis. Building on this observation, we propose TriangleMix, a training-free static attention pattern that uses dense attention in a subset of layers and switches to Triangle attention in the others. Extensive experiments show that TriangleMix preserves nearly lossless performance relative to dense attention while substantially reducing attention overhead in Triangle layers. For 128K inputs, Triangle attention achieves a 15.3x speedup in attention computation, significantly exceeding the acceleration of typical dynamic sparse methods (1.9x to 3.4x). Furthermore, TriangleMix can be seamlessly combined with dynamic sparsity approaches, delivering an additional 6% to 19% reduction in TTFT over using dynamic sparsity alone.

Yike Zhang, Jack H. Noble

Cochlear implant surgery is a treatment for individuals with severe hearing loss. It involves inserting an array of electrodes inside the cochlea to electrically stimulate the auditory nerve and restore hearing sensation. A crucial step in this procedure is mastoidectomy, a surgical intervention that removes part of the mastoid region of the temporal bone, providing a critical pathway to the cochlea for electrode placement. Accurate prediction of the mastoidectomy region from preoperative imaging assists presurgical planning, reduces surgical risks, and improves surgical outcomes. In previous work, a self-supervised network was introduced to predict the mastoidectomy region using only preoperative CT scans. While promising, the method suffered from suboptimal robustness, limiting its practical application. To address this limitation, we propose a novel weakly-supervised Mamba-based framework to predict accurate mastoidectomy regions directly from preoperative CT scans. Our approach utilizes a 3D T-Distribution loss function inspired by the Student-t distribution, which effectively handles the complex geometric variability inherent in mastoidectomy shapes. Weak supervision is achieved using the segmentation results from the prior self-supervised network to eliminate the need for manual data cleaning or labeling throughout the training process. The proposed method is extensively evaluated against state-of-the-art approaches, demonstrating superior performance in predicting accurate and clinically relevant mastoidectomy regions. Our findings highlight the robustness and efficiency of the weakly-supervised learning framework with the proposed novel 3D T-Distribution loss.

Yike Zhang, Eduardo Davalos Anaya, Jack H. Noble

This paper presents a novel method for monocular patient-to-image intraoperative registration, specifically designed to operate without any external hardware tracking equipment or fiducial point markers. Leveraging a synthetic microscopy surgical scene dataset with a wide range of transformations, our approach directly maps preoperative CT scans to 2D intraoperative surgical frames through a lightweight neural network for real-time cochlear implant surgery guidance via a zero-shot learning approach. Unlike traditional methods, our framework seamlessly integrates with monocular surgical microscopes, making it highly practical for clinical use without additional hardware dependencies and requirements. Our method estimates camera poses, which include a rotation matrix and a translation vector, by learning from the synthetic dataset, enabling accurate and efficient intraoperative registration. The proposed framework was evaluated on nine clinical cases using a patient-specific and cross-patient validation strategy. Our results suggest that our approach achieves clinically relevant accuracy in predicting 6D camera poses for registering 3D preoperative CT scans to 2D surgical scenes with an angular error within 10 degrees in most cases, while also addressing limitations of traditional methods, such as reliance on external tracking systems or fiducial markers.

Eduardo Davalos, Yike Zhang, Ashwin T. S. et al.

This study presents a novel framework for 3D gaze tracking tailored for mixed-reality settings, aimed at enhancing joint attention and collaborative efforts in team-based scenarios. Conventional gaze tracking, often limited by monocular cameras and traditional eye-tracking apparatus, struggles with simultaneous data synchronization and analysis from multiple participants in group contexts. Our proposed framework leverages state-of-the-art computer vision and machine learning techniques to overcome these obstacles, enabling precise 3D gaze estimation without dependence on specialized hardware or complex data fusion. Utilizing facial recognition and deep learning, the framework achieves real-time, tracking of gaze patterns across several individuals, addressing common depth estimation errors, and ensuring spatial and identity consistency within the dataset. Empirical results demonstrate the accuracy and reliability of our method in group environments. This provides mechanisms for significant advances in behavior and interaction analysis in educational and professional training applications in dynamic and unstructured environments.

Keqi Deng, Songjun Cao, Yike Zhang et al.

Recently, end-to-end automatic speech recognition models based on connectionist temporal classification (CTC) have achieved impressive results, especially when fine-tuned from wav2vec2.0 models. Due to the conditional independence assumption, CTC-based models are always weaker than attention-based encoder-decoder models and require the assistance of external language models (LMs). To solve this issue, we propose two knowledge transferring methods that leverage pre-trained LMs, such as BERT and GPT2, to improve CTC-based models. The first method is based on representation learning, in which the CTC-based models use the representation produced by BERT as an auxiliary learning target. The second method is based on joint classification learning, which combines GPT2 for text modeling with a hybrid CTC/attention architecture. Experiment on AISHELL-1 corpus yields a character error rate (CER) of 4.2% on the test set. When compared to the vanilla CTC-based models fine-tuned from the wav2vec2.0 models, our knowledge transferring method reduces CER by 16.1% relatively without external LMs.

Kun Wei, Yike Zhang, Sining Sun et al.

Conversational automatic speech recognition (ASR) is a task to recognize conversational speech including multiple speakers. Unlike sentence-level ASR, conversational ASR can naturally take advantages from specific characteristics of conversation, such as role preference and topical coherence. This paper proposes a conversational ASR model which explicitly learns conversation-level characteristics under the prevalent end-to-end neural framework. The highlights of the proposed model are twofold. First, a latent variational module (LVM) is attached to a conformer-based encoder-decoder ASR backbone to learn role preference and topical coherence. Second, a topic model is specifically adopted to bias the outputs of the decoder to words in the predicted topics. Experiments on two Mandarin conversational ASR tasks show that the proposed model achieves a maximum 12% relative character error rate (CER) reduction.

Keqi Deng, Songjun Cao, Yike Zhang et al.

Recently, self-supervised pretraining has achieved impressive results in end-to-end (E2E) automatic speech recognition (ASR). However, the dominant sequence-to-sequence (S2S) E2E model is still hard to fully utilize the self-supervised pre-training methods because its decoder is conditioned on acoustic representation thus cannot be pretrained separately. In this paper, we propose a pretrained Transformer (Preformer) S2S ASR architecture based on hybrid CTC/attention E2E models to fully utilize the pretrained acoustic models (AMs) and language models (LMs). In our framework, the encoder is initialized with a pretrained AM (wav2vec2.0). The Preformer leverages CTC as an auxiliary task during training and inference. Furthermore, we design a one-cross decoder (OCD), which relaxes the dependence on acoustic representations so that it can be initialized with pretrained LM (DistilGPT2). Experiments are conducted on the AISHELL-1 corpus and achieve a $4.6\%$ character error rate (CER) on the test set. Compared with our vanilla hybrid CTC/attention Transformer baseline, our proposed CTC/attention-based Preformer yields $27\%$ relative CER reduction. To the best of our knowledge, this is the first work to utilize both pretrained AM and LM in a S2S ASR system.

Songjun Cao, Yueteng Kang, Yanzhe Fu et al.

Recently self-supervised learning has emerged as an effective approach to improve the performance of automatic speech recognition (ASR). Under such a framework, the neural network is usually pre-trained with massive unlabeled data and then fine-tuned with limited labeled data. However, the non-streaming architecture like bidirectional transformer is usually adopted by the neural network to achieve competitive results, which can not be used in streaming scenarios. In this paper, we mainly focus on improving the performance of streaming transformer under the self-supervised learning framework. Specifically, we propose a novel two-stage training method during fine-tuning, which combines knowledge distilling and self-training. The proposed training method achieves 16.3% relative word error rate (WER) reduction on Librispeech noisy test set. Finally, by only using the 100h clean subset of Librispeech as the labeled data and the rest (860h) as the unlabeled data, our streaming transformer based model obtains competitive WERs 3.5/8.7 on Librispeech clean/noisy test sets.

Songjun Cao, Yike Zhang, Xiaobing Feng et al.

Nowadays voice search for points of interest (POI) is becoming increasingly popular. However, speech recognition for local POI has remained to be a challenge due to multi-dialect and massive POI. This paper improves speech recognition accuracy for local POI from two aspects. Firstly, a geographic acoustic model (Geo-AM) is proposed. The Geo-AM deals with multi-dialect problem using dialect-specific input feature and dialect-specific top layer. Secondly, a group of geo-specific language models (Geo-LMs) are integrated into our speech recognition system to improve recognition accuracy of long tail and homophone POI. During decoding, specific language models are selected on demand according to users' geographic location. Experiments show that the proposed Geo-AM achieves 6.5%$\sim$10.1% relative character error rate (CER) reduction on an accent testset and the proposed Geo-AM and Geo-LM totally achieve over 18.7% relative CER reduction on Tencent Map task.