Generative Graph Perturbations for Scene Graph Prediction

Inferring objects and their relationships from an image is useful in many applications at the intersection of vision and language. Due to a long tail data distribution, the task is challenging, with the inevitable appearance of zero-shot compositions of objects and relationships at test time. Current models often fail to properly understand a scene in such cases, as during training they only observe a tiny fraction of the distribution corresponding to the most frequent compositions. This motivates us to study whether increasing the diversity of the training distribution, by generating replacement for parts of real scene graphs, can lead to better generalization? We employ generative adversarial networks (GANs) conditioned on scene graphs to generate augmented visual features. To increase their diversity, we propose several strategies to perturb the conditioning. One of them is to use a language model, such as BERT, to synthesize plausible yet still unlikely scene graphs. By evaluating our model on Visual Genome, we obtain both positive and negative results. This prompts us to make several observations that can potentially lead to further improvements.

Graph Density-Aware Losses for Novel Compositions in Scene Graph Generation

Scene graph generation (SGG) aims to predict graph-structured descriptions of input images, in the form of objects and relationships between them. This task is becoming increasingly useful for progress at the interface of vision and language. Here, it is important - yet challenging - to perform well on novel (zero-shot) or rare (few-shot) compositions of objects and relationships. In this paper, we identify two key issues that limit such generalization. Firstly, we show that the standard loss used in this task is unintentionally a function of scene graph density. This leads to the neglect of individual edges in large sparse graphs during training, even though these contain diverse few-shot examples that are important for generalization. Secondly, the frequency of relationships can create a strong bias in this task, such that a blind model predicting the most frequent relationship achieves good performance. Consequently, some state-of-the-art models exploit this bias to improve results. We show that such models can suffer the most in their ability to generalize to rare compositions, evaluating two different models on the Visual Genome dataset and its more recent, improved version, GQA. To address these issues, we introduce a density-normalized edge loss, which provides more than a two-fold improvement in certain generalization metrics. Compared to other works in this direction, our enhancements require only a few lines of code and no added computational cost. We also highlight the difficulty of accurately evaluating models using existing metrics, especially on zero/few shots, and introduce a novel weighted metric.

A Simple and Scalable Shape Representation for 3D Reconstruction

Deep learning applied to the reconstruction of 3D shapes has seen growing interest. A popular approach to 3D reconstruction and generation in recent years has been the CNN encoder-decoder model usually applied in voxel space. However, this often scales very poorly with the resolution limiting the effectiveness of these models. Several sophisticated alternatives for decoding to 3D shapes have been proposed typically relying on complex deep learning architectures for the decoder model. In this work, we show that this additional complexity is not necessary, and that we can actually obtain high quality 3D reconstruction using a linear decoder, obtained from principal component analysis on the signed distance function (SDF) of the surface. This approach allows easily scaling to larger resolutions. We show in multiple experiments that our approach is competitive with state-of-the-art methods. It also allows the decoder to be fine-tuned on the target task using a loss designed specifically for SDF transforms, obtaining further gains.

Few-Shot Single-View 3-D Object Reconstruction with Compositional Priors

The impressive performance of deep convolutional neural networks in single-view 3D reconstruction suggests that these models perform non-trivial reasoning about the 3D structure of the output space. However, recent work has challenged this belief, showing that complex encoder-decoder architectures perform similarly to nearest-neighbor baselines or simple linear decoder models that exploit large amounts of per category data in standard benchmarks. On the other hand settings where 3D shape must be inferred for new categories with few examples are more natural and require models that generalize about shapes. In this work we demonstrate experimentally that naive baselines do not apply when the goal is to learn to reconstruct novel objects using very few examples, and that in a \emph{few-shot} learning setting, the network must learn concepts that can be applied to new categories, avoiding rote memorization. To address deficiencies in existing approaches to this problem, we propose three approaches that efficiently integrate a class prior into a 3D reconstruction model, allowing to account for intra-class variability and imposing an implicit compositional structure that the model should learn. Experiments on the popular ShapeNet database demonstrate that our method significantly outperform existing baselines on this task in the few-shot setting.

Online Learned Continual Compression with Stacked Quantization Module

We introduce and study the problem of Online Continual Compression, where one attempts to learn to compress and store a representative dataset from a non i.i.d data stream, while only observing each sample once. This problem is highly relevant for downstream online continual learning tasks, as well as standard learning methods under resource constrained data collection. To address this we propose a new architecture which Stacks Quantization Modules (SQM), consisting of a series of discrete autoencoders, each equipped with their own memory. Every added module is trained to reconstruct the latent space of the previous module using fewer bits, allowing the learned representation to become more compact as training progresses. This modularity has several advantages: 1) moderate compressions are quickly available early in training, which is crucial for remembering the early tasks, 2) as more data needs to be stored, earlier data becomes more compressed, freeing memory, 3) unlike previous methods, our approach does not require pretraining, even on challenging datasets. We show several potential applications of this method. We first replace the episodic memory used in Experience Replay with SQM, leading to significant gains on standard continual learning benchmarks using a fixed memory budget. We then apply our method to online compression of larger images like those from Imagenet, and show that it is also effective with other modalities, such as LiDAR data.

Online Continual Learning with Maximally Interfered Retrieval

Continual learning, the setting where a learning agent is faced with a never ending stream of data, continues to be a great challenge for modern machine learning systems. In particular the online or "single-pass through the data" setting has gained attention recently as a natural setting that is difficult to tackle. Methods based on replay, either generative or from a stored memory, have been shown to be effective approaches for continual learning, matching or exceeding the state of the art in a number of standard benchmarks. These approaches typically rely on randomly selecting samples from the replay memory or from a generative model, which is suboptimal. In this work we consider a controlled sampling of memories for replay. We retrieve the samples which are most interfered, i.e. whose prediction will be most negatively impacted by the foreseen parameters update. We show a formulation for this sampling criterion in both the generative replay and the experience replay setting, producing consistent gains in performance and greatly reduced forgetting.

VideoNavQA: Bridging the Gap between Visual and Embodied Question Answering

Embodied Question Answering (EQA) is a recently proposed task, where an agent is placed in a rich 3D environment and must act based solely on its egocentric input to answer a given question. The desired outcome is that the agent learns to combine capabilities such as scene understanding, navigation and language understanding in order to perform complex reasoning in the visual world. However, initial advancements combining standard vision and language methods with imitation and reinforcement learning algorithms have shown EQA might be too complex and challenging for these techniques. In order to investigate the feasibility of EQA-type tasks, we build the VideoNavQA dataset that contains pairs of questions and videos generated in the House3D environment. The goal of this dataset is to assess question-answering performance from nearly-ideal navigation paths, while considering a much more complete variety of questions than current instantiations of the EQA task. We investigate several models, adapted from popular VQA methods, on this new benchmark. This establishes an initial understanding of how well VQA-style methods can perform within this novel EQA paradigm.

Decoupled Greedy Learning of CNNs

A commonly cited inefficiency of neural network training by back-propagation is the update locking problem: each layer must wait for the signal to propagate through the network before updating. We consider and analyze a training procedure, Decoupled Greedy Learning (DGL), that addresses this problem more effectively and at scales beyond those of previous solutions. It is based on a greedy relaxation of the joint training objective, recently shown to be effective in the context of Convolutional Neural Networks (CNNs) on large-scale image classification. We consider an optimization of this objective that permits us to decouple the layer training, allowing for layers or modules in networks to be trained with a potentially linear parallelization in layers. We show theoretically and empirically that this approach converges. In addition, we empirically find that it can lead to better generalization than sequential greedy optimization and even standard end-to-end back-propagation. We show that an extension of this approach to asynchronous settings, where modules can operate with large communication delays, is possible with the use of a replay buffer. We demonstrate the effectiveness of DGL on the CIFAR-10 datasets against alternatives and on the large-scale ImageNet dataset, where we are able to effectively train VGG and ResNet-152 models.