Amir Zadeh

Paul Pu Liang, Amir Zadeh, Louis-Philippe Morency · cmu

Multimodal machine learning is a vibrant multi-disciplinary research field that aims to design computer agents with intelligent capabilities such as understanding, reasoning, and learning through integrating multiple communicative modalities, including linguistic, acoustic, visual, tactile, and physiological messages. With the recent interest in video understanding, embodied autonomous agents, text-to-image generation, and multisensor fusion in application domains such as healthcare and robotics, multimodal machine learning has brought unique computational and theoretical challenges to the machine learning community given the heterogeneity of data sources and the interconnections often found between modalities. However, the breadth of progress in multimodal research has made it difficult to identify the common themes and open questions in the field. By synthesizing a broad range of application domains and theoretical frameworks from both historical and recent perspectives, this paper is designed to provide an overview of the computational and theoretical foundations of multimodal machine learning. We start by defining three key principles of modality heterogeneity, connections, and interactions that have driven subsequent innovations, and propose a taxonomy of six core technical challenges: representation, alignment, reasoning, generation, transference, and quantification covering historical and recent trends. Recent technical achievements will be presented through the lens of this taxonomy, allowing researchers to understand the similarities and differences across new approaches. We end by motivating several open problems for future research as identified by our taxonomy.

Yingting Li, Ambuj Mehrish, Shuai Zhao et al. · cmu

Fine-tuning is widely used as the default algorithm for transfer learning from pre-trained models. Parameter inefficiency can however arise when, during transfer learning, all the parameters of a large pre-trained model need to be updated for individual downstream tasks. As the number of parameters grows, fine-tuning is prone to overfitting and catastrophic forgetting. In addition, full fine-tuning can become prohibitively expensive when the model is used for many tasks. To mitigate this issue, parameter-efficient transfer learning algorithms, such as adapters and prefix tuning, have been proposed as a way to introduce a few trainable parameters that can be plugged into large pre-trained language models such as BERT, and HuBERT. In this paper, we introduce the Speech UndeRstanding Evaluation (SURE) benchmark for parameter-efficient learning for various speech-processing tasks. Additionally, we introduce a new adapter, ConvAdapter, based on 1D convolution. We show that ConvAdapter outperforms the standard adapters while showing comparable performance against prefix tuning and LoRA with only 0.94% of trainable parameters on some of the task in SURE. We further explore the effectiveness of parameter efficient transfer learning for speech synthesis task such as Text-to-Speech (TTS).

Alex Wilf, Martin Q. Ma, Paul Pu Liang et al. · cmu

Creating artificial social intelligence - algorithms that can understand the nuances of multi-person interactions - is an exciting and emerging challenge in processing facial expressions and gestures from multimodal videos. Recent multimodal methods have set the state of the art on many tasks, but have difficulty modeling the complex face-to-face conversational dynamics across speaking turns in social interaction, particularly in a self-supervised setup. In this paper, we propose Face-to-Face Contrastive Learning (F2F-CL), a graph neural network designed to model social interactions using factorization nodes to contextualize the multimodal face-to-face interaction along the boundaries of the speaking turn. With the F2F-CL model, we propose to perform contrastive learning between the factorization nodes of different speaking turns within the same video. We experimentally evaluated the challenging Social-IQ dataset and show state-of-the-art results.

Dushyant Singh Chauhan, Gopendra Vikram Singh, Navonil Majumder et al.

Humor recognition in conversations is a challenging task that has recently gained popularity due to its importance in dialogue understanding, including in multimodal settings (i.e., text, acoustics, and visual). The few existing datasets for humor are mostly in English. However, due to the tremendous growth in multilingual content, there is a great demand to build models and systems that support multilingual information access. To this end, we propose a dataset for Multimodal Multiparty Hindi Humor (M2H2) recognition in conversations containing 6,191 utterances from 13 episodes of a very popular TV series "Shrimaan Shrimati Phir Se". Each utterance is annotated with humor/non-humor labels and encompasses acoustic, visual, and textual modalities. We propose several strong multimodal baselines and show the importance of contextual and multimodal information for humor recognition in conversations. The empirical results on M2H2 dataset demonstrate that multimodal information complements unimodal information for humor recognition. The dataset and the baselines are available at http://www.iitp.ac.in/~ai-nlp-ml/resources.html and https://github.com/declare-lab/M2H2-dataset.

Wei Han, Hui Chen, Alexander Gelbukh et al.

Multimodal sentiment analysis aims to extract and integrate semantic information collected from multiple modalities to recognize the expressed emotions and sentiment in multimodal data. This research area's major concern lies in developing an extraordinary fusion scheme that can extract and integrate key information from various modalities. However, one issue that may restrict previous work to achieve a higher level is the lack of proper modeling for the dynamics of the competition between the independence and relevance among modalities, which could deteriorate fusion outcomes by causing the collapse of modality-specific feature space or introducing extra noise. To mitigate this, we propose the Bi-Bimodal Fusion Network (BBFN), a novel end-to-end network that performs fusion (relevance increment) and separation (difference increment) on pairwise modality representations. The two parts are trained simultaneously such that the combat between them is simulated. The model takes two bimodal pairs as input due to the known information imbalance among modalities. In addition, we leverage a gated control mechanism in the Transformer architecture to further improve the final output. Experimental results on three datasets (CMU-MOSI, CMU-MOSEI, and UR-FUNNY) verifies that our model significantly outperforms the SOTA. The implementation of this work is available at https://github.com/declare-lab/multimodal-deep-learning.

Chia-Yu Hung, Qi Sun, Pengfei Hong et al.

Existing Visual-Language-Action (VLA) models have shown promising performance in zero-shot scenarios, demonstrating impressive task execution and reasoning capabilities. However, a significant challenge arises from the limitations of visual encoding, which can result in failures during tasks such as object grasping. Moreover, these models typically suffer from high computational overhead due to their large sizes, often exceeding 7B parameters. While these models excel in reasoning and task planning, the substantial computational overhead they incur makes them impractical for real-time robotic environments, where speed and efficiency are paramount. To address the limitations of existing VLA models, we propose NORA, a 3B-parameter model designed to reduce computational overhead while maintaining strong task performance. NORA adopts the Qwen-2.5-VL-3B multimodal model as its backbone, leveraging its superior visual-semantic understanding to enhance visual reasoning and action grounding. Additionally, our \model{} is trained on 970k real-world robot demonstrations and equipped with the FAST+ tokenizer for efficient action sequence generation. Experimental results demonstrate that NORA outperforms existing large-scale VLA models, achieving better task performance with significantly reduced computational overhead, making it a more practical solution for real-time robotic autonomy.

Shang Hong Sim, Tej Deep Pala, Vernon Toh et al.

Generating grounded and trustworthy responses remains a key challenge for large language models (LLMs). While retrieval-augmented generation (RAG) with citation-based grounding holds promise, instruction-tuned models frequently fail even in straightforward scenarios: missing explicitly stated answers, citing incorrectly, or refusing when evidence is available. In this work, we explore how reinforcement learning (RL) and internal reasoning can enhance grounding in LLMs. We use the GRPO (Group Relative Policy Optimization) method to train models using verifiable outcome-based rewards targeting answer correctness, citation sufficiency, and refusal quality, without requiring gold reasoning traces or expensive annotations. Through comprehensive experiments across ASQA, QAMPARI, ELI5, and ExpertQA we show that reasoning-augmented models significantly outperform instruction-only variants, especially in handling unanswerable queries and generating well-cited responses. A two-stage training setup, first optimizing answer and citation behavior and then refusal, further improves grounding by stabilizing the learning signal. Additionally, we revisit instruction tuning via GPT-4 distillation and find that combining it with GRPO enhances performance on long-form, generative QA tasks. Overall, our findings highlight the value of reasoning, stage-wise optimization, and outcome-driven RL for building more verifiable and reliable LLMs.

Han Lin, Jaemin Cho, Amir Zadeh et al. · allen-ai

There is growing interest in integrating high-fidelity visual synthesis capabilities into large language models (LLMs) without compromising their strong reasoning capabilities. Existing methods that directly train LLMs or bridge LLMs and diffusion models usually suffer from costly training since the backbone LLMs have not seen image representations during pretraining. We present Bifrost-1, a unified framework that bridges pretrained multimodal LLMs (MLLMs) and diffusion models using patch-level CLIP image embeddings as latent variables, which are natively aligned with the MLLM's CLIP visual encoder. These patch-level image embeddings are integrated into the diffusion model with a lightweight adaptation of its ControlNet. To retain the original multimodal reasoning capabilities of MLLMs, we equip the MLLM with a visual generation branch initialized from the original MLLM parameters when predicting the patch-level image embeddings. By seamlessly integrating pretrained MLLMs and diffusion models with patch-level CLIP latents, our framework enables high-fidelity controllable image generation with significant training efficiency. Our experiments demonstrate that Bifrost-1 achieves comparable or better performance than previous methods in terms of visual fidelity and multimodal understanding, with substantially lower compute during training. We also provide comprehensive ablation studies showing the effectiveness of our design choices.

Weisheng Jin, Maojia Song, Tej Deep Pala et al.

As large language models (LLMs) tackle increasingly complex tasks and longer documents, their computational and memory costs during inference become a major bottleneck. To address this, we propose PromptDistill, a novel, training-free method that improves inference efficiency while preserving generation quality. PromptDistill identifies and retains the most informative tokens by leveraging attention interactions in early layers, preserving their hidden states while reducing the computational burden in later layers. This allows the model to focus on essential contextual information without fully processing all tokens. Unlike previous methods such as H2O and SnapKV, which perform compression only after processing the entire input, or GemFilter, which selects a fixed portion of the initial prompt without considering contextual dependencies, PromptDistill dynamically allocates computational resources to the most relevant tokens while maintaining a global awareness of the input. Experiments using our method and baseline approaches with base models such as LLaMA 3.1 8B Instruct, Phi 3.5 Mini Instruct, and Qwen2 7B Instruct on benchmarks including LongBench, InfBench, and Needle in a Haystack demonstrate that PromptDistill significantly improves efficiency while having minimal impact on output quality compared to the original models. With a single-stage selection strategy, PromptDistill effectively balances performance and efficiency, outperforming prior methods like GemFilter, H2O, and SnapKV due to its superior ability to retain essential information. Specifically, compared to GemFilter, PromptDistill achieves an overall $1\%$ to $5\%$ performance improvement while also offering better time efficiency. Additionally, we explore multi-stage selection, which further improves efficiency while maintaining strong generation performance.

Amir Zadeh, Santiago Benoit, Louis-Philippe Morency

Variational Inference (VI) offers a method for approximating intractable likelihoods. In neural VI, inference of approximate posteriors is commonly done using an encoder. Alternatively, encoderless VI offers a framework for learning generative models from data without encountering suboptimalities caused by amortization via an encoder (e.g. in presence of missing or uncertain data). However, in absence of an encoder, such methods often suffer in convergence due to the slow nature of gradient steps required to learn the approximate posterior parameters. In this paper, we introduce Relay VI (RVI), a framework that dramatically improves both the convergence and performance of encoderless VI. In our experiments over multiple datasets, we study the effectiveness of RVI in terms of convergence speed, loss, representation power and missing data imputation. We find RVI to be a unique tool, often superior in both performance and convergence speed to previously proposed encoderless as well as amortized VI models (e.g. VAE).

Amir Zadeh, Santiago Benoit, Louis-Philippe Morency

In this paper we present an approach for training deep generative models solely based on solving determined systems of linear equations. A network that uses this approach, called a StarNet, has the following desirable properties: 1) training requires no gradient as solution to the system of linear equations is not stochastic, 2) is highly scalable when solving the system of linear equations w.r.t the latent codes, and similarly for the parameters of the model, and 3) it gives desirable least-square bounds for the estimation of latent codes and network parameters within each layer.

Jianing Yang, Yongxin Wang, Ruitao Yi et al.

Human communication is multimodal in nature; it is through multiple modalities such as language, voice, and facial expressions, that opinions and emotions are expressed. Data in this domain exhibits complex multi-relational and temporal interactions. Learning from this data is a fundamentally challenging research problem. In this paper, we propose Modal-Temporal Attention Graph (MTAG). MTAG is an interpretable graph-based neural model that provides a suitable framework for analyzing multimodal sequential data. We first introduce a procedure to convert unaligned multimodal sequence data into a graph with heterogeneous nodes and edges that captures the rich interactions across modalities and through time. Then, a novel graph fusion operation, called MTAG fusion, along with a dynamic pruning and read-out technique, is designed to efficiently process this modal-temporal graph and capture various interactions. By learning to focus only on the important interactions within the graph, MTAG achieves state-of-the-art performance on multimodal sentiment analysis and emotion recognition benchmarks, while utilizing significantly fewer model parameters.

Shagun Uppal, Sarthak Bhagat, Devamanyu Hazarika et al.

Deep Learning and its applications have cascaded impactful research and development with a diverse range of modalities present in the real-world data. More recently, this has enhanced research interests in the intersection of the Vision and Language arena with its numerous applications and fast-paced growth. In this paper, we present a detailed overview of the latest trends in research pertaining to visual and language modalities. We look at its applications in their task formulations and how to solve various problems related to semantic perception and content generation. We also address task-specific trends, along with their evaluation strategies and upcoming challenges. Moreover, we shed some light on multi-disciplinary patterns and insights that have emerged in the recent past, directing this field towards more modular and transparent intelligent systems. This survey identifies key trends gravitating recent literature in VisLang research and attempts to unearth directions that the field is heading towards.

Jianing Yang, Yuying Zhu, Yongxin Wang et al.

Question answering biases in video QA datasets can mislead multimodal model to overfit to QA artifacts and jeopardize the model's ability to generalize. Understanding how strong these QA biases are and where they come from helps the community measure progress more accurately and provide researchers insights to debug their models. In this paper, we analyze QA biases in popular video question answering datasets and discover pretrained language models can answer 37-48% questions correctly without using any multimodal context information, far exceeding the 20% random guess baseline for 5-choose-1 multiple-choice questions. Our ablation study shows biases can come from annotators and type of questions. Specifically, annotators that have been seen during training are better predicted by the model and reasoning, abstract questions incur more biases than factual, direct questions. We also show empirically that using annotator-non-overlapping train-test splits can reduce QA biases for video QA datasets.

Navonil Majumder, Rishabh Bhardwaj, Soujanya Poria et al.

Aspect-based sentiment analysis (ABSA), a popular research area in NLP has two distinct parts -- aspect extraction (AE) and labeling the aspects with sentiment polarity (ALSA). Although distinct, these two tasks are highly correlated. The work primarily hypothesize that transferring knowledge from a pre-trained AE model can benefit the performance of ALSA models. Based on this hypothesis, word embeddings are obtained during AE and subsequently, feed that to the ALSA model. Empirically, this work show that the added information significantly improves the performance of three different baseline ALSA models on two distinct domains. This improvement also translates well across domains between AE and ALSA tasks.

Amir Zadeh, Smon Hessner, Yao-Chong Lim et al.

Posterior inference in directed graphical models is commonly done using a probabilistic encoder (a.k.a inference model) conditioned on the input. Often this inference model is trained jointly with the probabilistic decoder (a.k.a generator model). If probabilistic encoder encounters complexities during training (e.g. suboptimal complxity or parameterization), then learning reaches a suboptimal objective; a phenomena commonly called inference suboptimality. In Variational Inference (VI), optimizing the ELBo using Stochastic Variational Inference (SVI) can eliminate the inference suboptimality (as demonstrated in this paper), however, this solution comes at a substantial computational cost when inference needs to be done on new data points. Essentially, a long sequential chain of gradient updates is required to fully optimize approximate posteriors. In this paper, we present an approach called Pseudo-Encoded Stochastic Variational Inference (PE-SVI), to reduce the inference complexity of SVI during test time. Our approach relies on finding a suitable initial start point for gradient operations, which naturally reduces the required gradient steps. Furthermore, this initialization allows for adopting larger step sizes (compared to random initialization used in SVI), which further reduces the inference time complexity. PE-SVI reaches the same ELBo objective as SVI using less than one percent of required steps, on average.

Amir Zadeh, Chengfeng Mao, Kelly Shi et al.

The complex world around us is inherently multimodal and sequential (continuous). Information is scattered across different modalities and requires multiple continuous sensors to be captured. As machine learning leaps towards better generalization to real world, multimodal sequential learning becomes a fundamental research area. Arguably, modeling arbitrarily distributed spatio-temporal dynamics within and across modalities is the biggest challenge in this research area. In this paper, we present a new transformer model, called the Factorized Multimodal Transformer (FMT) for multimodal sequential learning. FMT inherently models the intramodal and intermodal (involving two or more modalities) dynamics within its multimodal input in a factorized manner. The proposed factorization allows for increasing the number of self-attentions to better model the multimodal phenomena at hand; without encountering difficulties during training (e.g. overfitting) even on relatively low-resource setups. All the attention mechanisms within FMT have a full time-domain receptive field which allows them to asynchronously capture long-range multimodal dynamics. In our experiments we focus on datasets that contain the three commonly studied modalities of language, vision and acoustic. We perform a wide range of experiments, spanning across 3 well-studied datasets and 21 distinct labels. FMT shows superior performance over previously proposed models, setting new state of the art in the studied datasets.

Amir Zadeh, Tianjun Ma, Soujanya Poria et al.

Monoaural audio source separation is a challenging research area in machine learning. In this area, a mixture containing multiple audio sources is given, and a model is expected to disentangle the mixture into isolated atomic sources. In this paper, we first introduce a challenging new dataset for monoaural source separation called WildMix. WildMix is designed with the goal of extending the boundaries of source separation beyond what previous datasets in this area would allow. It contains diverse in-the-wild recordings from 25 different sound classes, combined with each other using arbitrary composition policies. Source separation often requires modeling long-range dependencies in both temporal and spectral domains. To this end, we introduce a novel trasnformer-based model called Spectro-Temporal Transformer (STT). STT utilizes a specialized encoder, called Spectro-Temporal Encoder (STE). STE highlights temporal and spectral components of sources within a mixture, using a self-attention mechanism. It subsequently disentangles them in a hierarchical manner. In our experiments, STT swiftly outperforms various previous baselines for monoaural source separation on the challenging WildMix dataset.

Wasifur Rahman, Md. Kamrul Hasan, Sangwu Lee et al.

Recent Transformer-based contextual word representations, including BERT and XLNet, have shown state-of-the-art performance in multiple disciplines within NLP. Fine-tuning the trained contextual models on task-specific datasets has been the key to achieving superior performance downstream. While fine-tuning these pre-trained models is straightforward for lexical applications (applications with only language modality), it is not trivial for multimodal language (a growing area in NLP focused on modeling face-to-face communication). Pre-trained models don't have the necessary components to accept two extra modalities of vision and acoustic. In this paper, we proposed an attachment to BERT and XLNet called Multimodal Adaptation Gate (MAG). MAG allows BERT and XLNet to accept multimodal nonverbal data during fine-tuning. It does so by generating a shift to internal representation of BERT and XLNet; a shift that is conditioned on the visual and acoustic modalities. In our experiments, we study the commonly used CMU-MOSI and CMU-MOSEI datasets for multimodal sentiment analysis. Fine-tuning MAG-BERT and MAG-XLNet significantly boosts the sentiment analysis performance over previous baselines as well as language-only fine-tuning of BERT and XLNet. On the CMU-MOSI dataset, MAG-XLNet achieves human-level multimodal sentiment analysis performance for the first time in the NLP community.

Amir Zadeh, Yao-Chong Lim, Paul Pu Liang et al.

Learning a generative model from partial data (data with missingness) is a challenging area of machine learning research. We study a specific implementation of the Auto-Encoding Variational Bayes (AEVB) algorithm, named in this paper as a Variational Auto-Decoder (VAD). VAD is a generic framework which uses Variational Bayes and Markov Chain Monte Carlo (MCMC) methods to learn a generative model from partial data. The main distinction between VAD and Variational Auto-Encoder (VAE) is the encoder component, as VAD does not have one. Using a proposed efficient inference method from a multivariate Gaussian approximate posterior, VAD models allow inference to be performed via simple gradient ascent rather than MCMC sampling from a probabilistic decoder. This technique reduces the inference computational cost, allows for using more complex optimization techniques during latent space inference (which are shown to be crucial due to a high degree of freedom in the VAD latent space), and keeps the framework simple to implement. Through extensive experiments over several datasets and different missing ratios, we show that encoders cannot efficiently marginalize the input volatility caused by imputed missing values. We study multimodal datasets in this paper, which is a particular area of impact for VAD models.

Yansen Wang, Ying Shen, Zhun Liu et al.

Humans convey their intentions through the usage of both verbal and nonverbal behaviors during face-to-face communication. Speaker intentions often vary dynamically depending on different nonverbal contexts, such as vocal patterns and facial expressions. As a result, when modeling human language, it is essential to not only consider the literal meaning of the words but also the nonverbal contexts in which these words appear. To better model human language, we first model expressive nonverbal representations by analyzing the fine-grained visual and acoustic patterns that occur during word segments. In addition, we seek to capture the dynamic nature of nonverbal intents by shifting word representations based on the accompanying nonverbal behaviors. To this end, we propose the Recurrent Attended Variation Embedding Network (RAVEN) that models the fine-grained structure of nonverbal subword sequences and dynamically shifts word representations based on nonverbal cues. Our proposed model achieves competitive performance on two publicly available datasets for multimodal sentiment analysis and emotion recognition. We also visualize the shifted word representations in different nonverbal contexts and summarize common patterns regarding multimodal variations of word representations.

Paul Pu Liang, Ziyin Liu, Amir Zadeh et al.

Computational modeling of human multimodal language is an emerging research area in natural language processing spanning the language, visual and acoustic modalities. Comprehending multimodal language requires modeling not only the interactions within each modality (intra-modal interactions) but more importantly the interactions between modalities (cross-modal interactions). In this paper, we propose the Recurrent Multistage Fusion Network (RMFN) which decomposes the fusion problem into multiple stages, each of them focused on a subset of multimodal signals for specialized, effective fusion. Cross-modal interactions are modeled using this multistage fusion approach which builds upon intermediate representations of previous stages. Temporal and intra-modal interactions are modeled by integrating our proposed fusion approach with a system of recurrent neural networks. The RMFN displays state-of-the-art performance in modeling human multimodal language across three public datasets relating to multimodal sentiment analysis, emotion recognition, and speaker traits recognition. We provide visualizations to show that each stage of fusion focuses on a different subset of multimodal signals, learning increasingly discriminative multimodal representations.

Paul Pu Liang, Amir Zadeh, Louis-Philippe Morency

Emotion recognition is a core research area at the intersection of artificial intelligence and human communication analysis. It is a significant technical challenge since humans display their emotions through complex idiosyncratic combinations of the language, visual and acoustic modalities. In contrast to traditional multimodal fusion techniques, we approach emotion recognition from both direct person-independent and relative person-dependent perspectives. The direct person-independent perspective follows the conventional emotion recognition approach which directly infers absolute emotion labels from observed multimodal features. The relative person-dependent perspective approaches emotion recognition in a relative manner by comparing partial video segments to determine if there was an increase or decrease in emotional intensity. Our proposed model integrates these direct and relative prediction perspectives by dividing the emotion recognition task into three easier subtasks. The first subtask involves a multimodal local ranking of relative emotion intensities between two short segments of a video. The second subtask uses local rankings to infer global relative emotion ranks with a Bayesian ranking algorithm. The third subtask incorporates both direct predictions from observed multimodal behaviors and relative emotion ranks from local-global rankings for final emotion prediction. Our approach displays excellent performance on an audio-visual emotion recognition benchmark and improves over other algorithms for multimodal fusion.

Yao-Hung Hubert Tsai, Paul Pu Liang, Amir Zadeh et al.

Learning multimodal representations is a fundamentally complex research problem due to the presence of multiple heterogeneous sources of information. Although the presence of multiple modalities provides additional valuable information, there are two key challenges to address when learning from multimodal data: 1) models must learn the complex intra-modal and cross-modal interactions for prediction and 2) models must be robust to unexpected missing or noisy modalities during testing. In this paper, we propose to optimize for a joint generative-discriminative objective across multimodal data and labels. We introduce a model that factorizes representations into two sets of independent factors: multimodal discriminative and modality-specific generative factors. Multimodal discriminative factors are shared across all modalities and contain joint multimodal features required for discriminative tasks such as sentiment prediction. Modality-specific generative factors are unique for each modality and contain the information required for generating data. Experimental results show that our model is able to learn meaningful multimodal representations that achieve state-of-the-art or competitive performance on six multimodal datasets. Our model demonstrates flexible generative capabilities by conditioning on independent factors and can reconstruct missing modalities without significantly impacting performance. Lastly, we interpret our factorized representations to understand the interactions that influence multimodal learning.

Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan et al.

Multimodal research is an emerging field of artificial intelligence, and one of the main research problems in this field is multimodal fusion. The fusion of multimodal data is the process of integrating multiple unimodal representations into one compact multimodal representation. Previous research in this field has exploited the expressiveness of tensors for multimodal representation. However, these methods often suffer from exponential increase in dimensions and in computational complexity introduced by transformation of input into tensor. In this paper, we propose the Low-rank Multimodal Fusion method, which performs multimodal fusion using low-rank tensors to improve efficiency. We evaluate our model on three different tasks: multimodal sentiment analysis, speaker trait analysis, and emotion recognition. Our model achieves competitive results on all these tasks while drastically reducing computational complexity. Additional experiments also show that our model can perform robustly for a wide range of low-rank settings, and is indeed much more efficient in both training and inference compared to other methods that utilize tensor representations.

Amir Zadeh, Paul Pu Liang, Navonil Mazumder et al.

Multi-view sequential learning is a fundamental problem in machine learning dealing with multi-view sequences. In a multi-view sequence, there exists two forms of interactions between different views: view-specific interactions and cross-view interactions. In this paper, we present a new neural architecture for multi-view sequential learning called the Memory Fusion Network (MFN) that explicitly accounts for both interactions in a neural architecture and continuously models them through time. The first component of the MFN is called the System of LSTMs, where view-specific interactions are learned in isolation through assigning an LSTM function to each view. The cross-view interactions are then identified using a special attention mechanism called the Delta-memory Attention Network (DMAN) and summarized through time with a Multi-view Gated Memory. Through extensive experimentation, MFN is compared to various proposed approaches for multi-view sequential learning on multiple publicly available benchmark datasets. MFN outperforms all the existing multi-view approaches. Furthermore, MFN outperforms all current state-of-the-art models, setting new state-of-the-art results for these multi-view datasets.

Minghai Chen, Sen Wang, Paul Pu Liang et al.

With the increasing popularity of video sharing websites such as YouTube and Facebook, multimodal sentiment analysis has received increasing attention from the scientific community. Contrary to previous works in multimodal sentiment analysis which focus on holistic information in speech segments such as bag of words representations and average facial expression intensity, we develop a novel deep architecture for multimodal sentiment analysis that performs modality fusion at the word level. In this paper, we propose the Gated Multimodal Embedding LSTM with Temporal Attention (GME-LSTM(A)) model that is composed of 2 modules. The Gated Multimodal Embedding alleviates the difficulties of fusion when there are noisy modalities. The LSTM with Temporal Attention performs word level fusion at a finer fusion resolution between input modalities and attends to the most important time steps. As a result, the GME-LSTM(A) is able to better model the multimodal structure of speech through time and perform better sentiment comprehension. We demonstrate the effectiveness of this approach on the publicly-available Multimodal Corpus of Sentiment Intensity and Subjectivity Analysis (CMU-MOSI) dataset by achieving state-of-the-art sentiment classification and regression results. Qualitative analysis on our model emphasizes the importance of the Temporal Attention Layer in sentiment prediction because the additional acoustic and visual modalities are noisy. We also demonstrate the effectiveness of the Gated Multimodal Embedding in selectively filtering these noisy modalities out. Our results and analysis open new areas in the study of sentiment analysis in human communication and provide new models for multimodal fusion.

Amir Zadeh, Paul Pu Liang, Soujanya Poria et al.

Human face-to-face communication is a complex multimodal signal. We use words (language modality), gestures (vision modality) and changes in tone (acoustic modality) to convey our intentions. Humans easily process and understand face-to-face communication, however, comprehending this form of communication remains a significant challenge for Artificial Intelligence (AI). AI must understand each modality and the interactions between them that shape human communication. In this paper, we present a novel neural architecture for understanding human communication called the Multi-attention Recurrent Network (MARN). The main strength of our model comes from discovering interactions between modalities through time using a neural component called the Multi-attention Block (MAB) and storing them in the hybrid memory of a recurrent component called the Long-short Term Hybrid Memory (LSTHM). We perform extensive comparisons on six publicly available datasets for multimodal sentiment analysis, speaker trait recognition and emotion recognition. MARN shows state-of-the-art performance on all the datasets.

Amir Zadeh, Minghai Chen, Soujanya Poria et al.

Multimodal sentiment analysis is an increasingly popular research area, which extends the conventional language-based definition of sentiment analysis to a multimodal setup where other relevant modalities accompany language. In this paper, we pose the problem of multimodal sentiment analysis as modeling intra-modality and inter-modality dynamics. We introduce a novel model, termed Tensor Fusion Network, which learns both such dynamics end-to-end. The proposed approach is tailored for the volatile nature of spoken language in online videos as well as accompanying gestures and voice. In the experiments, our model outperforms state-of-the-art approaches for both multimodal and unimodal sentiment analysis.

Edmund Tong, Amir Zadeh, Cara Jones et al.

Human trafficking is a global epidemic affecting millions of people across the planet. Sex trafficking, the dominant form of human trafficking, has seen a significant rise mostly due to the abundance of escort websites, where human traffickers can openly advertise among at-will escort advertisements. In this paper, we take a major step in the automatic detection of advertisements suspected to pertain to human trafficking. We present a novel dataset called Trafficking-10k, with more than 10,000 advertisements annotated for this task. The dataset contains two sources of information per advertisement: text and images. For the accurate detection of trafficking advertisements, we designed and trained a deep multimodal model called the Human Trafficking Deep Network (HTDN).

Amir Zadeh, Tadas Baltrušaitis, Louis-Philippe Morency

Constrained Local Models (CLMs) are a well-established family of methods for facial landmark detection. However, they have recently fallen out of favor to cascaded regression-based approaches. This is in part due to the inability of existing CLM local detectors to model the very complex individual landmark appearance that is affected by expression, illumination, facial hair, makeup, and accessories. In our work, we present a novel local detector -- Convolutional Experts Network (CEN) -- that brings together the advantages of neural architectures and mixtures of experts in an end-to-end framework. We further propose a Convolutional Experts Constrained Local Model (CE-CLM) algorithm that uses CEN as local detectors. We demonstrate that our proposed CE-CLM algorithm outperforms competitive state-of-the-art baselines for facial landmark detection by a large margin on four publicly-available datasets. Our approach is especially accurate and robust on challenging profile images.

Amir Zadeh, Rowan Zellers, Eli Pincus et al.

People are sharing their opinions, stories and reviews through online video sharing websites every day. Studying sentiment and subjectivity in these opinion videos is experiencing a growing attention from academia and industry. While sentiment analysis has been successful for text, it is an understudied research question for videos and multimedia content. The biggest setbacks for studies in this direction are lack of a proper dataset, methodology, baselines and statistical analysis of how information from different modality sources relate to each other. This paper introduces to the scientific community the first opinion-level annotated corpus of sentiment and subjectivity analysis in online videos called Multimodal Opinion-level Sentiment Intensity dataset (MOSI). The dataset is rigorously annotated with labels for subjectivity, sentiment intensity, per-frame and per-opinion annotated visual features, and per-milliseconds annotated audio features. Furthermore, we present baselines for future studies in this direction as well as a new multimodal fusion approach that jointly models spoken words and visual gestures.