Hai Pham

Hai Pham, Young Jin Kim, Subhabrata Mukherjee et al. · cmu, microsoft-research

Mixture-of-experts (MoE) architecture has been proven a powerful method for diverse tasks in training deep models in many applications. However, current MoE implementations are task agnostic, treating all tokens from different tasks in the same manner. In this work, we instead design a novel method that incorporates task information into MoE models at different granular levels with shared dynamic task-based adapters. Our experiments and analysis show the advantages of our approaches over the dense and canonical MoE models on multi-task multilingual machine translations. With task-specific adapters, our models can additionally generalize to new tasks efficiently.

Hai X. Pham, David T. Hoffmann, Ricardo Guerrero et al.

Contrastive vision-language (V&L) models remain a popular choice for various applications. However, several limitations have emerged, most notably the limited ability of V&L models to learn compositional representations. Prior methods often addressed this limitation by generating custom training data to obtain hard negative samples. Hard negatives have been shown to improve performance on compositionality tasks, but are often specific to a single benchmark, do not generalize, and can cause substantial degradation of basic V&L capabilities such as zero-shot or retrieval performance, rendering them impractical. In this work we follow a different approach. We identify two root causes that limit compositionality performance of V&Ls: 1) Long training captions do not require a compositional representation; and 2) The final global pooling in the text and image encoders lead to a complete loss of the necessary information to learn binding in the first place. As a remedy, we propose two simple solutions: 1) We obtain short concept centric caption parts using standard NLP software and align those with the image; and 2) We introduce a parameter-free cross-modal attention-pooling to obtain concept centric visual embeddings from the image encoder. With these two changes and simple auxiliary contrastive losses, we obtain SOTA performance on standard compositionality benchmarks, while maintaining or improving strong zero-shot and retrieval capabilities. This is achieved without increasing inference cost. We release the code for this work at https://github.com/SamsungLabs/concept_centric_clip.

Piotr Padlewski, Max Bain, Matthew Henderson et al. · pku

We introduce Vibe-Eval: a new open benchmark and framework for evaluating multimodal chat models. Vibe-Eval consists of 269 visual understanding prompts, including 100 of hard difficulty, complete with gold-standard responses authored by experts. Vibe-Eval is open-ended and challenging with dual objectives: (i) vibe checking multimodal chat models for day-to-day tasks and (ii) rigorously testing and probing the capabilities of present frontier models. Notably, our hard set contains >50% questions that all frontier models answer incorrectly. We explore the nuances of designing, evaluating, and ranking models on ultra challenging prompts. We also discuss trade-offs between human and automatic evaluation, and show that automatic model evaluation using Reka Core roughly correlates to human judgment. We offer free API access for the purpose of lightweight evaluation and plan to conduct formal human evaluations for public models that perform well on the Vibe-Eval's automatic scores. We release the evaluation code and data, see https://github.com/reka-ai/reka-vibe-eval

Reka Team, Aitor Ormazabal, Che Zheng et al. · pku

We introduce Reka Core, Flash, and Edge, a series of powerful multimodal language models trained from scratch by Reka. Reka models are able to process and reason with text, images, video, and audio inputs. This technical report discusses details of training some of these models and provides comprehensive evaluation results. We show that Reka Edge and Reka Flash are not only state-of-the-art but also outperform many much larger models, delivering outsized values for their respective compute class. Meanwhile, our most capable and largest model, Reka Core, approaches the best frontier models on both automatic evaluations and blind human evaluations. On image question answering benchmarks (e.g. MMMU, VQAv2), Core performs competitively to GPT4-V. Meanwhile, on multimodal chat, Core ranks as the second most preferred model under a blind third-party human evaluation setup, outperforming other models such as Claude 3 Opus. On text benchmarks, Core not only performs competitively to other frontier models on a set of well-established benchmarks (e.g. MMLU, GSM8K) but also outperforms GPT4-0613 on human evaluation. On video question answering (Perception-Test), Core outperforms Gemini Ultra. Models are shipped in production at http://chat.reka.ai . A showcase of non cherry picked qualitative examples can also be found at http://showcase.reka.ai .

Haosen Yang, Adrian Bulat, Isma Hadji et al.

Diffusion models are proficient at generating high-quality images. They are however effective only when operating at the resolution used during training. Inference at a scaled resolution leads to repetitive patterns and structural distortions. Retraining at higher resolutions quickly becomes prohibitive. Thus, methods enabling pre-existing diffusion models to operate at flexible test-time resolutions are highly desirable. Previous works suffer from frequent artifacts and often introduce large latency overheads. We propose two simple modules that combine to solve these issues. We introduce a Frequency Modulation (FM) module that leverages the Fourier domain to improve the global structure consistency, and an Attention Modulation (AM) module which improves the consistency of local texture patterns, a problem largely ignored in prior works. Our method, coined Fam diffusion, can seamlessly integrate into any latent diffusion model and requires no additional training. Extensive qualitative results highlight the effectiveness of our method in addressing structural and local artifacts, while quantitative results show state-of-the-art performance. Also, our method avoids redundant inference tricks for improved consistency such as patch-based or progressive generation, leading to negligible latency overheads.

Hai X. Pham, Isma Hadji, Xinnuo Xu et al.

In this paper, we focus on task-specific question answering (QA). To this end, we introduce a method for generating exhaustive and high-quality training data, which allows us to train compact (e.g., run on a mobile device), task-specific QA models that are competitive against GPT variants. The key technological enabler is a novel mechanism for automatic question-answer generation from procedural text which can ingest large amounts of textual instructions and produce exhaustive in-domain QA training data. While current QA data generation methods can produce well-formed and varied data, their non-exhaustive nature is sub-optimal for training a QA model. In contrast, we leverage the highly structured aspect of procedural text and represent each step and the overall flow of the procedure as graphs. We then condition on graph nodes to automatically generate QA pairs in an exhaustive and controllable manner. Comprehensive evaluations of our method show that: 1) small models trained with our data achieve excellent performance on the target QA task, even exceeding that of GPT3 and ChatGPT despite being several orders of magnitude smaller. 2) semantic coverage is the key indicator for downstream QA performance. Crucially, while large language models excel at syntactic diversity, this does not necessarily result in improvements on the end QA model. In contrast, the higher semantic coverage provided by our method is critical for QA performance.

Hai X. Pham, Ricardo Guerrero, Jiatong Li et al.

Despite the abundance of multi-modal data, such as image-text pairs, there has been little effort in understanding the individual entities and their different roles in the construction of these data instances. In this work, we endeavour to discover the entities and their corresponding importance in cooking recipes automaticall} as a visual-linguistic association problem. More specifically, we introduce a novel cross-modal learning framework to jointly model the latent representations of images and text in the food image-recipe association and retrieval tasks. This model allows one to discover complex functional and hierarchical relationships between images and text, and among textual parts of a recipe including title, ingredients and cooking instructions. Our experiments show that by making use of efficient tree-structured Long Short-Term Memory as the text encoder in our computational cross-modal retrieval framework, we are not only able to identify the main ingredients and cooking actions in the recipe descriptions without explicit supervision, but we can also learn more meaningful feature representations of food recipes, appropriate for challenging cross-modal retrieval and recipe adaption tasks.

Ricardo Guerrero, Hai Xuan Pham, Vladimir Pavlovic

Computational food analysis (CFA) naturally requires multi-modal evidence of a particular food, e.g., images, recipe text, etc. A key to making CFA possible is multi-modal shared representation learning, which aims to create a joint representation of the multiple views (text and image) of the data. In this work we propose a method for food domain cross-modal shared representation learning that preserves the vast semantic richness present in the food data. Our proposed method employs an effective transformer-based multilingual recipe encoder coupled with a traditional image embedding architecture. Here, we propose the use of imperfect multilingual translations to effectively regularize the model while at the same time adding functionality across multiple languages and alphabets. Experimental analysis on the public Recipe1M dataset shows that the representation learned via the proposed method significantly outperforms the current state-of-the-arts (SOTA) on retrieval tasks. Furthermore, the representational power of the learned representation is demonstrated through a generative food image synthesis model conditioned on recipe embeddings. Synthesized images can effectively reproduce the visual appearance of paired samples, indicating that the learned representation captures the joint semantics of both the textual recipe and its visual content, thus narrowing the modality gap.

Quang Minh Hoang, Trong Nghia Hoang, Hai Pham et al.

We introduce a new scalable approximation for Gaussian processes with provable guarantees which hold simultaneously over its entire parameter space. Our approximation is obtained from an improved sample complexity analysis for sparse spectrum Gaussian processes (SSGPs). In particular, our analysis shows that under a certain data disentangling condition, an SSGP's prediction and model evidence (for training) can well-approximate those of a full GP with low sample complexity. We also develop a new auto-encoding algorithm that finds a latent space to disentangle latent input coordinates into well-separated clusters, which is amenable to our sample complexity analysis. We validate our proposed method on several benchmarks with promising results supporting our theoretical analysis.

Hai Pham, Amrith Setlur, Saket Dingliwal et al.

Despite the advent of deep learning in computer vision, the general handwriting recognition problem is far from solved. Most existing approaches focus on handwriting datasets that have clearly written text and carefully segmented labels. In this paper, we instead focus on learning handwritten characters from maintenance logs, a constrained setting where data is very limited and noisy. We break the problem into two consecutive stages of word segmentation and word recognition respectively and utilize data augmentation techniques to train both stages. Extensive comparisons with popular baselines for scene-text detection and word recognition show that our system achieves a lower error rate and is more suited to handle noisy and difficult documents

Hai Pham, Paul Pu Liang, Thomas Manzini et al.

Multimodal sentiment analysis is a core research area that studies speaker sentiment expressed from the language, visual, and acoustic modalities. The central challenge in multimodal learning involves inferring joint representations that can process and relate information from these modalities. However, existing work learns joint representations by requiring all modalities as input and as a result, the learned representations may be sensitive to noisy or missing modalities at test time. With the recent success of sequence to sequence (Seq2Seq) models in machine translation, there is an opportunity to explore new ways of learning joint representations that may not require all input modalities at test time. In this paper, we propose a method to learn robust joint representations by translating between modalities. Our method is based on the key insight that translation from a source to a target modality provides a method of learning joint representations using only the source modality as input. We augment modality translations with a cycle consistency loss to ensure that our joint representations retain maximal information from all modalities. Once our translation model is trained with paired multimodal data, we only need data from the source modality at test time for final sentiment prediction. This ensures that our model remains robust from perturbations or missing information in the other modalities. We train our model with a coupled translation-prediction objective and it achieves new state-of-the-art results on multimodal sentiment analysis datasets: CMU-MOSI, ICT-MMMO, and YouTube. Additional experiments show that our model learns increasingly discriminative joint representations with more input modalities while maintaining robustness to missing or perturbed modalities.

Hai Pham, Thomas Manzini, Paul Pu Liang et al.

Multimodal machine learning is a core research area spanning the language, visual and acoustic modalities. The central challenge in multimodal learning involves learning representations that can process and relate information from multiple modalities. In this paper, we propose two methods for unsupervised learning of joint multimodal representations using sequence to sequence (Seq2Seq) methods: a \textit{Seq2Seq Modality Translation Model} and a \textit{Hierarchical Seq2Seq Modality Translation Model}. We also explore multiple different variations on the multimodal inputs and outputs of these seq2seq models. Our experiments on multimodal sentiment analysis using the CMU-MOSI dataset indicate that our methods learn informative multimodal representations that outperform the baselines and achieve improved performance on multimodal sentiment analysis, specifically in the Bimodal case where our model is able to improve F1 Score by twelve points. We also discuss future directions for multimodal Seq2Seq methods.

Hai X. Pham, Yuting Wang, Vladimir Pavlovic

This paper presents Generative Adversarial Talking Head (GATH), a novel deep generative neural network that enables fully automatic facial expression synthesis of an arbitrary portrait with continuous action unit (AU) coefficients. Specifically, our model directly manipulates image pixels to make the unseen subject in the still photo express various emotions controlled by values of facial AU coefficients, while maintaining her personal characteristics, such as facial geometry, skin color and hair style, as well as the original surrounding background. In contrast to prior work, GATH is purely data-driven and it requires neither a statistical face model nor image processing tricks to enact facial deformations. Additionally, our model is trained from unpaired data, where the input image, with its auxiliary identity label taken from abundance of still photos in the wild, and the target frame are from different persons. In order to effectively learn such model, we propose a novel weakly supervised adversarial learning framework that consists of a generator, a discriminator, a classifier and an action unit estimator. Our work gives rise to template-and-target-free expression editing, where still faces can be effortlessly animated with arbitrary AU coefficients provided by the user.

Hai X. Pham, Yuting Wang, Vladimir Pavlovic

We present a deep learning framework for real-time speech-driven 3D facial animation from just raw waveforms. Our deep neural network directly maps an input sequence of speech audio to a series of micro facial action unit activations and head rotations to drive a 3D blendshape face model. In particular, our deep model is able to learn the latent representations of time-varying contextual information and affective states within the speech. Hence, our model not only activates appropriate facial action units at inference to depict different utterance generating actions, in the form of lip movements, but also, without any assumption, automatically estimates emotional intensity of the speaker and reproduces her ever-changing affective states by adjusting strength of facial unit activations. For example, in a happy speech, the mouth opens wider than normal, while other facial units are relaxed; or in a surprised state, both eyebrows raise higher. Experiments on a diverse audiovisual corpus of different actors across a wide range of emotional states show interesting and promising results of our approach. Being speaker-independent, our generalized model is readily applicable to various tasks in human-machine interaction and animation.

Hai X. Pham, Chongyu Chen, Luc N. Dao et al.

We introduce a novel robust hybrid 3D face tracking framework from RGBD video streams, which is capable of tracking head pose and facial actions without pre-calibration or intervention from a user. In particular, we emphasize on improving the tracking performance in instances where the tracked subject is at a large distance from the cameras, and the quality of point cloud deteriorates severely. This is accomplished by the combination of a flexible 3D shape regressor and the joint 2D+3D optimization on shape parameters. Our approach fits facial blendshapes to the point cloud of the human head, while being driven by an efficient and rapid 3D shape regressor trained on generic RGB datasets. As an on-line tracking system, the identity of the unknown user is adapted on-the-fly resulting in improved 3D model reconstruction and consequently better tracking performance. The result is a robust RGBD face tracker, capable of handling a wide range of target scene depths, beyond those that can be afforded by traditional depth or RGB face trackers. Lastly, since the blendshape is not able to accurately recover the real facial shape, we use the tracked 3D face model as a prior in a novel filtering process to further refine the depth map for use in other tasks, such as 3D reconstruction.