Tsuhan Chen

Dexter Neo, Stefan Winkler, Tsuhan Chen

We present a new loss function that addresses the out-of-distribution (OOD) calibration problem. While many objective functions have been proposed to effectively calibrate models in-distribution, our findings show that they do not always fare well OOD. Based on the Principle of Maximum Entropy, we incorporate helpful statistical constraints observed during training, delivering better model calibration without sacrificing accuracy. We provide theoretical analysis and show empirically that our method works well in practice, achieving state-of-the-art calibration on both synthetic and real-world benchmarks.

Yunwei Bai, Ying Kiat Tan, Shiming Chen et al.

Few-shot learning (FSL) commonly requires a model to identify images (queries) that belong to classes unseen during training, based on a few labelled samples of the new classes (support set) as reference. So far, plenty of algorithms involve training data augmentation to improve the generalization capability of FSL models, but outlier queries or support images during inference can still pose great generalization challenges. In this work, to reduce the bias caused by the outlier samples, we generate additional test-class samples by combining original samples with suitable train-class samples via a generative image combiner. Then, we obtain averaged features via an augmentor, which leads to more typical representations through the averaging. We experimentally and theoretically demonstrate the effectiveness of our method, obtaining a test accuracy improvement proportion of around 10\% (e.g., from 46.86\% to 53.28\%) for trained FSL models. Importantly, given a pretrained image combiner, our method is training-free for off-the-shelf FSL models, whose performance can be improved without extra datasets nor further training of the models themselves. Codes are available at https://github.com/WendyBaiYunwei/FSL-Rectifier-Pub.

Dexter Neo, Tsuhan Chen

We present a novel extension to the family of Soft Actor-Critic (SAC) algorithms. We argue that based on the Maximum Entropy Principle, discrete SAC can be further improved via additional statistical constraints derived from a surrogate critic policy. Furthermore, our findings suggests that these constraints provide an added robustness against potential domain shifts, which are essential for safe deployment of reinforcement learning agents in the real-world. We provide theoretical analysis and show empirical results on low data regimes for both in-distribution and out-of-distribution variants of Atari 2600 games.

Yunwei Bai, Ying Kiat Tan, Yao Shu et al.

Diffusion models can be challenged in the low signal-to-noise regime, where they have to make pixel-level predictions despite the presence of high noise. The geometric intuition is akin to using the finest stroke for oil painting throughout, which may be ineffective. We therefore study stroke-size control as a controlled intervention that changes the effective roughness of the supervised target, predictions and perturbations across timesteps, in an attempt to ease the low signal-to-noise challenge. We analyze the advantages and trade-offs of the intervention both theoretically and empirically. Code will be released.

Yunwei Bai, Ying Kiat Tan, Yao Shu et al.

Few-shot learning (FSL) challenges model generalization to novel classes based on just a few shots of labeled examples, a testbed where traditional test-time augmentations fail to be effective. We introduce 1S-DAug, a one-shot generative augmentation operator that synthesizes diverse yet faithful variants from just one example image at test time. 1S-DAug couples traditional geometric perturbations with controlled noise injection and a denoising diffusion process conditioned on the original image. The generated images are then encoded and aggregated, alongside the original image, into a combined representation for more robust FSL predictions. Integrated as a training-free model-agnostic plugin, 1S-DAug consistently improves FSL across standard benchmarks of 4 different datasets without any model parameter update, including achieving up to 20\% proportional accuracy improvement on the miniImagenet 5-way-1-shot benchmark. Code will be released.

Dexter Neo, Tsuhan Chen

Large Vision-Language Models (LVLMs) have made remarkable developments along with the recent surge of large language models. Despite their advancements, LVLMs have a tendency to generate plausible yet inaccurate or inconsistent information based on the provided source content. This phenomenon, also known as ``hallucinations" can have serious downstream implications during the deployment of LVLMs. To address this, we present VORD a simple and effective method that alleviates hallucinations by calibrating token predictions based on ordinal relationships between modified image pairs. VORD is presented in two forms: 1.) a minimalist training-free variant which eliminates implausible tokens from modified image pairs, and 2.) a trainable objective function that penalizes unlikely tokens. Our experiments demonstrate that VORD delivers better calibration and effectively mitigates object hallucinations on a wide-range of LVLM benchmarks.

Dexter Neo, Tsuhan Chen

We present a soft benchmark for calibrating facial expression recognition (FER). While prior works have focused on identifying affective states, we find that FER models are uncalibrated. This is particularly true when out-of-distribution (OOD) shifts further exacerbate the ambiguity of facial expressions. While most OOD benchmarks provide hard labels, we argue that the ground-truth labels for evaluating FER models should be soft in order to better reflect the ambiguity behind facial behaviours. Our framework proposes soft labels that closely approximates the average information loss based on different types of OOD shifts. Finally, we show the benefits of calibration on five state-of-the-art FER algorithms tested on our benchmark.

Jiuxiang Gu, Jianfei Cai, Gang Wang et al.

The existing image captioning approaches typically train a one-stage sentence decoder, which is difficult to generate rich fine-grained descriptions. On the other hand, multi-stage image caption model is hard to train due to the vanishing gradient problem. In this paper, we propose a coarse-to-fine multi-stage prediction framework for image captioning, composed of multiple decoders each of which operates on the output of the previous stage, producing increasingly refined image descriptions. Our proposed learning approach addresses the difficulty of vanishing gradients during training by providing a learning objective function that enforces intermediate supervisions. Particularly, we optimize our model with a reinforcement learning approach which utilizes the output of each intermediate decoder's test-time inference algorithm as well as the output of its preceding decoder to normalize the rewards, which simultaneously solves the well-known exposure bias problem and the loss-evaluation mismatch problem. We extensively evaluate the proposed approach on MSCOCO and show that our approach can achieve the state-of-the-art performance.

Jiuxiang Gu, Gang Wang, Jianfei Cai et al.

Language Models based on recurrent neural networks have dominated recent image caption generation tasks. In this paper, we introduce a Language CNN model which is suitable for statistical language modeling tasks and shows competitive performance in image captioning. In contrast to previous models which predict next word based on one previous word and hidden state, our language CNN is fed with all the previous words and can model the long-range dependencies of history words, which are critical for image captioning. The effectiveness of our approach is validated on two datasets MS COCO and Flickr30K. Our extensive experimental results show that our method outperforms the vanilla recurrent neural network based language models and is competitive with the state-of-the-art methods.

Yuka Kihara, Matvey Soloviev, Tsuhan Chen

We present a new algorithm for multi-region segmentation of 2D images with objects that may partially occlude each other. Our algorithm is based on the observation hat human performance on this task is based both on prior knowledge about plausible shapes and taking into account the presence of occluding objects whose shape is already known - once an occluded region is identified, the shape prior can be used to guess the shape of the missing part. We capture the former aspect using a deep learning model of shape; for the latter, we simultaneously minimize the energy of all regions and consider only unoccluded pixels for data agreement. Existing algorithms incorporating object shape priors consider every object separately in turn and can't distinguish genuine deviation from the expected shape from parts missing due to occlusion. We show that our method significantly improves on the performance of a representative algorithm, as evaluated on both preprocessed natural and synthetic images. Furthermore, on the synthetic images, we recover the ground truth segmentation with good accuracy.

Amandianeze O. Nwana, Tsuhan Chen

One common trend in image tagging research is to focus on visually relevant tags, and this tends to ignore the personal and social aspect of tags, especially on photoblogging websites such as Flickr. Previous work has correctly identified that many of the tags that users provide on images are not visually relevant (i.e. representative of the salient content in the image) and they go on to treat such tags as noise, ignoring that the users chose to provide those tags over others that could have been more visually relevant. Another common assumption about user generated tags for images is that the order of these tags provides no useful information for the prediction of tags on future images. This assumption also tends to define usefulness in terms of what is visually relevant to the image. For general tagging or labeling applications that focus on providing visual information about image content, these assumptions are reasonable, but when considering personalized image tagging applications, these assumptions are at best too rigid, ignoring user choice and preferences. We challenge the aforementioned assumptions, and provide a machine learning approach to the problem of personalized image tagging with the following contributions: 1.) We reformulate the personalized image tagging problem as a search/retrieval ranking problem, 2.) We leverage the order of tags, which does not always reflect visual relevance, provided by the user in the past as a cue to their tag preferences, similar to click data, 3.) We propose a technique to augment sparse user tag data (semi-supervision), and 4.) We demonstrate the efficacy of our method on a subset of Flickr images, showing improvement over previous state-of-art methods.

Amandianeze O. Nwana, Tsuhan Chen

What makes a person pick certain tags over others when tagging an image? Does the order that a person presents tags for a given image follow an implicit bias that is personal? Can these biases be used to improve existing automated image tagging systems? We show that tag ordering, which has been largely overlooked by the image tagging community, is an important cue in understanding user tagging behavior and can be used to improve auto-tagging systems. Inspired by the assumption that people order their tags, we propose a new way of measuring tag preferences, and also propose a new personalized tagging objective function that explicitly considers a user's preferred tag orderings. We also provide a (partially) greedy algorithm that produces good solutions to our new objective and under certain conditions produces an optimal solution. We validate our method on a subset of Flickr images that spans 5000 users, over 5200 tags, and over 90,000 images. Our experiments show that exploiting personalized tag orders improves the average performance of state-of-art approaches both on per-image and per-user bases.

Jiuxiang Gu, Zhenhua Wang, Jason Kuen et al.

In the last few years, deep learning has led to very good performance on a variety of problems, such as visual recognition, speech recognition and natural language processing. Among different types of deep neural networks, convolutional neural networks have been most extensively studied. Leveraging on the rapid growth in the amount of the annotated data and the great improvements in the strengths of graphics processor units, the research on convolutional neural networks has been emerged swiftly and achieved state-of-the-art results on various tasks. In this paper, we provide a broad survey of the recent advances in convolutional neural networks. We detailize the improvements of CNN on different aspects, including layer design, activation function, loss function, regularization, optimization and fast computation. Besides, we also introduce various applications of convolutional neural networks in computer vision, speech and natural language processing.

Dong Ki Kim, Tsuhan Chen

Autonomous indoor navigation of Micro Aerial Vehicles (MAVs) possesses many challenges. One main reason is that GPS has limited precision in indoor environments. The additional fact that MAVs are not able to carry heavy weight or power consuming sensors, such as range finders, makes indoor autonomous navigation a challenging task. In this paper, we propose a practical system in which a quadcopter autonomously navigates indoors and finds a specific target, i.e., a book bag, by using a single camera. A deep learning model, Convolutional Neural Network (ConvNet), is used to learn a controller strategy that mimics an expert pilot's choice of action. We show our system's performance through real-time experiments in diverse indoor locations. To understand more about our trained network, we use several visualization techniques.

Amandianeze O Nwana, Salman Avestimehr, Tsuhan Chen

Current works on Information Centric Networking assume the spectrum of caching strategies under the Least Recently/ Frequently Used (LRFU) scheme as the de-facto standard, due to the ease of implementation and easier analysis of such strategies. In this paper we predict the popularity distribution of YouTube videos within a campus network. We explore two broad approaches in predicting the popularity of videos in the network: consensus approaches based on aggregate behavior in the network, and social approaches based on the information diffusion over an implicit network. We measure the performance of our approaches under a simple caching framework by picking the k most popular videos according to our predicted distribution and calculating the hit rate on the cache. We develop our approach by first incorporating video inter-arrival time (based on the power-law distribution governing the transmission time between two receivers of the same message in scale-free networks) to the baseline (LRFU), then combining with an information diffusion model over the inferred latent social graph that governs diffusion of videos in the network. We apply techniques from latent social network inference to learn the sharing probabilities between users in the network and apply a virus propagation model borrowed from mathematical epidemiology to estimate the number of times a video will be accessed in the future. Our approach gives rise to a 14% hit rate improvement over the baseline.