Pre-trained models are nowadays a fundamental component of machine learning research. In continual learning, they are commonly used to initialize the model before training on the stream of non-stationary data. However, pre-training is rarely applied during continual learning. We formalize and investigate the characteristics of the continual pre-training scenario in both language and vision environments, where a model is continually pre-trained on a stream of incoming data and only later fine-tuned to different downstream tasks. We show that continually pre-trained models are robust against catastrophic forgetting and we provide strong empirical evidence supporting the fact that self-supervised pre-training is more effective in retaining previous knowledge than supervised protocols. Code is provided at https://github.com/AndreaCossu/continual-pretraining-nlp-vision .
Training models continually to detect and classify objects, from new classes and new domains, remains an open problem. In this work, we conduct a thorough analysis of why and how object detection models forget catastrophically. We focus on distillation-based approaches in two-stage networks; the most-common strategy employed in contemporary continual object detection work.Distillation aims to transfer the knowledge of a model trained on previous tasks -- the teacher -- to a new model -- the student -- while it learns the new task. We show that this works well for the region proposal network, but that wrong, yet overly confident teacher predictions prevent student models from effective learning of the classification head. Our analysis provides a foundation that allows us to propose improvements for existing techniques by detecting incorrect teacher predictions, based on current ground-truth labels, and by employing an adaptive Huber loss as opposed to the mean squared error for the distillation loss in the classification heads. We evidence that our strategy works not only in a class incremental setting, but also in domain incremental settings, which constitute a realistic context, likely to be the setting of representative real-world problems.
As natural images usually contain multiple objects, multi-label image classification is more applicable "in the wild" than single-label classification. However, exhaustively annotating images with every object of interest is costly and time-consuming. We aim to train multi-label classifiers from single-label annotations only. We show that adding a consistency loss, ensuring that the predictions of the network are consistent over consecutive training epochs, is a simple yet effective method to train multi-label classifiers in a weakly supervised setting. We further extend this approach spatially, by ensuring consistency of the spatial feature maps produced over consecutive training epochs, maintaining per-class running-average heatmaps for each training image. We show that this spatial consistency loss further improves the multi-label mAP of the classifiers. In addition, we show that this method overcomes shortcomings of the "crop" data-augmentation by recovering correct supervision signal even when most of the single ground truth object is cropped out of the input image by the data augmentation. We demonstrate gains of the consistency and spatial consistency losses over the binary cross-entropy baseline, and over competing methods, on MS-COCO and Pascal VOC. We also demonstrate improved multi-label classification mAP on ImageNet-1K using the ReaL multi-label validation set.
In the online continual learning paradigm, agents must learn from a changing distribution while respecting memory and compute constraints. Experience Replay (ER), where a small subset of past data is stored and replayed alongside new data, has emerged as a simple and effective learning strategy. In this work, we focus on the change in representations of observed data that arises when previously unobserved classes appear in the incoming data stream, and new classes must be distinguished from previous ones. We shed new light on this question by showing that applying ER causes the newly added classes' representations to overlap significantly with the previous classes, leading to highly disruptive parameter updates. Based on this empirical analysis, we propose a new method which mitigates this issue by shielding the learned representations from drastic adaptation to accommodate new classes. We show that using an asymmetric update rule pushes new classes to adapt to the older ones (rather than the reverse), which is more effective especially at task boundaries, where much of the forgetting typically occurs. Empirical results show significant gains over strong baselines on standard continual learning benchmarks
Visual question answering is a vision-and-language multimodal task, that aims at predicting answers given samples from the question and image modalities. Most recent methods focus on learning a good joint embedding space of images and questions, either by improving the interaction between these two modalities, or by making it a more discriminant space. However, how informative this joint space is, has not been well explored. In this paper, we propose a novel regularization for VQA models, Constrained Optimization using Barlow's theory (COB), that improves the information content of the joint space by minimizing the redundancy. It reduces the correlation between the learned feature components and thereby disentangles semantic concepts. Our model also aligns the joint space with the answer embedding space, where we consider the answer and image+question as two different `views' of what in essence is the same semantic information. We propose a constrained optimization policy to balance the categorical and redundancy minimization forces. When built on the state-of-the-art GGE model, the resulting model improves VQA accuracy by 1.4% and 4% on the VQA-CP v2 and VQA v2 datasets respectively. The model also exhibits better interpretability.
This paper attacks the problem of language-guided navigation in a new perspective by using novel semantic navigation maps, which enables robots to carry out natural language instructions and move to a target position based on the map observations. We break down this problem into parts and introduce three different modules to solve the corresponding subproblems. Our approach leverages map information to provide Deterministic Path Candidate Proposals to reduce the solution space. Different from traditional methods that predict robots' movements toward the target step-by-step, we design an attention-based Language Driven Discriminator to evaluate path candidates and determine the best path as the final result. To represent the map observations along a path for a better modality alignment, a novel Path Feature Encoding scheme tailored for semantic navigation maps is proposed. Unlike traditional methods that tend to produce cumulative errors or be stuck in local decisions, our method which plans paths based on global information can greatly alleviate these problems. The proposed approach has noticeable performance gains, especially in long-distance navigation cases. Also, its training efficiency is significantly higher than of other methods.
The gap between simulation and the real-world restrains many machine learning breakthroughs in computer vision and reinforcement learning from being applicable in the real world. In this work, we tackle this gap for the specific case of camera-based navigation, formulating it as following a visual cue in the foreground with arbitrary backgrounds. The visual cue in the foreground can often be simulated realistically, such as a line, gate or cone. The challenge then lies in coping with the unknown backgrounds and integrating both. As such, the goal is to train a visual agent on data captured in an empty simulated environment except for this foreground cue and test this model directly in a visually diverse real world. In order to bridge this big gap, we show it's crucial to combine following techniques namely: Randomized augmentation of the fore- and background, regularization with both deep supervision and triplet loss and finally abstraction of the dynamics by using waypoints rather than direct velocity commands. The various techniques are ablated in our experimental results both qualitatively and quantitatively finally demonstrating a successful transfer from simulation to the real world.
Recognizing human actions is fundamentally a spatio-temporal reasoning problem, and should be, at least to some extent, invariant to the appearance of the human and the objects involved. Motivated by this hypothesis, in this work, we take an object-centric approach to action recognition. Multiple works have studied this setting before, yet it remains unclear (i) how well a carefully crafted, spatio-temporal layout-based method can recognize human actions, and (ii) how, and when, to fuse the information from layout and appearance-based models. The main focus of this paper is compositional/few-shot action recognition, where we advocate the usage of multi-head attention (proven to be effective for spatial reasoning) over spatio-temporal layouts, i.e., configurations of object bounding boxes. We evaluate different schemes to inject video appearance information to the system, and benchmark our approach on background cluttered action recognition. On the Something-Else and Action Genome datasets, we demonstrate (i) how to extend multi-head attention for spatio-temporal layout-based action recognition, (ii) how to improve the performance of appearance-based models by fusion with layout-based models, (iii) that even on non-compositional background-cluttered video datasets, a fusion between layout- and appearance-based models improves the performance.
Feature pyramids have become ubiquitous in multi-scale computer vision tasks such as object detection. Based on their importance, we divide a computer vision network into three parts: a backbone (generating a feature pyramid), a core (refining the feature pyramid) and a head (generating the final output). Most existing networks operating on feature pyramids, named cores, are shallow and mostly focus on communication-based processing in the form of top-down and bottom-up operations. We present a new core architecture called Trident Pyramid Network (TPN), that allows for a deeper design and for a better balance between communication-based processing and self-processing. We show consistent improvements when using our TPN core on the COCO object detection benchmark, outperforming the popular BiFPN baseline by 1.5 AP. Additionally, we empirically show that it is more beneficial to put additional computation into the TPN core, rather than into the backbone, by outperforming a ResNet-101+FPN baseline with our ResNet-50+TPN network by 1.7 AP, while operating under similar computation budgets. This emphasizes the importance of performing computation at the feature pyramid level in modern-day object detection systems. Code will be released.
Active visual exploration aims to assist an agent with a limited field of view to understand its environment based on partial observations made by choosing the best viewing directions in the scene. Recent methods have tried to address this problem either by using reinforcement learning, which is difficult to train, or by uncertainty maps, which are task-specific and can only be implemented for dense prediction tasks. In this paper, we propose the Glimpse-Attend-and-Explore model which: (a) employs self-attention to guide the visual exploration instead of task-specific uncertainty maps; (b) can be used for both dense and sparse prediction tasks; and (c) uses a contrastive stream to further improve the representations learned. Unlike previous works, we show the application of our model on multiple tasks like reconstruction, segmentation and classification. Our model provides encouraging results while being less dependent on dataset bias in driving the exploration. We further perform an ablation study to investigate the features and attention learned by our model. Finally, we show that our self-attention module learns to attend different regions of the scene by minimizing the loss on the downstream task. Code: https://github.com/soroushseifi/glimpse-attend-explore.