Valley2: Exploring Multimodal Models with Scalable Vision-Language Design

Recently, vision-language models have made remarkable progress, demonstrating outstanding capabilities in various tasks such as image captioning and video understanding. We introduce Valley2, a novel multimodal large language model designed to enhance performance across all domains and extend the boundaries of practical applications in e-commerce and short video scenarios. Notably, Valley2 achieves state-of-the-art (SOTA) performance on e-commerce benchmarks, surpassing open-source models of similar size by a large margin (79.66 vs. 72.76). Additionally, Valley2 ranks second on the OpenCompass leaderboard among models with fewer than 10B parameters, with an impressive average score of 67.4. The code and model weights are open-sourced at https://github.com/bytedance/Valley.

Reverse Modeling in Large Language Models

Humans are accustomed to reading and writing in a forward manner, and this natural bias extends to text understanding in auto-regressive large language models (LLMs). This paper investigates whether LLMs, like humans, struggle with reverse modeling, specifically with reversed text inputs. We found that publicly available pre-trained LLMs cannot understand such inputs. However, LLMs trained from scratch with both forward and reverse texts can understand them equally well during inference. Our case study shows that different-content texts result in different losses if input (to LLMs) in different directions -- some get lower losses for forward while some for reverse. This leads us to a simple and nice solution for data selection based on the loss differences between forward and reverse directions. Using our selected data in continued pretraining can boost LLMs' performance by a large margin across different language understanding benchmarks.

Common and Rare Fundus Diseases Identification Using Vision-Language Foundation Model with Knowledge of Over 400 Diseases

The current retinal artificial intelligence models were trained using data with a limited category of diseases and limited knowledge. In this paper, we present a retinal vision-language foundation model (RetiZero) with knowledge of over 400 fundus diseases. Specifically, we collected 341,896 fundus images paired with text descriptions from 29 publicly available datasets, 180 ophthalmic books, and online resources, encompassing over 400 fundus diseases across multiple countries and ethnicities. RetiZero achieved outstanding performance across various downstream tasks, including zero-shot retinal disease recognition, image-to-image retrieval, internal domain and cross-domain retinal disease classification, and few-shot fine-tuning. Specially, in the zero-shot scenario, RetiZero achieved a Top5 score of 0.8430 and 0.7561 on 15 and 52 fundus diseases respectively. In the image-retrieval task, RetiZero achieved a Top5 score of 0.9500 and 0.8860 on 15 and 52 retinal diseases respectively. Furthermore, clinical evaluations by ophthalmology experts from different countries demonstrate that RetiZero can achieve performance comparable to experienced ophthalmologists using zero-shot and image retrieval methods without requiring model retraining. These capabilities of retinal disease identification strengthen our RetiZero foundation model in clinical implementation.

VoCoT: Unleashing Visually Grounded Multi-Step Reasoning in Large Multi-Modal Models

While large multi-modal models (LMMs) have exhibited impressive capabilities across diverse tasks, their effectiveness in handling complex tasks has been limited by the prevailing single-step reasoning paradigm. To this end, this paper proposes VoCoT, a multi-step Visually grounded object-centric Chain-of-Thought reasoning framework tailored for inference with LMMs. VoCoT is characterized by two key features: (1) object-centric reasoning paths that revolve around cross-modal shared object-level information, and (2) visually grounded representation of object concepts in a multi-modal interleaved and aligned manner, which effectively bridges the modality gap within LMMs during long-term generation. Additionally, we construct an instruction dataset to facilitate LMMs in adapting to reasoning with VoCoT. By introducing VoCoT into the prevalent open-source LMM architecture, we introduce VolCano. With only 7B parameters and limited input resolution, VolCano demonstrates excellent performance across various scenarios, surpassing SOTA models, including GPT-4V, in tasks requiring complex reasoning. Our code, data and model will be available at https://github.com/RupertLuo/VoCoT.

CODA: A COst-efficient Test-time Domain Adaptation Mechanism for HAR

In recent years, emerging research on mobile sensing has led to novel scenarios that enhance daily life for humans, but dynamic usage conditions often result in performance degradation when systems are deployed in real-world settings. Existing solutions typically employ one-off adaptation schemes based on neural networks, which struggle to ensure robustness against uncertain drifting conditions in human-centric sensing scenarios. In this paper, we propose CODA, a COst-efficient Domain Adaptation mechanism for mobile sensing that addresses real-time drifts from the data distribution perspective with active learning theory, ensuring cost-efficient adaptation directly on the device. By incorporating a clustering loss and importance-weighted active learning algorithm, CODA retains the relationship between different clusters during cost-effective instance-level updates, preserving meaningful structure within the data distribution. We also showcase its generalization by seamlessly integrating it with Neural Network-based solutions for Human Activity Recognition tasks. Through meticulous evaluations across diverse datasets, including phone-based, watch-based, and integrated sensor-based sensing tasks, we demonstrate the feasibility and potential of online adaptation with CODA. The promising results achieved by CODA, even without learnable parameters, also suggest the possibility of realizing unobtrusive adaptation through specific application designs with sufficient feedback.

Learning Knowledge-Enhanced Contextual Language Representations for Domain Natural Language Understanding

Knowledge-Enhanced Pre-trained Language Models (KEPLMs) improve the performance of various downstream NLP tasks by injecting knowledge facts from large-scale Knowledge Graphs (KGs). However, existing methods for pre-training KEPLMs with relational triples are difficult to be adapted to close domains due to the lack of sufficient domain graph semantics. In this paper, we propose a Knowledge-enhanced lANGuAge Representation learning framework for various clOsed dOmains (KANGAROO) via capturing the implicit graph structure among the entities. Specifically, since the entity coverage rates of closed-domain KGs can be relatively low and may exhibit the global sparsity phenomenon for knowledge injection, we consider not only the shallow relational representations of triples but also the hyperbolic embeddings of deep hierarchical entity-class structures for effective knowledge fusion.Moreover, as two closed-domain entities under the same entity-class often have locally dense neighbor subgraphs counted by max point biconnected component, we further propose a data augmentation strategy based on contrastive learning over subgraphs to construct hard negative samples of higher quality. It makes the underlying KELPMs better distinguish the semantics of these neighboring entities to further complement the global semantic sparsity. In the experiments, we evaluate KANGAROO over various knowledge-aware and general NLP tasks in both full and few-shot learning settings, outperforming various KEPLM training paradigms performance in closed-domains significantly.

FashionLOGO: Prompting Multimodal Large Language Models for Fashion Logo Embeddings

Logo embedding plays a crucial role in various e-commerce applications by facilitating image retrieval or recognition, such as intellectual property protection and product search. However, current methods treat logo embedding as a purely visual problem, which may limit their performance in real-world scenarios. A notable issue is that the textual knowledge embedded in logo images has not been adequately explored. Therefore, we propose a novel approach that leverages textual knowledge as an auxiliary to improve the robustness of logo embedding. The emerging Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in both visual and textual understanding and could become valuable visual assistants in understanding logo images. Inspired by this observation, our proposed method, FashionLOGO, aims to utilize MLLMs to enhance fashion logo embedding. We explore how MLLMs can improve logo embedding by prompting them to generate explicit textual knowledge through three types of prompts, including image OCR, brief captions, and detailed descriptions prompts, in a zero-shot setting. We adopt a cross-attention transformer to enable image embedding queries to learn supplementary knowledge from textual embeddings automatically. To reduce computational costs, we only use the image embedding model in the inference stage, similar to traditional inference pipelines. Our extensive experiments on three real-world datasets demonstrate that FashionLOGO learns generalized and robust logo embeddings, achieving state-of-the-art performance in all benchmark datasets. Furthermore, we conduct comprehensive ablation studies to demonstrate the performance improvements resulting from the introduction of MLLMs.

Valley: Video Assistant with Large Language model Enhanced abilitY

Recently, several multi-modal models have been developed for joint image and language understanding, which have demonstrated impressive chat abilities by utilizing advanced large language models (LLMs). The process of developing such models is straightforward yet effective. It involves pre-training an adaptation module to align the semantics of the vision encoder and language model, followed by fine-tuning on the instruction-following data. However, despite the success of this pipeline in image and language understanding, its effectiveness in joint video and language understanding has not been widely explored. In this paper, we aim to develop a novel multi-modal foundation model capable of perceiving video, image, and language within a general framework. To achieve this goal, we introduce Valley: Video Assistant with Large Language model Enhanced ability. Specifically, our proposed Valley model is designed with a simple projection module that bridges video, image, and language modalities, and is further unified with a multi-lingual LLM. We also collect multi-source vision-text pairs and adopt a spatio-temporal pooling strategy to obtain a unified vision encoding of video and image input for pre-training. Furthermore, we generate multi-task instruction-following video data, including multi-shot captions, long video descriptions, action recognition, causal relationship inference, etc. To obtain the instruction-following data, we design diverse rounds of task-oriented conversations between humans and videos, facilitated by ChatGPT. Qualitative examples demonstrate that our proposed model has the potential to function as a highly effective multilingual video assistant that can make complex video understanding scenarios easy. Code, data, and models will be available at https://github.com/RupertLuo/Valley.

Meta-Learning Siamese Network for Few-Shot Text Classification

Few-shot learning has been used to tackle the problem of label scarcity in text classification, of which meta-learning based methods have shown to be effective, such as the prototypical networks (PROTO). Despite the success of PROTO, there still exist three main problems: (1) ignore the randomness of the sampled support sets when computing prototype vectors; (2) disregard the importance of labeled samples; (3) construct meta-tasks in a purely random manner. In this paper, we propose a Meta-Learning Siamese Network, namely, Meta-SN, to address these issues. Specifically, instead of computing prototype vectors from the sampled support sets, Meta-SN utilizes external knowledge (e.g. class names and descriptive texts) for class labels, which is encoded as the low-dimensional embeddings of prototype vectors. In addition, Meta-SN presents a novel sampling strategy for constructing meta-tasks, which gives higher sampling probabilities to hard-to-classify samples. Extensive experiments are conducted on six benchmark datasets to show the clear superiority of Meta-SN over other state-of-the-art models. For reproducibility, all the datasets and codes are provided at https://github.com/hccngu/Meta-SN.

SpanProto: A Two-stage Span-based Prototypical Network for Few-shot Named Entity Recognition

Few-shot Named Entity Recognition (NER) aims to identify named entities with very little annotated data. Previous methods solve this problem based on token-wise classification, which ignores the information of entity boundaries, and inevitably the performance is affected by the massive non-entity tokens. To this end, we propose a seminal span-based prototypical network (SpanProto) that tackles few-shot NER via a two-stage approach, including span extraction and mention classification. In the span extraction stage, we transform the sequential tags into a global boundary matrix, enabling the model to focus on the explicit boundary information. For mention classification, we leverage prototypical learning to capture the semantic representations for each labeled span and make the model better adapt to novel-class entities. To further improve the model performance, we split out the false positives generated by the span extractor but not labeled in the current episode set, and then present a margin-based loss to separate them from each prototype region. Experiments over multiple benchmarks demonstrate that our model outperforms strong baselines by a large margin.