Nemotron-4 340B Technical Report

We release the Nemotron-4 340B model family, including Nemotron-4-340B-Base, Nemotron-4-340B-Instruct, and Nemotron-4-340B-Reward. Our models are open access under the NVIDIA Open Model License Agreement, a permissive model license that allows distribution, modification, and use of the models and its outputs. These models perform competitively to open access models on a wide range of evaluation benchmarks, and were sized to fit on a single DGX H100 with 8 GPUs when deployed in FP8 precision. We believe that the community can benefit from these models in various research studies and commercial applications, especially for generating synthetic data to train smaller language models. Notably, over 98% of data used in our model alignment process is synthetically generated, showcasing the effectiveness of these models in generating synthetic data. To further support open research and facilitate model development, we are also open-sourcing the synthetic data generation pipeline used in our model alignment process.

DiverseDream: Diverse Text-to-3D Synthesis with Augmented Text Embedding

Text-to-3D synthesis has recently emerged as a new approach to sampling 3D models by adopting pretrained text-to-image models as guiding visual priors. An intriguing but underexplored problem with existing text-to-3D methods is that 3D models obtained from the sampling-by-optimization procedure tend to have mode collapses, and hence poor diversity in their results. In this paper, we provide an analysis and identify potential causes of such a limited diversity, and then devise a new method that considers the joint generation of different 3D models from the same text prompt, where we propose to use augmented text prompts via textual inversion of reference images to diversify the joint generation. We show that our method leads to improved diversity in text-to-3D synthesis qualitatively and quantitatively.

EmbryosFormer: Deformable Transformer and Collaborative Encoding-Decoding for Embryos Stage Development Classification

The timing of cell divisions in early embryos during the In-Vitro Fertilization (IVF) process is a key predictor of embryo viability. However, observing cell divisions in Time-Lapse Monitoring (TLM) is a time-consuming process and highly depends on experts. In this paper, we propose EmbryosFormer, a computational model to automatically detect and classify cell divisions from original time-lapse images. Our proposed network is designed as an encoder-decoder deformable transformer with collaborative heads. The transformer contracting path predicts per-image labels and is optimized by a classification head. The transformer expanding path models the temporal coherency between embryo images to ensure monotonic non-decreasing constraint and is optimized by a segmentation head. Both contracting and expanding paths are synergetically learned by a collaboration head. We have benchmarked our proposed EmbryosFormer on two datasets: a public dataset with mouse embryos with 8-cell stage and an in-house dataset with human embryos with 4-cell stage. Source code: https://github.com/UARK-AICV/Embryos.

Human View Synthesis using a Single Sparse RGB-D Input

Novel view synthesis for humans in motion is a challenging computer vision problem that enables applications such as free-viewpoint video. Existing methods typically use complex setups with multiple input views, 3D supervision, or pre-trained models that do not generalize well to new identities. Aiming to address these limitations, we present a novel view synthesis framework to generate realistic renders from unseen views of any human captured from a single-view sensor with sparse RGB-D, similar to a low-cost depth camera, and without actor-specific models. We propose an architecture to learn dense features in novel views obtained by sphere-based neural rendering, and create complete renders using a global context inpainting model. Additionally, an enhancer network leverages the overall fidelity, even in occluded areas from the original view, producing crisp renders with fine details. We show our method generates high-quality novel views of synthetic and real human actors given a single sparse RGB-D input. It generalizes to unseen identities, new poses and faithfully reconstructs facial expressions. Our approach outperforms prior human view synthesis methods and is robust to different levels of input sparsity.

SP-GPT2: Semantics Improvement in Vietnamese Poetry Generation

Automatic text generation has garnered growing attention in recent years as an essential step towards computer creativity. Generative Pretraining Transformer 2 (GPT2) is one of the state of the art approaches that have excellent successes. In this paper, we took the first step to investigate the power of GPT2 in traditional Vietnamese poetry generation. In the earlier time, our experiment with base GPT2 was quite good at generating the poem in the proper template. Though it can learn the patterns, including rhyme and tone rules, from the training data, like almost all other text generation approaches, the poems generated still has a topic drift and semantic inconsistency. To improve the cohesion within the poems, we proposed a new model SP-GPT2 (semantic poem GPT2) which was built on the top GPT2 model and an additional loss to constrain context throughout the entire poem. For better evaluation, we examined the methods by both automatic quantitative evaluation and human evaluation. Both automatic and human evaluation demonstrated that our approach can generate poems that have better cohesion without losing the quality due to additional loss. At the same time, we are the pioneers of this topic. We released the first computational scoring module for poems generated in the template containing the style rule dictionary. Additionally, we are the first to publish a Luc-Bat dataset, including 87609 Luc Bat poems, which is equivalent to about 2.6 million sentences, combined with about 83579 poems in other styles was also published for further exploration. The code is available at https://github.com/fsoft-ailab/Poem-Generator

Lightweight Monocular Depth with a Novel Neural Architecture Search Method

This paper presents a novel neural architecture search method, called LiDNAS, for generating lightweight monocular depth estimation models. Unlike previous neural architecture search (NAS) approaches, where finding optimized networks are computationally highly demanding, the introduced novel Assisted Tabu Search leads to efficient architecture exploration. Moreover, we construct the search space on a pre-defined backbone network to balance layer diversity and search space size. The LiDNAS method outperforms the state-of-the-art NAS approach, proposed for disparity and depth estimation, in terms of search efficiency and output model performance. The LiDNAS optimized models achieve results superior to compact depth estimation state-of-the-art on NYU-Depth-v2, KITTI, and ScanNet, while being 7%-500% more compact in size, i.e the number of model parameters.

Monocular Depth Estimation Primed by Salient Point Detection and Normalized Hessian Loss

Deep neural networks have recently thrived on single image depth estimation. That being said, current developments on this topic highlight an apparent compromise between accuracy and network size. This work proposes an accurate and lightweight framework for monocular depth estimation based on a self-attention mechanism stemming from salient point detection. Specifically, we utilize a sparse set of keypoints to train a FuSaNet model that consists of two major components: Fusion-Net and Saliency-Net. In addition, we introduce a normalized Hessian loss term invariant to scaling and shear along the depth direction, which is shown to substantially improve the accuracy. The proposed method achieves state-of-the-art results on NYU-Depth-v2 and KITTI while using 3.1-38.4 times smaller model in terms of the number of parameters than baseline approaches. Experiments on the SUN-RGBD further demonstrate the generalizability of the proposed method.

Boosting Monocular Depth Estimation with Lightweight 3D Point Fusion

In this paper, we address the problem of fusing monocular depth estimation with a conventional multi-view stereo or SLAM to exploit the best of both worlds, that is, the accurate dense depth of the first one and lightweightness of the second one. More specifically, we use a conventional pipeline to produce a sparse 3D point cloud that is fed to a monocular depth estimation network to enhance its performance. In this way, we can achieve accuracy similar to multi-view stereo with a considerably smaller number of weights. We also show that even as few as 32 points is sufficient to outperform the best monocular depth estimation methods, and around 200 points to gain full advantage of the additional information. Moreover, we demonstrate the efficacy of our approach by integrating it with a SLAM system built-in on mobile devices.

RGBD-Net: Predicting color and depth images for novel views synthesis

We address the problem of novel view synthesis from an unstructured set of reference images. A new method called RGBD-Net is proposed to predict the depth map and the color images at the target pose in a multi-scale manner. The reference views are warped to the target pose to obtain multi-scale plane sweep volumes, which are then passed to our first module, a hierarchical depth regression network which predicts the depth map of the novel view. Second, a depth-aware generator network refines the warped novel views and renders the final target image. These two networks can be trained with or without depth supervision. In experimental evaluation, RGBD-Net not only produces novel views with higher quality than the previous state-of-the-art methods, but also the obtained depth maps enable reconstruction of more accurate 3D point clouds than the existing multi-view stereo methods. The results indicate that RGBD-Net generalizes well to previously unseen data.

A Multi-Objective Learning to re-Rank Approach to Optimize Online Marketplaces for Multiple Stakeholders

Multi-objective recommender systems address the difficult task of recommending items that are relevant to multiple, possibly conflicting, criteria. However these systems are most often designed to address the objective of one single stakeholder, typically, in online commerce, the consumers whose input and purchasing decisions ultimately determine the success of the recommendation systems. In this work, we address the multi-objective, multi-stakeholder, recommendation problem involving one or more objective(s) per stakeholder. In addition to the consumer stakeholder, we also consider two other stakeholders; the suppliers who provide the goods and services for sale and the intermediary who is responsible for helping connect consumers to suppliers via its recommendation algorithms. We analyze the multi-objective, multi-stakeholder, problem from the point of view of the online marketplace intermediary whose objective is to maximize its commission through its recommender system. We define a multi-objective problem relating all our three stakeholders which we solve with a novel learning-to-re-rank approach that makes use of a novel regularization function based on the Kendall tau correlation metric and its kernel version; given an initial ranking of item recommendations built for the consumer, we aim to re-rank it such that the new ranking is also optimized for the secondary objectives while staying close to the initial ranking. We evaluate our approach on a real-world dataset of hotel recommendations provided by Expedia where we show the effectiveness of our approach against a business-rules oriented baseline model.