Vision-language (VL) pretrained models have achieved impressive performance on multimodal reasoning and zero-shot recognition tasks. Many of these VL models are pretrained on unlabeled image and caption pairs from the internet. In this paper, we study whether representations of primitive concepts--such as colors, shapes, or the attributes of object parts--emerge automatically within these pretrained VL models. We propose a two-step framework, Compositional Concept Mapping (CompMap), to investigate this. CompMap first asks a VL model to generate primitive concept activations with text prompts, and then learns to construct a composition model that maps the primitive concept activations (e.g. the likelihood of black tail or red wing) to composite concepts (e.g. a red-winged blackbird). We show that a composition model can be reliably learn from ground truth primitive concepts. We thus hypothesize that if primitive concepts indeed emerge in a VL pretrained model, its primitive concept activations can be used to learn a composition model similar to the one designed by experts. We propose a quantitative metric to measure the degree of similarity, and refer to the metric as the interpretability metric. We also measure the classification accuracy when using the primitive concept activations and the learned composition model to predict the composite concepts, and refer to it as the usefulness metric. Our study reveals that state-of-the-art VL pretrained models learn primitive concepts that are highly useful for fine-grained visual recognition on the CUB dataset, and compositional generalization tasks on the MIT-States dataset. However, we observe that the learned composition models have low interpretability in our qualitative analyses. Our results reveal the limitations of existing VL models, and the necessity of pretraining objectives that encourage the acquisition of primitive concepts.
Data augmentation is widely used for training a neural network given little labeled data. A common practice of augmentation training is applying a composition of multiple transformations sequentially to the data. Existing augmentation methods such as RandAugment randomly sample from a list of pre-selected transformations, while methods such as AutoAugment apply advanced search to optimize over an augmentation set of size $k^d$, which is the number of transformation sequences of length $d$, given a list of $k$ transformations. In this paper, we design efficient algorithms whose running time complexity is much faster than the worst-case complexity of $O(k^d)$, provably. We propose a new algorithm to search for a binary tree-structured composition of $k$ transformations, where each tree node corresponds to one transformation. The binary tree generalizes sequential augmentations, such as the SimCLR augmentation scheme for contrastive learning. Using a top-down, recursive search procedure, our algorithm achieves a runtime complexity of $O(2^d k)$, which is much faster than $O(k^d)$ as $k$ increases above $2$. We apply our algorithm to tackle data distributions with heterogeneous subpopulations by searching for one tree in each subpopulation and then learning a weighted combination, resulting in a forest of trees. We validate our proposed algorithms on numerous graph and image datasets, including a multi-label graph classification dataset we collected. The dataset exhibits significant variations in the sizes of graphs and their average degrees, making it ideal for studying data augmentation. We show that our approach can reduce the computation cost by 43% over existing search methods while improving performance by 4.3%. The tree structures can be used to interpret the relative importance of each transformation, such as identifying the important transformations on small vs. large graphs.
State-of-the-art machine learning systems are vulnerable to small perturbations to their input, where ``small'' is defined according to a threat model that assigns a positive threat to each perturbation. Most prior works define a task-agnostic, isotropic, and global threat, like the $\ell_p$ norm, where the magnitude of the perturbation fully determines the degree of the threat and neither the direction of the attack nor its position in space matter. However, common corruptions in computer vision, such as blur, compression, or occlusions, are not well captured by such threat models. This paper proposes a novel threat model called \texttt{Projected Displacement} (PD) to study robustness beyond existing isotropic and global threat models. The proposed threat model measures the threat of a perturbation via its alignment with \textit{unsafe directions}, defined as directions in the input space along which a perturbation of sufficient magnitude changes the ground truth class label. Unsafe directions are identified locally for each input based on observed training data. In this way, the PD threat model exhibits anisotropy and locality. Experiments on Imagenet-1k data indicate that, for any input, the set of perturbations with small PD threat includes \textit{safe} perturbations of large $\ell_p$ norm that preserve the true label, such as noise, blur and compression, while simultaneously excluding \textit{unsafe} perturbations that alter the true label. Unlike perceptual threat models based on embeddings of large-vision models, the PD threat model can be readily computed for arbitrary classification tasks without pre-training or finetuning. Further additional task annotation such as sensitivity to image regions or concept hierarchies can be easily integrated into the assessment of threat and thus the PD threat model presents practitioners with a flexible, task-driven threat specification.
Recognizing objects and scenes are two challenging but essential tasks in image understanding. In particular, the use of RGB-D sensors in handling these tasks has emerged as an important area of focus for better visual understanding. Meanwhile, deep neural networks, specifically convolutional neural networks (CNNs), have become widespread and have been applied to many visual tasks by replacing hand-crafted features with effective deep features. However, it is an open problem how to exploit deep features from a multi-layer CNN model effectively. In this paper, we propose a novel two-stage framework that extracts discriminative feature representations from multi-modal RGB-D images for object and scene recognition tasks. In the first stage, a pretrained CNN model has been employed as a backbone to extract visual features at multiple levels. The second stage maps these features into high level representations with a fully randomized structure of recursive neural networks (RNNs) efficiently. To cope with the high dimensionality of CNN activations, a random weighted pooling scheme has been proposed by extending the idea of randomness in RNNs. Multi-modal fusion has been performed through a soft voting approach by computing weights based on individual recognition confidences (i.e. SVM scores) of RGB and depth streams separately. This produces consistent class label estimation in final RGB-D classification performance. Extensive experiments verify that fully randomized structure in RNN stage encodes CNN activations to discriminative solid features successfully. Comparative experimental results on the popular Washington RGB-D Object and SUN RGB-D Scene datasets show that the proposed approach achieves superior or on-par performance compared to state-of-the-art methods both in object and scene recognition tasks. Code is available at https://github.com/acaglayan/CNN_randRNN.
Leveraging transfer learning has recently been shown to be an effective strategy for training large models with Differential Privacy (DP). Moreover, somewhat surprisingly, recent works have found that privately training just the last layer of a pre-trained model provides the best utility with DP. While past studies largely rely on algorithms like DP-SGD for training large models, in the specific case of privately learning from features, we observe that computational burden is low enough to allow for more sophisticated optimization schemes, including second-order methods. To that end, we systematically explore the effect of design parameters such as loss function and optimization algorithm. We find that, while commonly used logistic regression performs better than linear regression in the non-private setting, the situation is reversed in the private setting. We find that linear regression is much more effective than logistic regression from both privacy and computational aspects, especially at stricter epsilon values ($\epsilon < 1$). On the optimization side, we also explore using Newton's method, and find that second-order information is quite helpful even with privacy, although the benefit significantly diminishes with stricter privacy guarantees. While both methods use second-order information, least squares is effective at lower epsilons while Newton's method is effective at larger epsilon values. To combine the benefits of both, we propose a novel algorithm called DP-FC, which leverages feature covariance instead of the Hessian of the logistic regression loss and performs well across all $\epsilon$ values we tried. With this, we obtain new SOTA results on ImageNet-1k, CIFAR-100 and CIFAR-10 across all values of $\epsilon$ typically considered. Most remarkably, on ImageNet-1K, we obtain top-1 accuracy of 88\% under (8, $8 * 10^{-7}$)-DP and 84.3\% under (0.1, $8 * 10^{-7}$)-DP.
Purpose: This work addresses the detection of Helicobacter pylori (H. pylori) in histological images with immunohistochemical staining. This analysis is a time demanding task, currently done by an expert pathologist that visually inspects the samples. Given the effort required to localise the pathogen in images, a limited number of annotations might be available in an initial setting. Our goal is to design an approach that, using a limited set of annotations, is capable of obtaining results good enough to be used as a support tool. Methods: We propose to use autoencoders to learn the latent patterns of healthy patches and formulate a specific measure of the reconstruction error of the image in HSV space. ROC analysis is used to set the optimal threshold of this measure and the percentage of positive patches in a sample that determines the presence of H. pylori. Results: Our method has been tested on an own database of 245 Whole Slide Images (WSI) having 117 cases without H. pylori and different density of the bacteria in the remaining ones. The database has 1211 annotated patches, with only 163 positive patches. This dataset of positive annotations was used to train a baseline thresholding and an SVM using the features of a pre-trained RedNet18 and ViT models. A 10-fold cross-validation shows that our method has better performance with 91% accuracy, 86% sensitivity, 96% specificity and 0.97 AUC in the diagnosis of H. pylori. Conclusion: Unlike classification approaches, our shallow autoencoder with threshold adaptation for the detection of anomalous staining is able to achieve competitive results with a limited set of annotated data. This initial approach is good enough to be used as a guide for fast annotation of infected patches.
We propose focal modulation networks (FocalNets in short), where self-attention (SA) is completely replaced by a focal modulation mechanism for modeling token interactions in vision. Focal modulation comprises three components: (i) hierarchical contextualization, implemented using a stack of depth-wise convolutional layers, to encode visual contexts from short to long ranges, (ii) gated aggregation to selectively gather contexts for each query token based on its content, and (iii) element-wise modulation or affine transformation to inject the aggregated context into the query. Extensive experiments show FocalNets outperform the state-of-the-art SA counterparts (e.g., Swin and Focal Transformers) with similar computational costs on the tasks of image classification, object detection, and segmentation. Specifically, FocalNets with tiny and base size achieve 82.3% and 83.9% top-1 accuracy on ImageNet-1K. After pretrained on ImageNet-22K in 224 resolution, it attains 86.5% and 87.3% top-1 accuracy when finetuned with resolution 224 and 384, respectively. When transferred to downstream tasks, FocalNets exhibit clear superiority. For object detection with Mask R-CNN, FocalNet base trained with 1\times outperforms the Swin counterpart by 2.1 points and already surpasses Swin trained with 3\times schedule (49.0 v.s. 48.5). For semantic segmentation with UPerNet, FocalNet base at single-scale outperforms Swin by 2.4, and beats Swin at multi-scale (50.5 v.s. 49.7). Using large FocalNet and Mask2former, we achieve 58.5 mIoU for ADE20K semantic segmentation, and 57.9 PQ for COCO Panoptic Segmentation. Using huge FocalNet and DINO, we achieved 64.3 and 64.4 mAP on COCO minival and test-dev, respectively, establishing new SoTA on top of much larger attention-based models like Swinv2-G and BEIT-3. Code and checkpoints are available at https://github.com/microsoft/FocalNet.
This paper investigates the role of explicit thinking process in rule-based reinforcement fine-tuning (RFT) for MLLMs. We first propose CLS-RL for MLLM image classification, using verifiable rewards for fine-tuning. Experiments show CLS-RL significantly outperforms SFT and yields a cross-dataset generalization effect. We then rethink and question whether explicit thinking in RFT is always necessary. Challenging the convention that explicit thinking is crucial for the success of RFT, we introduce No-Thinking-RL, exploring RFT without thinking by introducing a simple equality accuracy reward. We evaluate No-Thinking-RL on 6 diverse tasks across different model sizes and types. Experimental results reveal three key findings: 1). Visual perception tasks do not require thinking during RFT, as No-Thinking-RL consistently outperforms or matches Thinking-based RFT across model sizes. 2).} Models with limited capabilities struggle to generate high-quality CoT for RFT, making Thinking-based RFT less effective than No-Thinking-RL. 3). There are inconsistencies between the answers in the thinking and answer tags for some responses of thinking-based RFT, which show lower accuracy than the overall accuracy. We hypothesize that explicit thinking before verifiable answers may hinder reward convergence and reduce performance. To test this hypothesis, we propose Think-After-Answer, which places thinking after the answer to mitigate this effect for experimental verification. Lastly, we conduct a pilot study to explore whether MLLMs can learn when to think during RFT, introducing an Adaptive-Thinking method. Experiments show that it converges to a specific prompt depending on model capability and task complexity, achieving comparable or better performance than both Thinking and No-Thinking-RL. This suggests MLLMs can adaptively decide to think or not based on their capabilities and task complexity.
The training of over-parameterized neural networks has received much study in recent literature. An important consideration is the regularization of over-parameterized networks due to their highly nonconvex and nonlinear geometry. In this paper, we study noise injection algorithms, which can regularize the Hessian of the loss, leading to regions with flat loss surfaces. Specifically, by injecting isotropic Gaussian noise into the weight matrices of a neural network, we can obtain an approximately unbiased estimate of the trace of the Hessian. However, naively implementing the noise injection via adding noise to the weight matrices before backpropagation presents limited empirical improvements. To address this limitation, we design a two-point estimate of the Hessian penalty, which injects noise into the weight matrices along both positive and negative directions of the random noise. In particular, this two-point estimate eliminates the variance of the first-order Taylor's expansion term on the Hessian. We show a PAC-Bayes generalization bound that depends on the trace of the Hessian (and the radius of the weight space), which can be measured from data. We conduct a detailed experimental study to validate our approach and show that it can effectively regularize the Hessian and improve generalization. First, our algorithm can outperform prior approaches on sharpness-reduced training, delivering up to a 2.4% test accuracy increase for fine-tuning ResNets on six image classification datasets. Moreover, the trace of the Hessian reduces by 15.8%, and the largest eigenvalue is reduced by 9.7% with our approach. We also find that the regularization of the Hessian can be combined with weight decay and data augmentation, leading to stronger regularization. Second, our approach remains effective for improving generalization in pretraining multimodal CLIP models and chain-of-thought fine-tuning.
In this research, I proposed a network structure for multi-view 3D object detection using camera-only data and a Bird's-Eye-View map. My work is based on a current key challenge domain adaptation and visual data transfer. Although many excellent camera-only 3D object detection has been continuously proposed, many research work risk dramatic performance drop when the networks are trained on the source domain but tested on a different target domain. Then I found it is very surprising that predictions on bounding boxes and classes are still replied to on 2D networks. Based on the domain gap assumption on various 3D datasets, I found they still shared a similar data extraction on the same BEV map size and camera data transfer. Therefore, to analyze the domain gap influence on the current method and to make good use of 3D space information among the dataset and the real world, I proposed a transfer learning method and Transformer construction to study the 3D object detection on NuScenes-mini and Lyft. Through multi-dataset training and a detection head from the Transformer, the network demonstrated good data migration performance and efficient detection performance by using 3D anchor query and 3D positional information. Relying on only a small amount of source data and the existing large model pre-training weights, the efficient network manages to achieve competitive results on the new target domain. Moreover, my study utilizes 3D information as available semantic information and 2D multi-view image features blending into the visual-language transfer design. In the final 3D anchor box prediction and object classification, my network achieved good results on standard metrics of 3D object detection, which differs from dataset-specific models on each training domain without any fine-tuning.
Neural Architecture Search (NAS) is quickly becoming the go-to approach to optimize the structure of Deep Learning (DL) models for complex tasks such as Image Classification or Object Detection. However, many other relevant applications of DL, especially at the edge, are based on time-series processing and require models with unique features, for which NAS is less explored. This work focuses in particular on Temporal Convolutional Networks (TCNs), a convolutional model for time-series processing that has recently emerged as a promising alternative to more complex recurrent architectures. We propose the first NAS tool that explicitly targets the optimization of the most peculiar architectural parameters of TCNs, namely dilation, receptive-field and number of features in each layer. The proposed approach searches for networks that offer good trade-offs between accuracy and number of parameters/operations, enabling an efficient deployment on embedded platforms. We test the proposed NAS on four real-world, edge-relevant tasks, involving audio and bio-signals. Results show that, starting from a single seed network, our method is capable of obtaining a rich collection of Pareto optimal architectures, among which we obtain models with the same accuracy as the seed, and 15.9-152x fewer parameters. Compared to three state-of-the-art NAS tools, ProxylessNAS, MorphNet and FBNetV2, our method explores a larger search space for TCNs (up to 10^12x) and obtains superior solutions, while requiring low GPU memory and search time. We deploy our NAS outputs on two distinct edge devices, the multicore GreenWaves Technology GAP8 IoT processor and the single-core STMicroelectronics STM32H7 microcontroller. With respect to the state-of-the-art hand-tuned models, we reduce latency and energy of up to 5.5x and 3.8x on the two targets respectively, without any accuracy loss.
Distribution shift over time occurs in many settings. Leveraging historical data is necessary to learn a model for the last time point when limited data is available in the final period, yet few methods have been developed specifically for this purpose. In this work, we construct a benchmark with different sequences of synthetic shifts to evaluate the effectiveness of 3 classes of methods that 1) learn from all data without adapting to the final period, 2) learn from historical data with no regard to the sequential nature and then adapt to the final period, and 3) leverage the sequential nature of historical data when tailoring a model to the final period. We call this benchmark Seq-to-Final to highlight the focus on using a sequence of time periods to learn a model for the final time point. Our synthetic benchmark allows users to construct sequences with different types of shift and compare different methods. We focus on image classification tasks using CIFAR-10 and CIFAR-100 as the base images for the synthetic sequences. We also evaluate the same methods on the Portraits dataset to explore the relevance to real-world shifts over time. Finally, we create a visualization to contrast the initializations and updates from different methods at the final time step. Our results suggest that, for the sequences in our benchmark, methods that disregard the sequential structure and adapt to the final time point tend to perform well. The approaches we evaluate that leverage the sequential nature do not offer any improvement. We hope that this benchmark will inspire the development of new algorithms that are better at leveraging sequential historical data or a deeper understanding of why methods that disregard the sequential nature are able to perform well.
In many machine learning tasks, learning a good representation of the data can be the key to building a well-performant solution. This is because most learning algorithms operate with the features in order to find models for the data. For instance, classification performance can improve if the data is mapped to a space where classes are easily separated, and regression can be facilitated by finding a manifold of data in the feature space. As a general rule, features are transformed by means of statistical methods such as principal component analysis, or manifold learning techniques such as Isomap or locally linear embedding. From a plethora of representation learning methods, one of the most versatile tools is the autoencoder. In this paper we aim to demonstrate how to influence its learned representations to achieve the desired learning behavior. To this end, we present a series of learning tasks: data embedding for visualization, image denoising, semantic hashing, detection of abnormal behaviors and instance generation. We model them from the representation learning perspective, following the state of the art methodologies in each field. A solution is proposed for each task employing autoencoders as the only learning method. The theoretical developments are put into practice using a selection of datasets for the different problems and implementing each solution, followed by a discussion of the results in each case study and a brief explanation of other six learning applications. We also explore the current challenges and approaches to explainability in the context of autoencoders. All of this helps conclude that, thanks to alterations in their structure as well as their objective function, autoencoders may be the core of a possible solution to many problems which can be modeled as a transformation of the feature space.
Large language models are effective at few-shot in-context learning (ICL). Recent advancements in multimodal foundation models have enabled unprecedentedly long context windows, presenting an opportunity to explore their capability to perform ICL with many more demonstrating examples. In this work, we evaluate the performance of multimodal foundation models scaling from few-shot to many-shot ICL. We benchmark GPT-4o and Gemini 1.5 Pro across 14 datasets spanning multiple domains (natural imagery, medical imagery, remote sensing, and molecular imagery) and tasks (image classification, visual QA, and object localization). We observe that many-shot ICL, including up to almost 2,000 demonstrating examples, leads to substantial improvements compared to few-shot (<100 examples) ICL across all of the datasets. Further, Gemini 1.5 Pro performance continues to improve log-linearly up to the maximum number of tested examples on many datasets. We also find open-weights multimodal foundation models like Llama 3.2-Vision do not benefit from the demonstrating examples, highlighting an important gap between open and closed multimodal foundation models. Given the high inference costs required for many-shot ICL, we also explore the impact of batching multiple queries in a single API call. We show that batching up to 50 queries can lead to performance improvements under zero-shot and many-shot ICL, with substantial gains in the zero-shot setting on multiple datasets, while drastically reducing per-query cost and latency. Finally, while GPT-4o and Gemini 1.5 Pro achieve similar zero-shot performance across the datasets, Gemini 1.5 Pro learns more quickly than GPT-4o on most datasets. Our results suggest that many-shot ICL could enable users to efficiently adapt multimodal foundation models to new applications and domains. Our codebase is publicly available at https://github.com/stanfordmlgroup/ManyICL .
In this paper we explore few-shot imitation learning for control problems, which involves learning to imitate a target policy by accessing a limited set of offline rollouts. This setting has been relatively under-explored despite its relevance to robotics and control applications. State-of-the-art methods developed to tackle few-shot imitation rely on meta-learning, which is expensive to train as it requires access to a distribution over tasks (rollouts from many target policies and variations of the base environment). Given this limitation we investigate an alternative approach, fine-tuning, a family of methods that pretrain on a single dataset and then fine-tune on unseen domain-specific data. Recent work has shown that fine-tuners outperform meta-learners in few-shot image classification tasks, especially when the data is out-of-domain. Here we evaluate to what extent this is true for control problems, proposing a simple yet effective baseline which relies on two stages: (i) training a base policy online via reinforcement learning (e.g. Soft Actor-Critic) on a single base environment, (ii) fine-tuning the base policy via behavioral cloning on a few offline rollouts of the target policy. Despite its simplicity this baseline is competitive with meta-learning methods on a variety of conditions and is able to imitate target policies trained on unseen variations of the original environment. Importantly, the proposed approach is practical and easy to implement, as it does not need any complex meta-training protocol. As a further contribution, we release an open source dataset called iMuJoCo (iMitation MuJoCo) consisting of 154 variants of popular OpenAI-Gym MuJoCo environments with associated pretrained target policies and rollouts, which can be used by the community to study few-shot imitation learning and offline reinforcement learning.
This paper focuses on Semi-Supervised Object Detection (SSOD). Knowledge Distillation (KD) has been widely used for semi-supervised image classification. However, adapting these methods for SSOD has the following obstacles. (1) The teacher model serves a dual role as a teacher and a student, such that the teacher predictions on unlabeled images may be very close to those of student, which limits the upper-bound of the student. (2) The class imbalance issue in SSOD hinders an efficient knowledge transfer from teacher to student. To address these problems, we propose a novel method Temporal Self-Ensembling Teacher (TSE-T) for SSOD. Differently from previous KD based methods, we devise a temporally evolved teacher model. First, our teacher model ensembles its temporal predictions for unlabeled images under stochastic perturbations. Second, our teacher model ensembles its temporal model weights with the student model weights by an exponential moving average (EMA) which allows the teacher gradually learn from the student. These self-ensembling strategies increase data and model diversity, thus improving teacher predictions on unlabeled images. Finally, we use focal loss to formulate consistency regularization term to handle the data imbalance problem, which is a more efficient manner to utilize the useful information from unlabeled images than a simple hard-thresholding method which solely preserves confident predictions. Evaluated on the widely used VOC and COCO benchmarks, the mAP of our method has achieved 80.73% and 40.52% on the VOC2007 test set and the COCO2014 minval5k set respectively, which outperforms a strong fully-supervised detector by 2.37% and 1.49%. Furthermore, our method sets the new state-of-the-art in SSOD on VOC2007 test set which outperforms the baseline SSOD method by 1.44%. The source code of this work is publicly available at http://github.com/syangdong/tse-t.
Recently, data-driven deep saliency models have achieved high performance and have outperformed classical saliency models, as demonstrated by results on datasets such as the MIT300 and SALICON. Yet, there remains a large gap between the performance of these models and the inter-human baseline. Some outstanding questions include what have these models learned, how and where they fail, and how they can be improved. This article attempts to answer these questions by analyzing the representations learned by individual neurons located at the intermediate layers of deep saliency models. To this end, we follow the steps of existing deep saliency models, that is borrowing a pre-trained model of object recognition to encode the visual features and learning a decoder to infer the saliency. We consider two cases when the encoder is used as a fixed feature extractor and when it is fine-tuned, and compare the inner representations of the network. To study how the learned representations depend on the task, we fine-tune the same network using the same image set but for two different tasks: saliency prediction versus scene classification. Our analyses reveal that: 1) some visual regions (e.g. head, text, symbol, vehicle) are already encoded within various layers of the network pre-trained for object recognition, 2) using modern datasets, we find that fine-tuning pre-trained models for saliency prediction makes them favor some categories (e.g. head) over some others (e.g. text), 3) although deep models of saliency outperform classical models on natural images, the converse is true for synthetic stimuli (e.g. pop-out search arrays), an evidence of significant difference between human and data-driven saliency models, and 4) we confirm that, after-fine tuning, the change in inner-representations is mostly due to the task and not the domain shift in the data.
Recent work in neural networks for image classification has seen a strong tendency towards increasing the spatial context. Whether achieved through large convolution kernels or self-attention, models scale poorly with the increased spatial context, such that the improved model accuracy often comes at significant costs. In this paper, we propose a module for studying the effective filter size of convolutional neural networks. To facilitate such a study, several challenges need to be addressed: 1) we need an effective means to train models with large filters (potentially as large as the input data) without increasing the number of learnable parameters 2) the employed convolution operation should be a plug-and-play module that can replace conventional convolutions in a CNN and allow for an efficient implementation in current frameworks 3) the study of filter sizes has to be decoupled from other aspects such as the network width or the number of learnable parameters 4) the cost of the convolution operation itself has to remain manageable i.e. we cannot naively increase the size of the convolution kernel. To address these challenges, we propose to learn the frequency representations of filter weights as neural implicit functions, such that the better scalability of the convolution in the frequency domain can be leveraged. Additionally, due to the implementation of the proposed neural implicit function, even large and expressive spatial filters can be parameterized by only a few learnable weights. Our analysis shows that, although the proposed networks could learn very large convolution kernels, the learned filters are well localized and relatively small in practice when transformed from the frequency to the spatial domain. We anticipate that our analysis of individually optimized filter sizes will allow for more efficient, yet effective, models in the future. https://github.com/GeJulia/NIFF.
The manual annotation for large-scale point clouds costs a lot of time and is usually unavailable in harsh real-world scenarios. Inspired by the great success of the pre-training and fine-tuning paradigm in both vision and language tasks, we argue that pre-training is one potential solution for obtaining a scalable model to 3D point cloud downstream tasks as well. In this paper, we, therefore, explore a new self-supervised learning method, called Mixing and Disentangling (MD), for 3D point cloud representation learning. As the name implies, we mix two input shapes and demand the model learning to separate the inputs from the mixed shape. We leverage this reconstruction task as the pretext optimization objective for self-supervised learning. There are two primary advantages: 1) Compared to prevailing image datasets, eg, ImageNet, point cloud datasets are de facto small. The mixing process can provide a much larger online training sample pool. 2) On the other hand, the disentangling process motivates the model to mine the geometric prior knowledge, eg, key points. To verify the effectiveness of the proposed pretext task, we build one baseline network, which is composed of one encoder and one decoder. During pre-training, we mix two original shapes and obtain the geometry-aware embedding from the encoder, then an instance-adaptive decoder is applied to recover the original shapes from the embedding. Albeit simple, the pre-trained encoder can capture the key points of an unseen point cloud and surpasses the encoder trained from scratch on downstream tasks. The proposed method has improved the empirical performance on both ModelNet-40 and ShapeNet-Part datasets in terms of point cloud classification and segmentation tasks. We further conduct ablation studies to explore the effect of each component and verify the generalization of our proposed strategy by harnessing different backbones.
In recent years, Convolutional Neural Networks (CNNs) have become the standard class of deep neural network for image processing, classification and segmentation tasks. However, the large strides in accuracy obtained by CNNs have been derived from increasing the complexity of network topologies, which incurs sizeable performance and energy penalties in the training and inference of CNNs. Many recent works have validated the effectiveness of parameter quantization, which consists in reducing the bit width of the network's parameters, to enable the attainment of considerable performance and energy efficiency gains without significantly compromising accuracy. However, it is difficult to compare the relative effectiveness of different quantization methods. To address this problem, we introduce RedBit, an open-source framework that provides a transparent, extensible and easy-to-use interface to evaluate the effectiveness of different algorithms and parameter configurations on network accuracy. We use RedBit to perform a comprehensive survey of five state-of-the-art quantization methods applied to the MNIST, CIFAR-10 and ImageNet datasets. We evaluate a total of 2300 individual bit width combinations, independently tuning the width of the network's weight and input activation parameters, from 32 bits down to 1 bit (e.g., 8/8, 2/2, 1/32, 1/1, for weights/activations). Upwards of 20000 hours of computing time in a pool of state-of-the-art GPUs were used to generate all the results in this paper. For 1-bit quantization, the accuracy losses for the MNIST, CIFAR-10 and ImageNet datasets range between [0.26%, 0.79%], [9.74%, 32.96%] and [10.86%, 47.36%] top-1, respectively. We actively encourage the reader to download the source code and experiment with RedBit, and to submit their own observed results to our public repository, available at https://github.com/IT-Coimbra/RedBit.