News

Congratulations to Harry Yang for passing his defense on May 7, 2019! Let us hear what he would like to say about his defense and an abstract of his thesis.

“In the thesis, we tackle the problem of translating faces and bodies between different identities without paired training data: we cannot directly train a translation module using supervised signals in this case. Instead, we propose to train a conditional variational auto-encoder (CVAE) to disentangle different latent factors such as identity and expressions. In order to achieve effective disentanglement, we further use multi-view information such as keypoints and facial landmarks to train multiple CVAEs. By relying on these simplified representations of the data we are using a more easily disentangled representation to guide the disentanglement of image itself. Experiments demonstrate the effectiveness of our method in multiple face and body datasets. We also show that our model is a more robust image classifier and adversarial example detector comparing with traditional multi-class neural networks.

“To address the issue of scaling to new identities and also generate better-quality results, we further propose an alternative approach that uses self-supervised learning based on StyleGAN to factorize out different attributes of face images, such as hair color, facial expressions, skin color, and others. Using pre-trained StyleGAN combined with iterative style inference we can easily manipulate the facial expressions or combine the facial expressions of any two people, without the need of training a specific new model for each of the identity involved. This is one of the first scalable and high-quality approach for generating DeepFake data, which serves as a critical first step to learn a more robust and general classifier against adversarial examples.”

Harry also shared about his Ph.D. experience:

“Firstly, I would [...]

Our long-term goal is to build intelligent systems that can perceive their visual environment and understand the linguistic information, and further make an accurate translation inference to another language. However, most multi-modal translation algorithms are not significantly better than an off-the-shelf text-only machine translation (MT) model. There remains an open question about how translation models should take advantage of visual context, because from the perspective of information theory, the mutual information of two random variables I(X; Y) will always be no greater than I(X, Z; Y) where Z is the additional visual input. This conclusion makes us believe that the visual content will hopefully help the translation systems.

Since the standard paradigm of multi-modal translation always considers the problem as a supervised learning task, the parallel corpus is usually sufficient to train a good translation model, and the gain from the extra image input is very limited. We however argue that the text-only UMT is fundamentally an ill-posed problem, since there are potentially many ways to associate target with source sentences. Intuitively, since the visual content and language are closely related, the image can play the role of a pivot “language” to bridge the two languages without paralleled corpus, making the problem “more well-defined” by reducing the problem to supervised learning.

We tackle the unsupervised translation with a multi-modal framework which includes two sequence-to-sequence encoder-decoder models and one shared image feature extractor in order to achieve the unsupervised translation. We employ transformer in both the text encoder and decoder of our model and design a novel joint attention mechanism to simulate the relationships among the language and visual domains.

Succinctly, our contributions are three-fold:

We formulate the multi-modal MT problem as unsupervised setting that fits the real [...]

Congratulations to Ron Salloum for passing his PhD defense on April 10, 2019. His PhD thesis is entitled “A Data-Driven Approach to Image Splicing Localization”.

Abstract:

The availability of low-cost and user-friendly editing software has made it significantly easier to manipulate images. Thus, there has been an increasing interest in developing forensic techniques to detect and localize image manipulations or forgeries. Splicing, which is one of the most common types of image forgery, involves copying a region from one image (referred to as the donor image) and pasting it onto another image (referred to as the host image). Forgers often use splicing to give a false impression that there is an additional object present in the image, or to remove an object from the image.

Many of the current splicing detection methods only determine whether a given image has been spliced and do not attempt to localize the spliced region. Relatively few methods attempt to tackle the splicing localization problem, which refers to the problem of determining which pixels in an image have been manipulated as a result of a splicing operation.

In my dissertation, I present two different splicing localization methods that we have developed. The first is the Multi-task Fully Convolutional Network (MFCN), which is a neural-network-based method that outperforms previous methods on many datasets. The second proposed method is based on cPCA++ (where cPCA stands for contrastive Principal Component Analysis), which is a new data visualization and clustering technique that we have developed. The cPCA++ method is more efficient than the MFCN and achieves comparable performance.

PhD Experience:

Pursuing my PhD degree was a very challenging but rewarding experience. I really enjoyed my time in the Media Communications Laboratory and had the opportunity to work on exciting research projects. [...]

With the rise of visualization, animation and autonomous driving applications, the demand for 3D point cloud analysis and understanding has rapidly increased. Point Cloud is a kind of data obtained from lidar scanning which contains abundant 3D information. Our research directions about point cloud in autonomous driving are object detection, segmentation and classification.

Due to its unstructured and unordered properties, people usually transfer point cloud into other data types such as mesh, voxel and multi-view. But the transformation must cause information lost. Recently, several deep-learning-solutions such as PointNet/Pointnet++ [1, 2] tailored to point clouds provide a more efficient and flexible way to handle 3D data. Some successful results for object classification and parts and semantic scene segmentation have been demonstrated. However, object and scene understanding with Convolutional Neural Networks (CNNs) on 3D volumetric data is still limited due to its high memory requirement and computational cost. This brings a challenge for autonomous driving since it requires real-time and concise processing of the observed scenes and objects.

An interpretable CNN design based on the feedforward (FF) methodology [3] without any backpropagation (BP) was recently proposed by the Media Communications Lab at USC. The FF design offers a complementary approach to CNN filter weights selection. We are now designing a feed-forward (FF) network for both object classification and indoor scene segmentation. The advantages of the FF design methodology are multiple folds. It is completely interpretable. It demands much less training complexity and training data. Furthermore, it can be generalized to weakly supervised or unsupervised learning scenarios in a straightforward manner. The latter is extremely important in real world application scenarios since data labeling is very tedious and expensive.

References:

R. Qi, H. Su, K. Mo, and L. J. Guibas. [...]

Deep learning has shown its effectiveness in various computer vision tasks. However, a large amount of labeled data is usually needed for deep learning approaches. Active learning can help reduce the labeling efforts by choosing the most informative samples to label and thus achieves a comparable performance with less labeled data.

There are two major types of active learning strategy: uncertainty based and diversity based.

The core idea of uncertainty based methods is to label those samples that are most uncertain to the existing model trained on current labeled set. For example, an image with a prediction of 50 percent cat is empirically considered to be more valuable than an image with a prediction of 99 percent cat, where the former has larger uncertainty. Besides uncertainty metrics from information theory like entropy, Beluch et al. [1] proposes to use an ensemble to estimate the uncertainty of unlabeled images and achieves good results in ImageNet dataset.

In contrast, diversity based methods rely on an assumption that a more diverse set of images chosen as the training set can lead to better performance. Sener et al. [2] formalizes the active learning problem into a core-set problem and achieves competitive performance in CIFAR-10 dataset. Mixed-integer programming is used to solve their objective function.

Our current research focuses on balancing the two factors (uncertainty and diversity) in a explainable way.

References:
[1] Beluch, William H., et al. “The power of ensembles for active learning in image classification.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.
[2] Sener, Ozan, and Silvio Savarese. “Active learning for convolutional neural networks: A core-set approach.” (2018).

Author: Yeji Shen