In this paper, we propose the Continuous Time Fractional Topic Model (cFTM), a new method for dynamic topic modeling. This approach incorporates fractional Brownian motion~(fBm) to effectively identify positive or negative correlations in topic and word distribution over time, revealing long-term dependency or roughness. Our theoretical analysis shows that the cFTM can capture these long-term dependency or roughness in both topic and word distributions, mirroring the main characteristics of fBm. Moreover, we prove that the parameter estimation process for the cFTM is on par with that of LDA, traditional topic models. To demonstrate the cFTM's property, we conduct empirical study using economic news articles. The results from these tests support the model's ability to identify and track long-term dependency or roughness in topics over time.
Recently, there has been growing evidence that if the width and depth of a neural network are scaled toward the so-called rich feature learning limit ($\mu$P and its depth extension), then some hyperparameters - such as the learning rate - exhibit transfer from small to very large models, thus reducing the cost of hyperparameter tuning. From an optimization perspective, this phenomenon is puzzling, as it implies that the loss landscape is remarkably consistent across very different model sizes. In this work, we find empirical evidence that learning rate transfer can be attributed to the fact that under $\mu$P and its depth extension, the largest eigenvalue of the training loss Hessian (i.e. the sharpness) is largely independent of the width and depth of the network for a sustained period of training time. On the other hand, we show that under the neural tangent kernel (NTK) regime, the sharpness exhibits very different dynamics at different scales, thus preventing learning rate transfer. But what causes these differences in the sharpness dynamics? Through a connection between the spectra of the Hessian and the NTK matrix, we argue that the cause lies in the presence (for $\mu$P) or progressive absence (for the NTK regime) of feature learning, which results in a different evolution of the NTK, and thus of the sharpness. We corroborate our claims with a substantial suite of experiments, covering a wide range of datasets and architectures: from ResNets and Vision Transformers trained on benchmark vision datasets to Transformers-based language models trained on WikiText
With the explosive growth of medical data and the rapid development of artificial intelligence technology, precision medicine has emerged as a key to enhancing the quality and efficiency of healthcare services. In this context, Large Language Models (LLMs) play an increasingly vital role in medical knowledge acquisition and question-answering systems. To further improve the performance of these systems in the medical domain, we introduce an innovative method that jointly trains an Information Retrieval (IR) system and an LLM during the fine-tuning phase. This approach, which we call Joint Medical LLM and Retrieval Training (JMLR), is designed to overcome the challenges faced by traditional models in handling medical question-answering tasks. By employing a synchronized training mechanism, JMLR reduces the demand for computational resources and enhances the model's ability to leverage medical knowledge for reasoning and answering questions. Our experimental results demonstrate that JMLR-13B (81.2% on Amboos, 61.3% on MedQA) outperforms models using conventional pre-training and fine-tuning Meditron-70B (76.4% on AMBOSS, 60.3% on MedQA). For models of the same 7B scale, JMLR-7B(68.7% on Amboos, 51.7% on MedQA) significantly outperforms other public models (Meditron-7B: 50.1%, 47.9%), proving its superiority in terms of cost (our training time: 37 hours, traditional method: 144 hours), efficiency, and effectiveness in medical question-answering tasks. Through this work, we provide a new and efficient knowledge enhancement tool for healthcare, demonstrating the great potential of integrating IR and LLM training in precision medical information retrieval and question-answering systems.
3D modeling holds significant importance in the realms of AR/VR and gaming, allowing for both artistic creativity and practical applications. However, the process is often time-consuming and demands a high level of skill. In this paper, we present a novel approach to create volumetric representations of 3D characters from consistent turnaround concept art, which serves as the standard input in the 3D modeling industry. While Neural Radiance Field (NeRF) has been a game-changer in image-based 3D reconstruction, to the best of our knowledge, there is no known research that optimizes the pipeline for concept art. To harness the potential of concept art, with its defined body poses and specific view angles, we propose encoding it as priors for our model. We train the network to make use of these priors for various 3D points through a learnable view-direction-attended multi-head self-attention layer. Additionally, we demonstrate that a combination of ray sampling and surface sampling enhances the inference capabilities of our network. Our model is able to generate high-quality 360-degree views of characters. Subsequently, we provide a simple guideline to better leverage our model to extract the 3D mesh. It is important to note that our model's inferencing capabilities are influenced by the training data's characteristics, primarily focusing on characters with a single head, two arms, and two legs. Nevertheless, our methodology remains versatile and adaptable to concept art from diverse subject matters, without imposing any specific assumptions on the data.
3D Object Detectors (3D-OD) are crucial for understanding the environment in many robotic tasks, especially autonomous driving. Including 3D information via Lidar sensors improves accuracy greatly. However, such detectors perform poorly on domains they were not trained on, i.e. different locations, sensors, weather, etc., limiting their reliability in safety-critical applications. There exist methods to adapt 3D-ODs to these domains; however, these methods treat 3D-ODs as a black box, neglecting underlying architectural decisions and source-domain training strategies. Instead, we dive deep into the details of 3D-ODs, focusing our efforts on fundamental factors that influence robustness prior to domain adaptation. We systematically investigate four design choices (and the interplay between them) often overlooked in 3D-OD robustness and domain adaptation: architecture, voxel encoding, data augmentations, and anchor strategies. We assess their impact on the robustness of nine state-of-the-art 3D-ODs across six benchmarks encompassing three types of domain gaps - sensor type, weather, and location. Our main findings are: (1) transformer backbones with local point features are more robust than 3D CNNs, (2) test-time anchor size adjustment is crucial for adaptation across geographical locations, significantly boosting scores without retraining, (3) source-domain augmentations allow the model to generalize to low-resolution sensors, and (4) surprisingly, robustness to bad weather is improved when training directly on more clean weather data than on training with bad weather data. We outline our main conclusions and findings to provide practical guidance on developing more robust 3D-ODs.
Reinforcement learning has witnessed significant advancements, particularly with the emergence of model-based approaches. Among these, $Q$-learning has proven to be a powerful algorithm in model-free settings. However, the extension of $Q$-learning to a model-based framework remains relatively unexplored. In this paper, we delve into the sample complexity of $Q$-learning when integrated with a model-based approach. Through theoretical analyses and empirical evaluations, we seek to elucidate the conditions under which model-based $Q$-learning excels in terms of sample efficiency compared to its model-free counterpart.
With recent advances in computer vision, it appears that autonomous driving will be part of modern society sooner rather than later. However, there are still a significant number of concerns to address. Although modern computer vision techniques demonstrate superior performance, they tend to prioritize accuracy over efficiency, which is a crucial aspect of real-time applications. Large object detection models typically require higher computational power, which is achieved by using more sophisticated onboard hardware. For autonomous driving, these requirements translate to increased fuel costs and, ultimately, a reduction in mileage. Further, despite their computational demands, the existing object detectors are far from being real-time. In this research, we assess the robustness of our previously proposed, highly efficient pedestrian detector LSFM on well-established autonomous driving benchmarks, including diverse weather conditions and nighttime scenes. Moreover, we extend our LSFM model for general object detection to achieve real-time object detection in traffic scenes. We evaluate its performance, low latency, and generalizability on traffic object detection datasets. Furthermore, we discuss the inadequacy of the current key performance indicator employed by object detection systems in the context of autonomous driving and propose a more suitable alternative that incorporates real-time requirements.
We study the problem of symbolic music generation (e.g., generating piano rolls), with a technical focus on non-differentiable rule guidance. Musical rules are often expressed in symbolic form on note characteristics, such as note density or chord progression, many of which are non-differentiable which pose a challenge when using them for guided diffusion. We propose Stochastic Control Guidance (SCG), a novel guidance method that only requires forward evaluation of rule functions that can work with pre-trained diffusion models in a plug-and-play way, thus achieving training-free guidance for non-differentiable rules for the first time. Additionally, we introduce a latent diffusion architecture for symbolic music generation with high time resolution, which can be composed with SCG in a plug-and-play fashion. Compared to standard strong baselines in symbolic music generation, this framework demonstrates marked advancements in music quality and rule-based controllability, outperforming current state-of-the-art generators in a variety of settings. For detailed demonstrations, code and model checkpoints, please visit our project website: https://scg-rule-guided-music.github.io/.
Reference features from a template or historical frames are crucial for visual object tracking. Prior works utilize all features from a fixed template or memory for visual object tracking. However, due to the dynamic nature of videos, the required reference historical information for different search regions at different time steps is also inconsistent. Therefore, using all features in the template and memory can lead to redundancy and impair tracking performance. To alleviate this issue, we propose a novel tracking paradigm, consisting of a relevance attention mechanism and a global representation memory, which can adaptively assist the search region in selecting the most relevant historical information from reference features. Specifically, the proposed relevance attention mechanism in this work differs from previous approaches in that it can dynamically choose and build the optimal global representation memory for the current frame by accessing cross-frame information globally. Moreover, it can flexibly read the relevant historical information from the constructed memory to reduce redundancy and counteract the negative effects of harmful information. Extensive experiments validate the effectiveness of the proposed method, achieving competitive performance on five challenging datasets with 71 FPS.
Recent recommender systems started to use rating elicitation, which asks new users to rate a small seed itemset for inferring their preferences, to improve the quality of initial recommendations. The key challenge of the rating elicitation is to choose the seed items which can best infer the new users' preference. This paper proposes a novel end-to-end Deep learning framework for Rating Elicitation (DRE), that chooses all the seed items at a time with consideration of the non-linear interactions. To this end, it first defines categorical distributions to sample seed items from the entire itemset, then it trains both the categorical distributions and a neural reconstruction network to infer users' preferences on the remaining items from CF information of the sampled seed items. Through the end-to-end training, the categorical distributions are learned to select the most representative seed items while reflecting the complex non-linear interactions. Experimental results show that DRE outperforms the state-of-the-art approaches in the recommendation quality by accurately inferring the new users' preferences and its seed itemset better represents the latent space than the seed itemset obtained by the other methods.