All NVIDIA Inception startups enter the program as Community members. By invitation, select startups may be elevated to Premier status if they can demonstrate that they have closed at least their Series A, are in production, use our tools and platforms and pass an internal review. NVIDIA Inception Premier members benefit from a dedicated relationship manager, DevRel support, increased marketing, and access to Premier-only events (such as business connections or VC introductions).

We present a framework for learning multimodal representations from unlabeled data using convolution-free Transformer architectures. Specifically, our Video-Audio-Text Transformer (VATT) takes raw signals as inputs and extracts multimodal representations that are rich enough to benefit a variety of downstream tasks. We train VATT end-to-end from scratch using multimodal contrastive losses and evaluate its performance by the downstream tasks of video action recognition, audio event classification, image classification, and text-to-video retrieval. Furthermore, we study a modality-agnostic single-backbone Transformer by sharing weights among the three modalities. We show that the convolution-free VATT outperforms state-of-the-art ConvNet-based architectures in the downstream tasks. Especially, VATT's vision Transformer achieves the top-1 accuracy of 82.1% on Kinetics-400, 83.6% on Kinetics-600,and 41.1% on Moments in Time, new records while avoiding supervised pre-training. Transferring to image classification leads to 78.7% top-1 accuracy on ImageNet compared to 64.7% by training the same Transformer from scratch, showing the generalizability of our model despite the domain gap between videos and images. VATT's audio Transformer also sets a new record on waveform-based audio event recognition by achieving the mAP of 39.4% on AudioSet without any supervised pre-training.

Deep probabilistic generative models have achieved incredible success in many fields of application. Among such models, variational autoencoders (VAEs) have proved their ability in modeling a generative process by learning a latent representation of the input. In this paper, we propose a novel VAE defined in the quaternion domain, which exploits the properties of quaternion algebra to improve performance while significantly reducing the number of parameters required by the network. The success of the proposed quaternion VAE with respect to traditional VAEs relies on the ability to leverage the internal relations between quaternion-valued input features and on the properties of second-order statistics which allow to define the latent variables in the augmented quaternion domain. In order to show the advantages due to such properties, we define a plain convolutional VAE in the quaternion domain and we evaluate its performance with respect to its real-valued counterpart on the CelebA face dataset.

Deep neural networks (DNNs) are increasingly important in safety-critical systems, for example in their perception layer to analyze images. Unfortunately, there is a lack of methods to ensure the functional safety of DNN-based components. We observe three major challenges with existing practices regarding DNNs in safety-critical systems: (1) scenarios that are underrepresented in the test set may lead to serious safety violation risks, but may, however, remain unnoticed; (2) characterizing such high-risk scenarios is critical for safety analysis; (3) retraining DNNs to address these risks is poorly supported when causes of violations are difficult to determine. To address these problems in the context of DNNs analyzing images, we propose HUDD, an approach that automatically supports the identification of root causes for DNN errors. HUDD identifies root causes by applying a clustering algorithm to heatmaps capturing the relevance of every DNN neuron on the DNN outcome. Also, HUDD retrains DNNs with images that are automatically selected based on their relatedness to the identified image clusters. We evaluated HUDD with DNNs from the automotive domain. HUDD was able to identify all the distinct root causes of DNN errors, thus supporting safety analysis. Also, our retraining approach has shown to be more effective at improving DNN accuracy than existing approaches.

Natural language-based vehicle retrieval is a task to find a target vehicle within a given image based on a natural language description as a query. This technology can be applied to various areas including police searching for a suspect vehicle. However, it is challenging due to the ambiguity of language descriptions and the difficulty of processing multi-modal data. To tackle this problem, we propose a deep neural network called SBNet that performs natural language-based segmentation for vehicle retrieval. We also propose two task-specific modules to improve performance: a substitution module that helps features from different domains to be embedded in the same space and a future prediction module that learns temporal information. SBnet has been trained using the CityFlow-NL dataset that contains 2,498 tracks of vehicles with three unique natural language descriptions each and tested 530 unique vehicle tracks and their corresponding query sets. SBNet achieved a significant improvement over the baseline in the natural language-based vehicle tracking track in the AI City Challenge 2021.

Modern machine learning increasingly requires training on a large collection of data from multiple sources, not all of which can be trusted. A particularly concerning scenario is when a small fraction of poisoned data changes the behavior of the trained model when triggered by an attacker-specified watermark. Such a compromised model will be deployed unnoticed as the model is accurate otherwise. There have been promising attempts to use the intermediate representations of such a model to separate corrupted examples from clean ones. However, these defenses work only when a certain spectral signature of the poisoned examples is large enough for detection. There is a wide range of attacks that cannot be protected against by the existing defenses. We propose a novel defense algorithm using robust covariance estimation to amplify the spectral signature of corrupted data. This defense provides a clean model, completely removing the backdoor, even in regimes where previous methods have no hope of detecting the poisoned examples. Code and pre-trained models are available at https://github.com/SewoongLab/spectre-defense .

We present the system that we have developed for the identification and verification of pills using images that are taken by the VeriMedi device. The VeriMedi device is an Internet of Things device that takes pictures of a filled pill vial from the bottom of the vial and uses the solution that is presented in this research to identify the pills in the vials. The solution has two serially connected deep learning solutions which do segmentation and identification. The segmentation solution creates the masks for each pill in the vial image by using the Mask R-CNN model, then segments and crops the pills and blurs the background. After that, the segmented pill images are sent to the identification solution where a Deep Metric Learning model that is trained with Proxy Anchor Loss (PAL) function generates embedding vectors for each pill image. The generated embedding vectors are fed into a one-layer fully connected network that is trained with the exact solution to predict each single pill image. Then, the aggregation/verification function aggregates the multiple predictions coming from multiple single pill images and verifies the correctness of the final prediction with respect to predefined rules. Besides, we enhanced the PAL with a better proxy initialization that increased the performance of the models and let the model learn the new classes of images continually without retraining the model with the whole dataset. When the model that is trained with initial classes is retrained only with new classes, the accuracy of the model increases for both old and new classes. The identification solution that we have presented in this research can also be reused for other problem domains which require continual learning and/or Fine-Grained Visual Categorization.

Vulnerability detection has always been the most important task in the field of software security. With the development of technology, in the face of massive source code, automated analysis and detection of vulnerabilities has become a current research hotspot. For special text files such as source code, using some of the hottest NLP technologies to build models and realize the automatic analysis and detection of source code has become one of the most anticipated studies in the field of vulnerability detection. This article does a brief survey of some recent new documents and technologies, such as CodeBERT, and summarizes the previous technologies.

We develop a deep generative model built on a fully differentiable simulator for multi-agent trajectory prediction. Agents are modeled with conditional recurrent variational neural networks (CVRNNs), which take as input an ego-centric birdview image representing the current state of the world and output an action, consisting of steering and acceleration, which is used to derive the subsequent agent state using a kinematic bicycle model. The full simulation state is then differentiably rendered for each agent, initiating the next time step. We achieve state-of-the-art results on the INTERACTION dataset, using standard neural architectures and a standard variational training objective, producing realistic multi-modal predictions without any ad-hoc diversity-inducing losses. We conduct ablation studies to examine individual components of the simulator, finding that both the kinematic bicycle model and the continuous feedback from the birdview image are crucial for achieving this level of performance. We name our model ITRA, for "Imagining the Road Ahead".

Adding to the literature on the data-driven detection of bid-rigging cartels, we propose a novel approach based on deep learning (a subfield of artificial intelligence) that flags cartel participants based on their pairwise bidding interactions with other firms. More concisely, we combine a so-called convolutional neural network for image recognition with graphs that in a pairwise manner plot the normalized bid values of some reference firm against the normalized bids of any other firms participating in the same tenders as the reference firm. Based on Japanese and Swiss procurement data, we construct such graphs for both collusive and competitive episodes (i.e when a bid-rigging cartel is or is not active) and use a subset of graphs to train the neural network such that it learns distinguishing collusive from competitive bidding patterns. We use the remaining graphs to test the neural network's out-of-sample performance in correctly classifying collusive and competitive bidding interactions. We obtain a very decent average accuracy of around 90% or slightly higher when either applying the method within Japanese, Swiss, or mixed data (in which Swiss and Japanese graphs are pooled). When using data from one country for training to test the trained model's performance in the other country (i.e. transnationally), predictive performance decreases (likely due to institutional differences in procurement procedures across countries), but often remains satisfactorily high. All in all, the generally quite high accuracy of the convolutional neural network despite being trained in a rather small sample of a few 100 graphs points to a large potential of deep learning approaches for flagging and fighting bid-rigging cartels.