How can deep neural networks encode information that corresponds to words in human speech into raw acoustic data? This paper proposes two neural network architectures for modeling unsupervised lexical learning from raw acoustic inputs, ciwGAN (Categorical InfoWaveGAN) and fiwGAN (Featural InfoWaveGAN), that combine a Deep Convolutional GAN architecture for audio data (WaveGAN; arXiv:1705.07904) with an information theoretic extension of GAN -- InfoGAN (arXiv:1606.03657), and propose a new latent space structure that can model featural learning simultaneously with a higher level classification and allows for a very low-dimension vector representation of lexical items. Lexical learning is modeled as emergent from an architecture that forces a deep neural network to output data such that unique information is retrievable from its acoustic outputs. The networks trained on lexical items from TIMIT learn to encode unique information corresponding to lexical items in the form of categorical variables in their latent space. By manipulating these variables, the network outputs specific lexical items. The network occasionally outputs innovative lexical items that violate training data, but are linguistically interpretable and highly informative for cognitive modeling and neural network interpretability. Innovative outputs suggest that phonetic and phonological representations learned by the network can be productively recombined and directly paralleled to productivity in human speech: a fiwGAN network trained on `suit' and `dark' outputs innovative `start', even though it never saw `start' or even a [st] sequence in the training data. We also argue that setting latent featural codes to values well beyond training range results in almost categorical generation of prototypical lexical items and reveals underlying values of each latent code.
This paper studies the multi-modal recommendation problem, where the item multi-modality information (e.g., images and textual descriptions) is exploited to improve the recommendation accuracy. Besides the user-item interaction graph, existing state-of-the-art methods usually use auxiliary graphs (e.g., user-user or item-item relation graph) to augment the learned representations of users and/or items. These representations are often propagated and aggregated on auxiliary graphs using graph convolutional networks, which can be prohibitively expensive in computation and memory, especially for large graphs. Moreover, existing multi-modal recommendation methods usually leverage randomly sampled negative examples in Bayesian Personalized Ranking (BPR) loss to guide the learning of user/item representations, which increases the computational cost on large graphs and may also bring noisy supervision signals into the training process. To tackle the above issues, we propose a novel self-supervised multi-modal recommendation model, dubbed BM3, which requires neither augmentations from auxiliary graphs nor negative samples. Specifically, BM3 first bootstraps latent contrastive views from the representations of users and items with a simple dropout augmentation. It then jointly optimizes three multi-modal objectives to learn the representations of users and items by reconstructing the user-item interaction graph and aligning modality features under both inter- and intra-modality perspectives. BM3 alleviates both the need for contrasting with negative examples and the complex graph augmentation from an additional target network for contrastive view generation. We show BM3 outperforms prior recommendation models on three datasets with number of nodes ranging from 20K to 200K, while achieving a 2-9X reduction in training time. Our code is available at https://github.com/enoche/BM3.
Diffusion models have demonstrated remarkable capabilities in image synthesis and related generative tasks. Nevertheless, their practicality for real-world applications is constrained by substantial computational costs and latency issues. Quantization is a dominant way to compress and accelerate diffusion models, where post-training quantization (PTQ) and quantization-aware training (QAT) are two main approaches, each bearing its own properties. While PTQ exhibits efficiency in terms of both time and data usage, it may lead to diminished performance in low bit-width. On the other hand, QAT can alleviate performance degradation but comes with substantial demands on computational and data resources. In this paper, we introduce a data-free and parameter-efficient fine-tuning framework for low-bit diffusion models, dubbed EfficientDM, to achieve QAT-level performance with PTQ-like efficiency. Specifically, we propose a quantization-aware variant of the low-rank adapter (QALoRA) that can be merged with model weights and jointly quantized to low bit-width. The fine-tuning process distills the denoising capabilities of the full-precision model into its quantized counterpart, eliminating the requirement for training data. We also introduce scale-aware optimization and temporal learned step-size quantization to further enhance performance. Extensive experimental results demonstrate that our method significantly outperforms previous PTQ-based diffusion models while maintaining similar time and data efficiency. Specifically, there is only a 0.05 sFID increase when quantizing both weights and activations of LDM-4 to 4-bit on ImageNet 256x256. Compared to QAT-based methods, our EfficientDM also boasts a 16.2x faster quantization speed with comparable generation quality. Code is available at \href{https://github.com/ThisisBillhe/EfficientDM}{this hrl}.
Generative methods have recently seen significant improvements by generating in a lower-dimensional latent representation of the data. However, many of the generative methods applied in the latent space remain complex and difficult to train. Further, it is not entirely clear why transitioning to a lower-dimensional latent space can improve generative quality. In this work, we introduce a new and simple generative method grounded in topology theory -- Generative Topological Networks (GTNs) -- which also provides insights into why lower-dimensional latent-space representations might be better-suited for data generation. GTNs are simple to train -- they employ a standard supervised learning approach and do not suffer from common generative pitfalls such as mode collapse, posterior collapse or the need to pose constraints on the neural network architecture. We demonstrate the use of GTNs on several datasets, including MNIST, CelebA, CIFAR-10 and the Hands and Palm Images dataset by training GTNs on a lower-dimensional latent representation of the data. We show that GTNs can improve upon VAEs and that they are quick to converge, generating realistic samples in early epochs. Further, we use the topological considerations behind the development of GTNs to offer insights into why generative models may benefit from operating on a lower-dimensional latent space, highlighting the important link between the intrinsic dimension of the data and the dimension in which the data is generated. Particularly, we demonstrate that generating in high dimensional ambient spaces may be a contributing factor to out-of-distribution samples generated by diffusion models. We also highlight other topological properties that are important to consider when using and designing generative models. Our code is available at: https://github.com/alonalj/GTN
Text-to-image models are trained using large datasets collected by scraping image-text pairs from the internet. These datasets often include private, copyrighted, and licensed material. Training models on such datasets enables them to generate images with such content, which might violate copyright laws and individual privacy. This phenomenon is termed imitation -- generation of images with content that has recognizable similarity to its training images. In this work we study the relationship between a concept's frequency in the training dataset and the ability of a model to imitate it. We seek to determine the point at which a model was trained on enough instances to imitate a concept -- the imitation threshold. We posit this question as a new problem: Finding the Imitation Threshold (FIT) and propose an efficient approach that estimates the imitation threshold without incurring the colossal cost of training multiple models from scratch. We experiment with two domains -- human faces and art styles -- for which we create four datasets, and evaluate three text-to-image models which were trained on two pretraining datasets. Our results reveal that the imitation threshold of these models is in the range of 200-600 images, depending on the domain and the model. The imitation threshold can provide an empirical basis for copyright violation claims and acts as a guiding principle for text-to-image model developers that aim to comply with copyright and privacy laws. We release the code and data at \url{https://github.com/vsahil/MIMETIC-2.git} and the project's website is hosted at \url{https://how-many-van-goghs-does-it-take.github.io}.
Holistically understanding an object and its 3D movable parts through visual perception models is essential for enabling an autonomous agent to interact with the world. For autonomous driving, the dynamics and states of vehicle parts such as doors, the trunk, and the bonnet can provide meaningful semantic information and interaction states, which are essential to ensuring the safety of the self-driving vehicle. Existing visual perception models mainly focus on coarse parsing such as object bounding box detection or pose estimation and rarely tackle these situations. In this paper, we address this important autonomous driving problem by solving three critical issues. First, to deal with data scarcity, we propose an effective training data generation process by fitting a 3D car model with dynamic parts to vehicles in real images before reconstructing human-vehicle interaction (VHI) scenarios. Our approach is fully automatic without any human interaction, which can generate a large number of vehicles in uncommon states (VUS) for training deep neural networks (DNNs). Second, to perform fine-grained vehicle perception, we present a multi-task network for VUS parsing and a multi-stream network for VHI parsing. Third, to quantitatively evaluate the effectiveness of our data augmentation approach, we build the first VUS dataset in real traffic scenarios (e.g., getting on/out or placing/removing luggage). Experimental results show that our approach advances other baseline methods in 2D detection and instance segmentation by a big margin (over 8%). In addition, our network yields large improvements in discovering and understanding these uncommon cases. Moreover, we have released the source code, the dataset, and the trained model on Github (https://github.com/zongdai/EditingForDNN).
Image inpainting aims to fill the missing hole of the input. It is hard to solve this task efficiently when facing high-resolution images due to two reasons: (1) Large reception field needs to be handled for high-resolution image inpainting. (2) The general encoder and decoder network synthesizes many background pixels synchronously due to the form of the image matrix. In this paper, we try to break the above limitations for the first time thanks to the recent development of continuous implicit representation. In detail, we down-sample and encode the degraded image to produce the spatial-adaptive parameters for each spatial patch via an attentional Fast Fourier Convolution(FFC)-based parameter generation network. Then, we take these parameters as the weights and biases of a series of multi-layer perceptron(MLP), where the input is the encoded continuous coordinates and the output is the synthesized color value. Thanks to the proposed structure, we only encode the high-resolution image in a relatively low resolution for larger reception field capturing. Then, the continuous position encoding will be helpful to synthesize the photo-realistic high-frequency textures by re-sampling the coordinate in a higher resolution. Also, our framework enables us to query the coordinates of missing pixels only in parallel, yielding a more efficient solution than the previous methods. Experiments show that the proposed method achieves real-time performance on the 2048$\times$2048 images using a single GTX 2080 Ti GPU and can handle 4096$\times$4096 images, with much better performance than existing state-of-the-art methods visually and numerically. The code is available at: https://github.com/NiFangBaAGe/CoordFill.
Many of the commonly used datasets for face recognition development are collected from the internet without proper user consent. Due to the increasing focus on privacy in the social and legal frameworks, the use and distribution of these datasets are being restricted and strongly questioned. These databases, which have a realistically high variability of data per identity, have enabled the success of face recognition models. To build on this success and to align with privacy concerns, synthetic databases, consisting purely of synthetic persons, are increasingly being created and used in the development of face recognition solutions. In this work, we present a three-player generative adversarial network (GAN) framework, namely IDnet, that enables the integration of identity information into the generation process. The third player in our IDnet aims at forcing the generator to learn to generate identity-separable face images. We empirically proved that our IDnet synthetic images are of higher identity discrimination in comparison to the conventional two-player GAN, while maintaining a realistic intra-identity variation. We further studied the identity link between the authentic identities used to train the generator and the generated synthetic identities, showing very low similarities between these identities. We demonstrated the applicability of our IDnet data in training face recognition models by evaluating these models on a wide set of face recognition benchmarks. In comparison to the state-of-the-art works in synthetic-based face recognition, our solution achieved comparable results to a recent rendering-based approach and outperformed all existing GAN-based approaches. The training code and the synthetic face image dataset are publicly available ( https://github.com/fdbtrs/Synthetic-Face-Recognition ).
Recently, there has been a surge in face personalization techniques, benefiting from the advanced capabilities of pretrained text-to-image diffusion models. Among these, a notable method is Textual Inversion, which generates personalized images by inverting given images into textual embeddings. However, methods based on Textual Inversion still struggle with balancing the trade-off between reconstruction quality and editability. In this study, we examine this issue through the lens of initialization. Upon closely examining traditional initialization methods, we identified a significant disparity between the initial and learned embeddings in terms of both scale and orientation. The scale of the learned embedding can be up to 100 times greater than that of the initial embedding. Such a significant change in the embedding could increase the risk of overfitting, thereby compromising the editability. Driven by this observation, we introduce a novel initialization method, termed Cross Initialization, that significantly narrows the gap between the initial and learned embeddings. This method not only improves both reconstruction and editability but also reduces the optimization steps from 5000 to 320. Furthermore, we apply a regularization term to keep the learned embedding close to the initial embedding. We show that when combined with Cross Initialization, this regularization term can effectively improve editability. We provide comprehensive empirical evidence to demonstrate the superior performance of our method compared to the baseline methods. Notably, in our experiments, Cross Initialization is the only method that successfully edits an individual's facial expression. Additionally, a fast version of our method allows for capturing an input image in roughly 26 seconds, while surpassing the baseline methods in terms of both reconstruction and editability. Code will be made publicly available.
Pursuing artificial intelligence for biomedical science, a.k.a. AI Scientist, draws increasing attention, where one common approach is to build a copilot agent driven by Large Language Models (LLMs). However, to evaluate such systems, people either rely on direct Question-Answering (QA) to the LLM itself, or in a biomedical experimental manner. How to precisely benchmark biomedical agents from an AI Scientist perspective remains largely unexplored. To this end, we draw inspiration from one most important abilities of scientists, understanding the literature, and introduce BioKGBench. In contrast to traditional evaluation benchmark that only focuses on factual QA, where the LLMs are known to have hallucination issues, we first disentangle "Understanding Literature" into two atomic abilities, i) "Understanding" the unstructured text from research papers by performing scientific claim verification, and ii) Ability to interact with structured Knowledge-Graph Question-Answering (KGQA) as a form of "Literature" grounding. We then formulate a novel agent task, dubbed KGCheck, using KGQA and domain-based Retrieval-Augmented Generation (RAG) to identify the factual errors of existing large-scale knowledge graph databases. We collect over two thousand data for two atomic tasks and 225 high-quality annotated data for the agent task. Surprisingly, we discover that state-of-the-art agents, both daily scenarios and biomedical ones, have either failed or inferior performance on our benchmark. We then introduce a simple yet effective baseline, dubbed BKGAgent. On the widely used popular knowledge graph, we discover over 90 factual errors which provide scenarios for agents to make discoveries and demonstrate the effectiveness of our approach. The code and data are available at https://github.com/westlake-autolab/BioKGBench.
Accurate retrieval of surface solar radiation (SSR) from satellite imagery critically depends on estimating the background reflectance that a spaceborne sensor would observe under clear-sky conditions. Deviations from this baseline can then be used to detect cloud presence and guide radiative transfer models in inferring atmospheric attenuation. Operational retrieval algorithms typically approximate background reflectance using monthly statistics, assuming surface properties vary slowly relative to atmospheric conditions. However, this approach fails in mountainous regions where intermittent snow cover and changing snow surfaces are frequent. We propose an attention-based emulator for SSR retrieval that implicitly learns to infer clear-sky surface reflectance from raw satellite image sequences. Built on the Temporo-Spatial Vision Transformer, our approach eliminates the need for hand-crafted features such as explicit albedo maps or cloud masks. The emulator is trained on instantaneous SSR estimates from the HelioMont algorithm over Switzerland, a region characterized by complex terrain and dynamic snow cover. Inputs include multi-spectral SEVIRI imagery from the Meteosat Second Generation platform, augmented with static topographic features and solar geometry. The target variable is HelioMont's SSR, computed as the sum of its direct and diffuse horizontal irradiance components, given at a spatial resolution of 1.7 km. We show that, when provided a sufficiently long temporal context, the model matches the performances of albedo-informed models, highlighting the model's ability to internally learn and exploit latent surface reflectance dynamics. Our geospatial analysis shows this effect is most powerful in mountainous regions and improves generalization in both simple and complex topographic settings. Code and datasets are publicly available at https://github.com/frischwood/HeMu-dev.git
Deriving robust control policies for realistic urban navigation scenarios is not a trivial task. In an end-to-end approach, these policies must map high-dimensional images from the vehicle's cameras to low-level actions such as steering and throttle. While pure Reinforcement Learning (RL) approaches are based exclusively on engineered rewards, Generative Adversarial Imitation Learning (GAIL) agents learn from expert demonstrations while interacting with the environment, which favors GAIL on tasks for which a reward signal is difficult to derive, such as autonomous driving. However, training deep networks directly from raw images on RL tasks is known to be unstable and troublesome. To deal with that, this work proposes a hierarchical GAIL-based architecture (hGAIL) which decouples representation learning from the driving task to solve the autonomous navigation of a vehicle. The proposed architecture consists of two modules: a GAN (Generative Adversarial Net) which generates an abstract mid-level input representation, which is the Bird's-Eye View (BEV) from the surroundings of the vehicle; and the GAIL which learns to control the vehicle based on the BEV predictions from the GAN as input. hGAIL is able to learn both the policy and the mid-level representation simultaneously as the agent interacts with the environment. Our experiments made in the CARLA simulation environment have shown that GAIL exclusively from cameras (without BEV) fails to even learn the task, while hGAIL, after training exclusively on one city, was able to autonomously navigate successfully in 98% of the intersections of a new city not used in training phase. Videos and code available at: https://sites.google.com/view/hgail
Recently, Deep reinforcement learning (DRL) models have shown promising results in solving routing problems. However, most DRL solvers are commonly proposed to solve node routing problems, such as the Traveling Salesman Problem (TSP). Meanwhile, there has been limited research on applying neural methods to arc routing problems, such as the Chinese Postman Problem (CPP), since they often feature irregular and complex solution spaces compared to TSP. To fill these gaps, this paper proposes a novel DRL framework to address the CPP with load-dependent costs (CPP-LC) (Corberan et al., 2018), which is a complex arc routing problem with load constraints. The novelty of our method is two-fold. First, we formulate the CPP-LC as a Markov Decision Process (MDP) sequential model. Subsequently, we introduce an autoregressive model based on DRL, namely Arc-DRL, consisting of an encoder and decoder to address the CPP-LC challenge effectively. Such a framework allows the DRL model to work efficiently and scalably to arc routing problems. Furthermore, we propose a new bio-inspired meta-heuristic solution based on Evolutionary Algorithm (EA) for CPP-LC. Extensive experiments show that Arc-DRL outperforms existing meta-heuristic methods such as Iterative Local Search (ILS) and Variable Neighborhood Search (VNS) proposed by (Corberan et al., 2018) on large benchmark datasets for CPP-LC regarding both solution quality and running time; while the EA gives the best solution quality with much more running time. We release our C++ implementations for metaheuristics such as EA, ILS and VNS along with the code for data generation and our generated data at https://github.com/HySonLab/Chinese_Postman_Problem
Video Question Answering (VideoQA) has emerged as a vital tool to evaluate agents' ability to understand human daily behaviors. Despite the recent success of large vision language models in many multi-modal tasks, complex situation reasoning over videos involving multiple human-object interaction events still remains challenging. In contrast, humans can easily tackle it by using a series of episode memories as anchors to quickly locate question-related key moments for reasoning. To mimic this effective reasoning strategy, we propose the Glance-Focus model. One simple way is to apply an action detection model to predict a set of actions as key memories. However, these actions within a closed set vocabulary are hard to generalize to various video domains. Instead of that, we train an Encoder-Decoder to generate a set of dynamic event memories at the glancing stage. Apart from using supervised bipartite matching to obtain the event memories, we further design an unsupervised memory generation method to get rid of dependence on event annotations. Next, at the focusing stage, these event memories act as a bridge to establish the correlation between the questions with high-level event concepts and low-level lengthy video content. Given the question, the model first focuses on the generated key event memory, then focuses on the most relevant moment for reasoning through our designed multi-level cross-attention mechanism. We conduct extensive experiments on four Multi-Event VideoQA benchmarks including STAR, EgoTaskQA, AGQA, and NExT-QA. Our proposed model achieves state-of-the-art results, surpassing current large models in various challenging reasoning tasks. The code and models are available at https://github.com/ByZ0e/Glance-Focus.
Effective weed control plays a crucial role in optimizing crop yield and enhancing agricultural product quality. However, the reliance on herbicide application not only poses a critical threat to the environment but also promotes the emergence of resistant weeds. Fortunately, recent advances in precision weed management enabled by ML and DL provide a sustainable alternative. Despite great progress, existing algorithms are mainly developed based on supervised learning approaches, which typically demand large-scale datasets with manual-labeled annotations, which is time-consuming and labor-intensive. As such, label-efficient learning methods, especially semi-supervised learning, have gained increased attention in the broader domain of computer vision and have demonstrated promising performance. These methods aim to utilize a small number of labeled data samples along with a great number of unlabeled samples to develop high-performing models comparable to the supervised learning counterpart trained on a large amount of labeled data samples. In this study, we assess the effectiveness of a semi-supervised learning framework for multi-class weed detection, employing two well-known object detection frameworks, namely FCOS and Faster-RCNN. Specifically, we evaluate a generalized student-teacher framework with an improved pseudo-label generation module to produce reliable pseudo-labels for the unlabeled data. To enhance generalization, an ensemble student network is employed to facilitate the training process. Experimental results show that the proposed approach is able to achieve approximately 76\% and 96\% detection accuracy as the supervised methods with only 10\% of labeled data in CottenWeedDet3 and CottonWeedDet12, respectively. We offer access to the source code, contributing a valuable resource for ongoing semi-supervised learning research in weed detection and beyond.
Large pre-trained language models have become popular for many applications and form an important backbone of many downstream tasks in natural language processing (NLP). Applying 'explainable artificial intelligence' (XAI) techniques to enrich such models' outputs is considered crucial for assuring their quality and shedding light on their inner workings. However, large language models are trained on a plethora of data containing a variety of biases, such as gender biases, affecting model weights and, potentially, behavior. Currently, it is unclear to what extent such biases also impact model explanations in possibly unfavorable ways. We create a gender-controlled text dataset, GECO, in which otherwise identical sentences appear in male and female forms. This gives rise to ground-truth 'world explanations' for gender classification tasks, enabling the objective evaluation of the correctness of XAI methods. We also provide GECOBench, a rigorous quantitative evaluation framework benchmarking popular XAI methods, applying them to pre-trained language models fine-tuned to different degrees. This allows us to investigate how pre-training induces undesirable bias in model explanations and to what extent fine-tuning can mitigate such explanation bias. We show a clear dependency between explanation performance and the number of fine-tuned layers, where XAI methods are observed to particularly benefit from fine-tuning or complete retraining of embedding layers. Remarkably, this relationship holds for models achieving similar classification performance on the same task. With that, we highlight the utility of the proposed gender-controlled dataset and novel benchmarking approach for research and development of novel XAI methods. All code including dataset generation, model training, evaluation and visualization is available at: https://github.com/braindatalab/gecobench
Object detection using LiDAR point clouds relies on a large amount of human-annotated samples when training the underlying detectors' deep neural networks. However, generating 3D bounding box annotation for a large-scale dataset could be costly and time-consuming. Alternatively, unsupervised domain adaptation (UDA) enables a given object detector to operate on a novel new data, with unlabeled training dataset, by transferring the knowledge learned from training labeled \textit{source domain} data to the new unlabeled \textit{target domain}. Pseudo label strategies, which involve training the 3D object detector using target-domain predicted bounding boxes from a pre-trained model, are commonly used in UDA. However, these pseudo labels often introduce noise, impacting performance. In this paper, we introduce the Domain Adaptive LIdar (DALI) object detection framework to address noise at both distribution and instance levels. Firstly, a post-training size normalization (PTSN) strategy is developed to mitigate bias in pseudo label size distribution by identifying an unbiased scale after network training. To address instance-level noise between pseudo labels and corresponding point clouds, two pseudo point clouds generation (PPCG) strategies, ray-constrained and constraint-free, are developed to generate pseudo point clouds for each instance, ensuring the consistency between pseudo labels and pseudo points during training. We demonstrate the effectiveness of our method on the publicly available and popular datasets KITTI, Waymo, and nuScenes. We show that the proposed DALI framework achieves state-of-the-art results and outperforms leading approaches on most of the domain adaptation tasks. Our code is available at \href{https://github.com/xiaohulugo/T-RO2024-DALI}{https://github.com/xiaohulugo/T-RO2024-DALI}.
Recent virtual try-on approaches have advanced by fine-tuning the pre-trained text-to-image diffusion models to leverage their powerful generative ability. However, the use of text prompts in virtual try-on is still underexplored. This paper tackles a text-editable virtual try-on task that changes the clothing item based on the provided clothing image while editing the wearing style (e.g., tucking style, fit) according to the text descriptions. In the text-editable virtual try-on, three key aspects exist: (i) designing rich text descriptions for paired person-clothing data to train the model, (ii) addressing the conflicts where textual information of the existing person's clothing interferes the generation of the new clothing, and (iii) adaptively adjust the inpainting mask aligned with the text descriptions, ensuring proper editing areas while preserving the original person's appearance irrelevant to the new clothing. To address these aspects, we propose PromptDresser, a text-editable virtual try-on model that leverages large multimodal model (LMM) assistance to enable high-quality and versatile manipulation based on generative text prompts. Our approach utilizes LMMs via in-context learning to generate detailed text descriptions for person and clothing images independently, including pose details and editing attributes using minimal human cost. Moreover, to ensure the editing areas, we adjust the inpainting mask depending on the text prompts adaptively. We found that our approach, utilizing detailed text prompts, not only enhances text editability but also effectively conveys clothing details that are difficult to capture through images alone, thereby enhancing image quality. Our code is available at https://github.com/rlawjdghek/PromptDresser.
Vision transformers have generated significant interest in the computer vision community because of their flexibility in exploiting contextual information, whether it is sharply confined local, or long range global. However, they are known to be data hungry. This has motivated the research in self-supervised transformer pretraining, which does not need to decode the semantic information conveyed by labels to link it to the image properties, but rather focuses directly on extracting a concise representation of the image data that reflects the notion of similarity, and is invariant to nuisance factors. The key vehicle for the self-learning process used by the majority of self-learning methods is the generation of multiple views of the training data and the creation of pretext tasks which use these views to define the notion of image similarity, and data integrity. However, this approach lacks the natural propensity to extract contextual information. We propose group masked model learning (GMML), a self-supervised learning (SSL) mechanism for pretraining vision transformers with the ability to extract the contextual information present in all the concepts in an image. GMML achieves this by manipulating randomly groups of connected tokens, ensuingly covering a meaningful part of a semantic concept, and then recovering the hidden semantic information from the visible part of the concept. GMML implicitly introduces a novel data augmentation process. Unlike most of the existing SSL approaches, GMML does not require momentum encoder, nor rely on careful implementation details such as large batches and gradient stopping, which are all artefacts of most of the current self-supervised learning techniques. The source code is publicly available for the community to train on bigger corpora: https://github.com/Sara-Ahmed/GMML.
This work addresses fair generative models. Dataset biases have been a major cause of unfairness in deep generative models. Previous work had proposed to augment large, biased datasets with small, unbiased reference datasets. Under this setup, a weakly-supervised approach has been proposed, which achieves state-of-the-art quality and fairness in generated samples. In our work, based on this setup, we propose a simple yet effective approach. Specifically, first, we propose fairTL, a transfer learning approach to learn fair generative models. Under fairTL, we pre-train the generative model with the available large, biased datasets and subsequently adapt the model using the small, unbiased reference dataset. We find that our fairTL can learn expressive sample generation during pre-training, thanks to the large (biased) dataset. This knowledge is then transferred to the target model during adaptation, which also learns to capture the underlying fair distribution of the small reference dataset. Second, we propose fairTL++, where we introduce two additional innovations to improve upon fairTL: (i) multiple feedback and (ii) Linear-Probing followed by Fine-Tuning (LP-FT). Taking one step further, we consider an alternative, challenging setup when only a pre-trained (potentially biased) model is available but the dataset that was used to pre-train the model is inaccessible. We demonstrate that our proposed fairTL and fairTL++ remain very effective under this setup. We note that previous work requires access to the large, biased datasets and is incapable of handling this more challenging setup. Extensive experiments show that fairTL and fairTL++ achieve state-of-the-art in both quality and fairness of generated samples. The code and additional resources can be found at bearwithchris.github.io/fairTL/.