Many fine-grained classification tasks, like rare animal identification, have limited training data and consequently classifiers trained on these datasets often fail to generalize to variations in the domain like changes in weather or location. As such, we explore how natural language descriptions of the domains seen in training data can be used with large vision models trained on diverse pretraining datasets to generate useful variations of the training data. We introduce ALIA (Automated Language-guided Image Augmentation), a method which utilizes large vision and language models to automatically generate natural language descriptions of a dataset's domains and augment the training data via language-guided image editing. To maintain data integrity, a model trained on the original dataset filters out minimal image edits and those which corrupt class-relevant information. The resulting dataset is visually consistent with the original training data and offers significantly enhanced diversity. On fine-grained and cluttered datasets for classification and detection, ALIA surpasses traditional data augmentation and text-to-image generated data by up to 15\%, often even outperforming equivalent additions of real data. Code is avilable at https://github.com/lisadunlap/ALIA.
Machine Learning (ML) has recently been a skyrocketing field in Computer Science. As computer hardware engineers, we are enthusiastic about hardware implementations of popular software ML architectures to optimize their performance, reliability, and resource usage. In this project, we designed a highly-configurable, real-time device for recognizing handwritten letters and digits using an Altera DE1 FPGA Kit. We followed various engineering standards, including IEEE-754 32-bit Floating-Point Standard, Video Graphics Array (VGA) display protocol, Universal Asynchronous Receiver-Transmitter (UART) protocol, and Inter-Integrated Circuit (I2C) protocols to achieve the project goals. These significantly improved our design in compatibility, reusability, and simplicity in verifications. Following these standards, we designed a 32-bit floating-point (FP) instruction set architecture (ISA). We developed a 5-stage RISC processor in System Verilog to manage image processing, matrix multiplications, ML classifications, and user interfaces. Three different ML architectures were implemented and evaluated on our design: Linear Classification (LC), a 784-64-10 fully connected neural network (NN), and a LeNet-like Convolutional Neural Network (CNN) with ReLU activation layers and 36 classes (10 for the digits and 26 for the case-insensitive letters). The training processes were done in Python scripts, and the resulting kernels and weights were stored in hex files and loaded into the FPGA's SRAM units. Convolution, pooling, data management, and various other ML features were guided by firmware in our custom assembly language. This paper documents the high-level design block diagrams, interfaces between each System Verilog module, implementation details of our software and firmware components, and further discussions on potential impacts.
Diffusion models are a class of flexible generative models trained with an approximation to the log-likelihood objective. However, most use cases of diffusion models are not concerned with likelihoods, but instead with downstream objectives such as human-perceived image quality or drug effectiveness. In this paper, we investigate reinforcement learning methods for directly optimizing diffusion models for such objectives. We describe how posing denoising as a multi-step decision-making problem enables a class of policy gradient algorithms, which we refer to as denoising diffusion policy optimization (DDPO), that are more effective than alternative reward-weighted likelihood approaches. Empirically, DDPO is able to adapt text-to-image diffusion models to objectives that are difficult to express via prompting, such as image compressibility, and those derived from human feedback, such as aesthetic quality. Finally, we show that DDPO can improve prompt-image alignment using feedback from a vision-language model without the need for additional data collection or human annotation.
Unsupervised learning has grown in popularity because of the difficulty of collecting annotated data and the development of modern frameworks that allow us to learn from unlabeled data. Existing studies, however, either disregard variations at different levels of similarity or only consider negative samples from one batch. We argue that image pairs should have varying degrees of similarity, and the negative samples should be allowed to be drawn from the entire dataset. In this work, we propose Search-based Unsupervised Visual Representation Learning (SUVR) to learn better image representations in an unsupervised manner. We first construct a graph from the image dataset by the similarity between images, and adopt the concept of graph traversal to explore positive samples. In the meantime, we make sure that negative samples can be drawn from the full dataset. Quantitative experiments on five benchmark image classification datasets demonstrate that SUVR can significantly outperform strong competing methods on unsupervised embedding learning. Qualitative experiments also show that SUVR can produce better representations in which similar images are clustered closer together than unrelated images in the latent space.
Can we customize a deep generative model which can generate images that can match the texture of some given image? When you see an image of a church, you may wonder if you can get similar pictures for that church. Here we present a method, for adapting GANs with one reference image, and then we can generate images that have similar textures to the given image. Specifically, we modify the weights of the pre-trained GAN model, guided by the reference image given by the user. We use a patch discriminator adversarial loss to encourage the output of the model to match the texture on the given image, also we use a laplacian adversarial loss to ensure diversity and realism, and alleviate the contradiction between the two losses. Experiments show that the proposed method can make the outputs of GANs match the texture of the given image as well as keep diversity and realism.
Diffusion models, such as Stable Diffusion, have shown incredible performance on text-to-image generation. Since text-to-image generation often requires models to generate visual concepts with fine-grained details and attributes specified in text prompts, can we leverage the powerful representations learned by pre-trained diffusion models for discriminative tasks such as image-text matching? To answer this question, we propose a novel approach, Discriminative Stable Diffusion (DSD), which turns pre-trained text-to-image diffusion models into few-shot discriminative learners. Our approach uses the cross-attention score of a Stable Diffusion model to capture the mutual influence between visual and textual information and fine-tune the model via attention-based prompt learning to perform image-text matching. By comparing DSD with state-of-the-art methods on several benchmark datasets, we demonstrate the potential of using pre-trained diffusion models for discriminative tasks with superior results on few-shot image-text matching.
Semi-supervised learning (SSL) has attracted much attention since it reduces the expensive costs of collecting adequate well-labeled training data, especially for deep learning methods. However, traditional SSL is built upon an assumption that labeled and unlabeled data should be from the same distribution e.g., classes and domains. However, in practical scenarios, unlabeled data would be from unseen classes or unseen domains, and it is still challenging to exploit them by existing SSL methods. Therefore, in this paper, we proposed a unified framework to leverage these unseen unlabeled data for open-scenario semi-supervised medical image classification. We first design a novel scoring mechanism, called dual-path outliers estimation, to identify samples from unseen classes. Meanwhile, to extract unseen-domain samples, we then apply an effective variational autoencoder (VAE) pre-training. After that, we conduct domain adaptation to fully exploit the value of the detected unseen-domain samples to boost semi-supervised training. We evaluated our proposed framework on dermatology and ophthalmology tasks. Extensive experiments demonstrate our model can achieve superior classification performance in various medical SSL scenarios.
Large pre-trained vision-language models have shown great prominence in transferring pre-acquired knowledge to various domains and downstream tasks with appropriate prompting or tuning. Existing prevalent tuning methods can be generally categorized into three genres: 1) prompt engineering by creating suitable prompt texts, which is time-consuming and requires domain expertise; 2) or simply fine-tuning the whole model, which is extremely inefficient; 3) prompt tuning through parameterized prompt embeddings with the text encoder. Nevertheless, all methods rely on the text encoder for bridging the modality gap between vision and language. In this work, we question the necessity of the cumbersome text encoder for a more lightweight and efficient tuning paradigm as well as more representative prompt embeddings closer to the image representations. To achieve this, we propose a Concept Embedding Search (ConES) approach by optimizing prompt embeddings -- without the need of the text encoder -- to capture the 'concept' of the image modality through a variety of task objectives. By dropping the text encoder, we are able to significantly speed up the learning process, \eg, from about an hour to just ten minutes in our experiments for personalized text-to-image generation without impairing the generation quality. Moreover, our proposed approach is orthogonal to current existing tuning methods since the searched concept embeddings can be further utilized in the next stage of fine-tuning the pre-trained large models for boosting performance. Extensive experiments show that our approach can beat the prompt tuning and textual inversion methods in a variety of downstream tasks including objection detection, instance segmentation, and image generation. Our approach also shows better generalization capability for unseen concepts in specialized domains, such as the medical domain.
Generative Adversarial Networks (GAN) is a model for data synthesis, which creates plausible data through the competition of generator and discriminator. Although GAN application to image synthesis is extensively studied, it has inherent limitations to natural language generation. Because natural language is composed of discrete tokens, a generator has difficulty updating its gradient through backpropagation; therefore, most text-GAN studies generate sentences starting with a random token based on a reward system. Thus, the generators of previous studies are pre-trained in an autoregressive way before adversarial training, causing data memorization that synthesized sentences reproduce the training data. In this paper, we synthesize sentences using a framework similar to the original GAN. More specifically, we propose Text Embedding Space Generative Adversarial Networks (TESGAN) which generate continuous text embedding spaces instead of discrete tokens to solve the gradient backpropagation problem. Furthermore, TESGAN conducts unsupervised learning which does not directly refer to the text of the training data to overcome the data memorization issue. By adopting this novel method, TESGAN can synthesize new sentences, showing the potential of unsupervised learning for text synthesis. We expect to see extended research combining Large Language Models with a new perspective of viewing text as an continuous space.
Mammogram image is important for breast cancer screening, and typically obtained in a dual-view form, i.e., cranio-caudal (CC) and mediolateral oblique (MLO), to provide complementary information. However, previous methods mostly learn features from the two views independently, which violates the clinical knowledge and ignores the importance of dual-view correlation. In this paper, we propose a dual-view correlation hybrid attention network (DCHA-Net) for robust holistic mammogram classification. Specifically, DCHA-Net is carefully designed to extract and reinvent deep features for the two views, and meanwhile to maximize the underlying correlations between them. A hybrid attention module, consisting of local relation and non-local attention blocks, is proposed to alleviate the spatial misalignment of the paired views in the correlation maximization. A dual-view correlation loss is introduced to maximize the feature similarity between corresponding strip-like regions with equal distance to the chest wall, motivated by the fact that their features represent the same breast tissues, and thus should be highly-correlated. Experimental results on two public datasets, i.e., INbreast and CBIS-DDSM, demonstrate that DCHA-Net can well preserve and maximize feature correlations across views, and thus outperforms the state-of-the-arts for classifying a whole mammogram as malignant or not.