Unified Autoregressive Visual Generation and Understanding with Continuous Tokens

We present UniFluid, a unified autoregressive framework for joint visual generation and understanding leveraging continuous visual tokens. Our unified autoregressive architecture processes multimodal image and text inputs, generating discrete tokens for text and continuous tokens for image. We find though there is an inherent trade-off between the image generation and understanding task, a carefully tuned training recipe enables them to improve each other. By selecting an appropriate loss balance weight, the unified model achieves results comparable to or exceeding those of single-task baselines on both tasks. Furthermore, we demonstrate that employing stronger pre-trained LLMs and random-order generation during training is important to achieve high-fidelity image generation within this unified framework. Built upon the Gemma model series, UniFluid exhibits competitive performance across both image generation and understanding, demonstrating strong transferability to various downstream tasks, including image editing for generation, as well as visual captioning and question answering for understanding.

HAR-DoReMi: Optimizing Data Mixture for Self-Supervised Human Activity Recognition Across Heterogeneous IMU Datasets

Cross-dataset Human Activity Recognition (HAR) suffers from limited model generalization, hindering its practical deployment. To address this critical challenge, inspired by the success of DoReMi in Large Language Models (LLMs), we introduce a data mixture optimization strategy for pre-training HAR models, aiming to improve the recognition performance across heterogeneous datasets. However, directly applying DoReMi to the HAR field encounters new challenges due to the continuous, multi-channel and intrinsic heterogeneous characteristics of IMU sensor data. To overcome these limitations, we propose a novel framework HAR-DoReMi, which introduces a masked reconstruction task based on Mean Squared Error (MSE) loss. By raplacing the discrete language sequence prediction task, which relies on the Negative Log-Likelihood (NLL) loss, in the original DoReMi framework, the proposed framework is inherently more appropriate for handling the continuous and multi-channel characteristics of IMU data. In addition, HAR-DoReMi integrates the Mahony fusion algorithm into the self-supervised HAR pre-training, aiming to mitigate the heterogeneity of varying sensor orientation. This is achieved by estimating the sensor orientation within each dataset and facilitating alignment with a unified coordinate system, thereby improving the cross-dataset generalization ability of the HAR model. Experimental evaluation on multiple cross-dataset HAR transfer tasks demonstrates that HAR-DoReMi improves the accuracy by an average of 6.51%, compared to the current state-of-the-art method with only approximately 30% to 50% of the data usage. These results confirm the effectiveness of HAR-DoReMi in improving the generalization and data efficiency of pre-training HAR models, underscoring its significant potential to facilitate the practical deployment of HAR technology.

PTMs-TSCIL Pre-Trained Models Based Class-Incremental Learning

Class-incremental learning (CIL) for time series data faces critical challenges in balancing stability against catastrophic forgetting and plasticity for new knowledge acquisition, particularly under real-world constraints where historical data access is restricted. While pre-trained models (PTMs) have shown promise in CIL for vision and NLP domains, their potential in time series class-incremental learning (TSCIL) remains underexplored due to the scarcity of large-scale time series pre-trained models. Prompted by the recent emergence of large-scale pre-trained models (PTMs) for time series data, we present the first exploration of PTM-based Time Series Class-Incremental Learning (TSCIL). Our approach leverages frozen PTM backbones coupled with incrementally tuning the shared adapter, preserving generalization capabilities while mitigating feature drift through knowledge distillation. Furthermore, we introduce a Feature Drift Compensation Network (DCN), designed with a novel two-stage training strategy to precisely model feature space transformations across incremental tasks. This allows for accurate projection of old class prototypes into the new feature space. By employing DCN-corrected prototypes, we effectively enhance the unified classifier retraining, mitigating model feature drift and alleviating catastrophic forgetting. Extensive experiments on five real-world datasets demonstrate state-of-the-art performance, with our method yielding final accuracy gains of 1.4%-6.1% across all datasets compared to existing PTM-based approaches. Our work establishes a new paradigm for TSCIL, providing insights into stability-plasticity optimization for continual learning systems.

ReLU-KAN: New Kolmogorov-Arnold Networks that Only Need Matrix Addition, Dot Multiplication, and ReLU

Limited by the complexity of basis function (B-spline) calculations, Kolmogorov-Arnold Networks (KAN) suffer from restricted parallel computing capability on GPUs. This paper proposes a novel ReLU-KAN implementation that inherits the core idea of KAN. By adopting ReLU (Rectified Linear Unit) and point-wise multiplication, we simplify the design of KAN's basis function and optimize the computation process for efficient CUDA computing. The proposed ReLU-KAN architecture can be readily implemented on existing deep learning frameworks (e.g., PyTorch) for both inference and training. Experimental results demonstrate that ReLU-KAN achieves a 20x speedup compared to traditional KAN with 4-layer networks. Furthermore, ReLU-KAN exhibits a more stable training process with superior fitting ability while preserving the "catastrophic forgetting avoidance" property of KAN. You can get the code in https://github.com/quiqi/relu_kan

HARMamba: Efficient Wearable Sensor Human Activity Recognition Based on Bidirectional Selective SSM

Wearable sensor human activity recognition (HAR) is a crucial area of research in activity sensing. While transformer-based temporal deep learning models have been extensively studied and implemented, their large number of parameters present significant challenges in terms of system computing load and memory usage, rendering them unsuitable for real-time mobile activity recognition applications. Recently, an efficient hardware-aware state space model (SSM) called Mamba has emerged as a promising alternative. Mamba demonstrates strong potential in long sequence modeling, boasts a simpler network architecture, and offers an efficient hardware-aware design. Leveraging SSM for activity recognition represents an appealing avenue for exploration. In this study, we introduce HARMamba, which employs a more lightweight selective SSM as the foundational model architecture for activity recognition. The goal is to address the computational resource constraints encountered in real-time activity recognition scenarios. Our approach involves processing sensor data flow by independently learning each channel and segmenting the data into "patches". The marked sensor sequence's position embedding serves as the input token for the bidirectional state space model, ultimately leading to activity categorization through the classification head. Compared to established activity recognition frameworks like Transformer-based models, HARMamba achieves superior performance while also reducing computational and memory overhead. Furthermore, our proposed method has been extensively tested on four public activity datasets: PAMAP2, WISDM, UNIMIB, and UCI, demonstrating impressive performance in activity recognition tasks.

P2LHAP:Wearable sensor-based human activity recognition, segmentation and forecast through Patch-to-Label Seq2Seq Transformer

Traditional deep learning methods struggle to simultaneously segment, recognize, and forecast human activities from sensor data. This limits their usefulness in many fields such as healthcare and assisted living, where real-time understanding of ongoing and upcoming activities is crucial. This paper introduces P2LHAP, a novel Patch-to-Label Seq2Seq framework that tackles all three tasks in a efficient single-task model. P2LHAP divides sensor data streams into a sequence of "patches", served as input tokens, and outputs a sequence of patch-level activity labels including the predicted future activities. A unique smoothing technique based on surrounding patch labels, is proposed to identify activity boundaries accurately. Additionally, P2LHAP learns patch-level representation by sensor signal channel-independent Transformer encoders and decoders. All channels share embedding and Transformer weights across all sequences. Evaluated on three public datasets, P2LHAP significantly outperforms the state-of-the-art in all three tasks, demonstrating its effectiveness and potential for real-world applications.

Gemini: A Family of Highly Capable Multimodal Models

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of Gemini models in cross-modal reasoning and language understanding will enable a wide variety of use cases and we discuss our approach toward deploying them responsibly to users.

IG Captioner: Information Gain Captioners are Strong Zero-shot Classifiers

Generative training has been demonstrated to be powerful for building visual-language models. However, on zero-shot discriminative benchmarks, there is still a performance gap between models trained with generative and discriminative objectives. In this paper, we aim to narrow this gap by improving the efficacy of generative training on classification tasks, without any finetuning processes or additional modules. Specifically, we focus on narrowing the gap between the generative captioner and the CLIP classifier. We begin by analysing the predictions made by the captioner and classifier and observe that the caption generation inherits the distribution bias from the language model trained with pure text modality, making it less grounded on the visual signal. To tackle this problem, we redesign the scoring objective for the captioner to alleviate the distributional bias and focus on measuring the gain of information brought by the visual inputs. We further design a generative training objective to match the evaluation objective. We name our model trained and evaluated from the novel procedures as Information Gain (IG) captioner. We pretrain the models on the public Laion-5B dataset and perform a series of discriminative evaluations. For the zero-shot classification on ImageNet, IG captioner achieves $> 18\%$ improvements over the standard captioner, achieving comparable performances with the CLIP classifier. IG captioner also demonstrated strong performance on zero-shot image-text retrieval tasks on MSCOCO and Flickr30K. We hope this paper inspires further research towards unifying generative and discriminative training procedures for visual-language models.

A Multi-Task Deep Learning Approach for Sensor-based Human Activity Recognition and Segmentation

Sensor-based human activity segmentation and recognition are two important and challenging problems in many real-world applications and they have drawn increasing attention from the deep learning community in recent years. Most of the existing deep learning works were designed based on pre-segmented sensor streams and they have treated activity segmentation and recognition as two separate tasks. In practice, performing data stream segmentation is very challenging. We believe that both activity segmentation and recognition may convey unique information which can complement each other to improve the performance of the two tasks. In this paper, we firstly proposes a new multitask deep neural network to solve the two tasks simultaneously. The proposed neural network adopts selective convolution and features multiscale windows to segment activities of long or short time durations. First, multiple windows of different scales are generated to center on each unit of the feature sequence. Then, the model is trained to predict, for each window, the activity class and the offset to the true activity boundaries. Finally, overlapping windows are filtered out by non-maximum suppression, and adjacent windows of the same activity are concatenated to complete the segmentation task. Extensive experiments were conducted on eight popular benchmarking datasets, and the results show that our proposed method outperforms the state-of-the-art methods both for activity recognition and segmentation.

TencentPretrain: A Scalable and Flexible Toolkit for Pre-training Models of Different Modalities

Recently, the success of pre-training in text domain has been fully extended to vision, audio, and cross-modal scenarios. The proposed pre-training models of different modalities are showing a rising trend of homogeneity in their model structures, which brings the opportunity to implement different pre-training models within a uniform framework. In this paper, we present TencentPretrain, a toolkit supporting pre-training models of different modalities. The core feature of TencentPretrain is the modular design. The toolkit uniformly divides pre-training models into 5 components: embedding, encoder, target embedding, decoder, and target. As almost all of common modules are provided in each component, users can choose the desired modules from different components to build a complete pre-training model. The modular design enables users to efficiently reproduce existing pre-training models or build brand-new one. We test the toolkit on text, vision, and audio benchmarks and show that it can match the performance of the original implementations.