James J. Clark

Danilo Vucetic, Mohammadreza Tayaranian, Maryam Ziaeefard et al.

Resource-constrained devices are increasingly the deployment targets of machine learning applications. Static models, however, do not always suffice for dynamic environments. On-device training of models allows for quick adaptability to new scenarios. With the increasing size of deep neural networks, as noted with the likes of BERT and other natural language processing models, comes increased resource requirements, namely memory, computation, energy, and time. Furthermore, training is far more resource intensive than inference. Resource-constrained on-device learning is thus doubly difficult, especially with large BERT-like models. By reducing the memory usage of fine-tuning, pre-trained BERT models can become efficient enough to fine-tune on resource-constrained devices. We propose Freeze And Reconfigure (FAR), a memory-efficient training regime for BERT-like models that reduces the memory usage of activation maps during fine-tuning by avoiding unnecessary parameter updates. FAR reduces fine-tuning time on the DistilBERT model and CoLA dataset by 30%, and time spent on memory operations by 47%. More broadly, reductions in metric performance on the GLUE and SQuAD datasets are around 1% on average.

Ali Edalati, Marzieh Tahaei, Ivan Kobyzev et al.

Fine-tuning a Pre-trained Language Model (PLM) on a specific downstream task has been a well-known paradigm in Natural Language Processing. However, with the ever-growing size of PLMs, training the entire model on several downstream tasks becomes very expensive and resource-hungry. Recently, different Parameter Efficient Tuning (PET) techniques are proposed to improve the efficiency of fine-tuning PLMs. One popular category of PET methods is the low-rank adaptation methods which insert learnable truncated SVD modules into the original model either sequentially or in parallel. However, low-rank decomposition suffers from limited representation power. In this work, we address this problem using the Kronecker product instead of the low-rank representation. We introduce KronA, a Kronecker product-based adapter module for efficient fine-tuning of Transformer-based PLMs. We apply the proposed methods for fine-tuning T5 on the GLUE benchmark to show that incorporating the Kronecker-based modules can outperform state-of-the-art PET methods.

Danilo Vucetic, Mohammadreza Tayaranian, Maryam Ziaeefard et al.

Fine-tuning BERT-based models is resource-intensive in memory, computation, and time. While many prior works aim to improve inference efficiency via compression techniques, e.g., pruning, these works do not explicitly address the computational challenges of training to downstream tasks. We introduce Learner modules and priming, novel methods for fine-tuning that exploit the overparameterization of pre-trained language models to gain benefits in convergence speed and resource utilization. Learner modules navigate the double bind of 1) training efficiently by fine-tuning a subset of parameters, and 2) training effectively by ensuring quick convergence and high metric scores. Our results on DistilBERT demonstrate that learners perform on par with or surpass the baselines. Learners train 7x fewer parameters than state-of-the-art methods on GLUE. On CoLA, learners fine-tune 20% faster, and have significantly lower resource utilization.

Manoosh Samiei, James J. Clark

Visual Search is referred to the task of finding a target object among a set of distracting objects in a visual display. In this paper, based on an independent analysis of the COCO-Search18 dataset, we investigate how the performance of human participants during visual search is affected by different parameters such as the size and eccentricity of the target object. We also study the correlation between the error rate of participants and search performance. Our studies show that a bigger and more eccentric target is found faster with fewer number of fixations. Our code for the graphics are publicly available at https://github.com/ManooshSamiei/COCOSearch18_Analysis.

Ibtihel Amara, Maryam Ziaeefard, Brett H. Meyer et al.

Knowledge distillation (KD) is an effective tool for compressing deep classification models for edge devices. However, the performance of KD is affected by the large capacity gap between the teacher and student networks. Recent methods have resorted to a multiple teacher assistant (TA) setting for KD, which sequentially decreases the size of the teacher model to relatively bridge the size gap between these models. This paper proposes a new technique called Curriculum Expert Selection for Knowledge Distillation (CES-KD) to efficiently enhance the learning of a compact student under the capacity gap problem. This technique is built upon the hypothesis that a student network should be guided gradually using stratified teaching curriculum as it learns easy (hard) data samples better and faster from a lower (higher) capacity teacher network. Specifically, our method is a gradual TA-based KD technique that selects a single teacher per input image based on a curriculum driven by the difficulty in classifying the image. In this work, we empirically verify our hypothesis and rigorously experiment with CIFAR-10, CIFAR-100, CINIC-10, and ImageNet datasets and show improved accuracy on VGG-like models, ResNets, and WideResNets architectures.

Manoosh Samiei, James J. Clark

Most studies in computational modeling of visual attention encompass task-free observation of images. Free-viewing saliency considers limited scenarios of daily life. Most visual activities are goal-oriented and demand a great amount of top-down attention control. Visual search task demands more top-down control of attention, compared to free-viewing. In this paper, we present two approaches to model visual attention and distraction of observers during visual search. Our first approach adapts a light-weight free-viewing saliency model to predict eye fixation density maps of human observers over pixels of search images, using a two-stream convolutional encoder-decoder network, trained and evaluated on COCO-Search18 dataset. This method predicts which locations are more distracting when searching for a particular target. Our network achieves good results on standard saliency metrics (AUC-Judd=0.95, AUC-Borji=0.85, sAUC=0.84, NSS=4.64, KLD=0.93, CC=0.72, SIM=0.54, and IG=2.59). Our second approach is object-based and predicts the distractor and target objects during visual search. Distractors are all objects except the target that observers fixate on during search. This method uses a Mask-RCNN segmentation network pre-trained on MS-COCO and fine-tuned on COCO-Search18 dataset. We release our segmentation annotations of targets and distractors in COCO-Search18 for three target categories: bottle, bowl, and car. The average scores over the three categories are: F1-score=0.64, MAP(iou:0.5)=0.57, MAR(iou:0.5)=0.73. Our implementation code in Tensorflow is publicly available at https://github.com/ManooshSamiei/Distraction-Visual-Search .

Ibtihel Amara, Nazanin Sepahvand, Brett H. Meyer et al.

We address the challenge of producing trustworthy and accurate compact models for edge devices. While Knowledge Distillation (KD) has improved model compression in terms of achieving high accuracy performance, calibration of these compact models has been overlooked. We introduce BD-KD (Balanced Divergence Knowledge Distillation), a framework for logit-based online KD. BD-KD enhances both accuracy and model calibration simultaneously, eliminating the need for post-hoc recalibration techniques, which add computational overhead to the overall training pipeline and degrade performance. Our method encourages student-centered training by adjusting the conventional online distillation loss on both the student and teacher losses, employing sample-wise weighting of forward and reverse Kullback-Leibler divergence. This strategy balances student network confidence and boosts performance. Experiments across CIFAR10, CIFAR100, TinyImageNet, and ImageNet datasets, and various architectures demonstrate improved calibration and accuracy compared to recent online KD methods.

Samrudhdhi B. Rangrej, Chetan L. Srinidhi, James J. Clark

Most hard attention models initially observe a complete scene to locate and sense informative glimpses, and predict class-label of a scene based on glimpses. However, in many applications (e.g., aerial imaging), observing an entire scene is not always feasible due to the limited time and resources available for acquisition. In this paper, we develop a Sequential Transformers Attention Model (STAM) that only partially observes a complete image and predicts informative glimpse locations solely based on past glimpses. We design our agent using DeiT-distilled and train it with a one-step actor-critic algorithm. Furthermore, to improve classification performance, we introduce a novel training objective, which enforces consistency between the class distribution predicted by a teacher model from a complete image and the class distribution predicted by our agent using glimpses. When the agent senses only 4% of the total image area, the inclusion of the proposed consistency loss in our training objective yields 3% and 8% higher accuracy on ImageNet and fMoW datasets, respectively. Moreover, our agent outperforms previous state-of-the-art by observing nearly 27% and 42% fewer pixels in glimpses on ImageNet and fMoW.

Rezvan Sherkati, James J. Clark

We present a new method for image salience prediction, Clustered Saliency Prediction. This method divides subjects into clusters based on their personal features and their known saliency maps, and generates an image salience model conditioned on the cluster label. We test our approach on a public dataset of personalized saliency maps and cluster the subjects using selected importance weights for personal feature factors. We propose the Multi-Domain Saliency Translation model which uses image stimuli and universal saliency maps to predict saliency maps for each cluster. For obtaining universal saliency maps, we applied various state-of-the-art methods, DeepGaze IIE, ML-Net and SalGAN, and compared their effectiveness in our system. We show that our Clustered Saliency Prediction technique outperforms the universal saliency prediction models. Also, we demonstrate the effectiveness of our clustering method by comparing the results of Clustered Saliency Prediction using clusters obtained by our algorithm with some baseline methods. Finally, we propose an approach to assign new people to their most appropriate cluster and prove its usefulness in the experiments.

Mohammadreza Tayaranian, Seyyed Hasan Mozafari, Brett H. Meyer et al.

Transformer-based language models have shown state-of-the-art performance on a variety of natural language understanding tasks. To achieve this performance, these models are first pre-trained on general corpus and then fine-tuned on downstream tasks. Previous work studied the effect of pruning the training set of the downstream tasks on the performance of the model on its evaluation set. In this work, we propose an automatic dataset pruning method for the training set of fine-tuning tasks. Our method is based on the model's success rate in correctly classifying each training data point. Unlike previous work which relies on user feedback to determine subset size, our method automatically extracts training subsets that are adapted for each pair of model and fine-tuning task. Our method provides multiple subsets for use in dataset pruning that navigate the trade-off between subset size and evaluation accuracy. Our largest subset, which we also refer to as the winning ticket subset, is on average $3 \times$ smaller than the original training set of the fine-tuning task. Our experiments on 5 downstream tasks and 2 language models show that, on average, fine-tuning on the winning ticket subsets results in a $0.1 \%$ increase in the evaluation performance of the model.

Saverio Vadacchino, Raghav Mehta, Nazanin Mohammadi Sepahvand et al.

Segmentation of enhancing tumours or lesions from MRI is important for detecting new disease activity in many clinical contexts. However, accurate segmentation requires the inclusion of medical images (e.g., T1 post contrast MRI) acquired after injecting patients with a contrast agent (e.g., Gadolinium), a process no longer thought to be safe. Although a number of modality-agnostic segmentation networks have been developed over the past few years, they have been met with limited success in the context of enhancing pathology segmentation. In this work, we present HAD-Net, a novel offline adversarial knowledge distillation (KD) technique, whereby a pre-trained teacher segmentation network, with access to all MRI sequences, teaches a student network, via hierarchical adversarial training, to better overcome the large domain shift presented when crucial images are absent during inference. In particular, we apply HAD-Net to the challenging task of enhancing tumour segmentation when access to post-contrast imaging is not available. The proposed network is trained and tested on the BraTS 2019 brain tumour segmentation challenge dataset, where it achieves performance improvements in the ranges of 16% - 26% over (a) recent modality-agnostic segmentation methods (U-HeMIS, U-HVED), (b) KD-Net adapted to this problem, (c) the pre-trained student network and (d) a non-hierarchical version of the network (AD-Net), in terms of Dice scores for enhancing tumour (ET). The network also shows improvements in tumour core (TC) Dice scores. Finally, the network outperforms both the baseline student network and AD-Net in terms of uncertainty quantification for enhancing tumour segmentation based on the BraTs 2019 uncertainty challenge metrics. Our code is publicly available at: https://github.com/SaverioVad/HAD_Net

Youyuan Zhang, Xuan Ju, James J. Clark

Diffusion models have demonstrated remarkable capabilities in text-to-image and text-to-video generation, opening up possibilities for video editing based on textual input. However, the computational cost associated with sequential sampling in diffusion models poses challenges for efficient video editing. Existing approaches relying on image generation models for video editing suffer from time-consuming one-shot fine-tuning, additional condition extraction, or DDIM inversion, making real-time applications impractical. In this work, we propose FastVideoEdit, an efficient zero-shot video editing approach inspired by Consistency Models (CMs). By leveraging the self-consistency property of CMs, we eliminate the need for time-consuming inversion or additional condition extraction, reducing editing time. Our method enables direct mapping from source video to target video with strong preservation ability utilizing a special variance schedule. This results in improved speed advantages, as fewer sampling steps can be used while maintaining comparable generation quality. Experimental results validate the state-of-the-art performance and speed advantages of FastVideoEdit across evaluation metrics encompassing editing speed, temporal consistency, and text-video alignment.

Ye Yuan, Youyuan Zhang, Can Chen et al.

Offline model-based optimization (MBO) aims to maximize a black-box objective function using only an offline dataset of designs and scores. These tasks span various domains, such as robotics, material design, and protein and molecular engineering. A common approach involves training a surrogate model using existing designs and their corresponding scores, and then generating new designs through gradient-based updates with respect to the surrogate model. This method suffers from the out-of-distribution issue, where the surrogate model may erroneously predict high scores for unseen designs. To address this challenge, we introduce a novel method, Design Editing for Offline Model-based Optimization (DEMO), which leverages a diffusion prior to calibrate overly optimized designs. DEMO first generates pseudo design candidates by performing gradient ascent with respect to a surrogate model. While these pseudo design candidates contain information beyond the offline dataset, they might be invalid or have erroneously high predicted scores. Therefore, to address this challenge while utilizing the information provided by pseudo design candidates, we propose an editing process to refine these pseudo design candidates. We introduce noise to the pseudo design candidates and subsequently denoise them with a diffusion prior trained on the offline dataset, ensuring they align with the distribution of valid designs. Empirical evaluations on seven offline MBO tasks show that, with properly tuned hyperparameters, DEMOs score is competitive with the best previously reported scores in the literature.

Youyuan Zhang, Zehua Liu, Zenan Li et al. · utoronto

In this paper, we consider the conditional generation problem by guiding off-the-shelf unconditional diffusion models with differentiable loss functions in a plug-and-play fashion. While previous research has primarily focused on balancing the unconditional diffusion model and the guided loss through a tuned weight hyperparameter, we propose a novel framework that distinctly decouples these two components. Specifically, we introduce two variables ${x}$ and ${z}$, to represent the generated samples governed by the unconditional generation model and the guidance function, respectively. This decoupling reformulates conditional generation into two manageable subproblems, unified by the constraint ${x} = {z}$. Leveraging this setup, we develop a new algorithm based on the Alternating Direction Method of Multipliers (ADMM) to adaptively balance these components. Additionally, we establish the equivalence between the diffusion reverse step and the proximal operator of ADMM and provide a detailed convergence analysis of our algorithm under certain mild assumptions. Our experiments demonstrate that our proposed method ADMMDiff consistently generates high-quality samples while ensuring strong adherence to the conditioning criteria. It outperforms existing methods across a range of conditional generation tasks, including image generation with various guidance and controllable motion synthesis.

Mohammadreza Tayaranian, Seyyed Hasan Mozafari, James J. Clark et al.

SWIN transformer is a prominent vision transformer model that has state-of-the-art accuracy in image classification tasks. Despite this success, its unique architecture causes slower inference compared with similar deep neural networks. Integer quantization of the model is one of the methods used to improve its inference latency. However, state-of-the-art has not been able to fully quantize the model. In this work, we improve upon the inference latency of the state-of-the-art methods by removing the floating-point operations, which are associated with the GELU activation in Swin Transformer. While previous work proposed to replace the non-integer operations with linear approximation functions, we propose to replace GELU with ReLU activation. The advantage of ReLU over previous methods is its low memory and computation complexity. We use iterative knowledge distillation to compensate for the lost accuracy due to replacing GELU with ReLU. We quantize our GELU-less SWIN transformer and show that on an RTX 4090 NVIDIA GPU we can improve the inference latency of the quantized SWIN transformer by at least $11\%$ while maintaining an accuracy drop of under $0.5\%$ on the ImageNet evaluation dataset.

Andrei Chubarau, Yinan Wang, James J. Clark

We introduce Neural Radiance and Gaze Fields (NeRGs) as a novel approach for representing visual attention patterns in 3D scenes. Our system renders a 2D view of a 3D scene with a pre-trained Neural Radiance Field (NeRF) and visualizes the gaze field for arbitrary observer positions, which may be decoupled from the render camera perspective. We achieve this by augmenting a standard NeRF with an additional neural network that models the gaze probability distribution. The output of a NeRG is a rendered image of the scene viewed from the camera perspective and a pixel-wise salience map representing conditional probability that an observer fixates on a given surface within the 3D scene as visible in the rendered image. Much like how NeRFs perform novel view synthesis, NeRGs enable the reconstruction of gaze patterns from arbitrary perspectives within complex 3D scenes. To ensure consistent gaze reconstructions, we constrain gaze prediction on the 3D structure of the scene and model gaze occlusion due to intervening surfaces when the observer's viewpoint is decoupled from the rendering camera. For training, we leverage ground truth head pose data from skeleton tracking data or predictions from 2D salience models. We demonstrate the effectiveness of NeRGs in a real-world convenience store setting, where head pose tracking data is available.

Rambod Azimi, Yijian Kong, Dusan Gostimirovic et al.

Integrated silicon photonic devices, which manipulate light to transmit and process information on a silicon-on-insulator chip, are highly sensitive to structural variations. Minor deviations during nanofabrication-the precise process of building structures at the nanometer scale-such as over- or under-etching, corner rounding, and unintended defects, can significantly impact performance. To address these challenges, we introduce SEMU-Net, a comprehensive set of methods that automatically segments scanning electron microscope images (SEM) and uses them to train two deep neural network models based on U-Net and its variants. The predictor model anticipates fabrication-induced variations, while the corrector model adjusts the design to address these issues, ensuring that the final fabricated structures closely align with the intended specifications. Experimental results show that the segmentation U-Net reaches an average IoU score of 99.30%, while the corrector attention U-Net in a tandem architecture achieves an average IoU score of 98.67%.

Lulan Shen, Ali Edalati, Brett Meyer et al.

This paper describes a simple yet effective technique for refining a pretrained classifier network. The proposed AdCorDA method is based on modification of the training set and making use of the duality between network weights and layer inputs. We call this input space training. The method consists of two stages - adversarial correction followed by domain adaptation. Adversarial correction uses adversarial attacks to correct incorrect training-set classifications. The incorrectly classified samples of the training set are removed and replaced with the adversarially corrected samples to form a new training set, and then, in the second stage, domain adaptation is performed back to the original training set. Extensive experimental validations show significant accuracy boosts of over 5% on the CIFAR-100 dataset. The technique can be straightforwardly applied to refinement of weight-quantized neural networks, where experiments show substantial enhancement in performance over the baseline. The adversarial correction technique also results in enhanced robustness to adversarial attacks.

Lulan Shen, Ali Edalati, Brett Meyer et al.

It is necessary to develop efficient DNNs deployed on edge devices with limited computation resources. However, the compressed networks often execute new tasks in the target domain, which is different from the source domain where the original network is trained. It is important to investigate the robustness of compressed networks in two types of data distribution shifts: domain shifts and adversarial perturbations. In this study, we discover that compressed models are less robust to distribution shifts than their original networks. Interestingly, larger networks are more vulnerable to losing robustness than smaller ones, even when they are compressed to a similar size as the smaller networks. Furthermore, compact networks obtained by knowledge distillation are much more robust to distribution shifts than pruned networks. Finally, post-training quantization is a reliable method for achieving significant robustness to distribution shifts, and it outperforms both pruned and distilled models in terms of robustness.

Amir Ardakani, Arash Ardakani, Brett Meyer et al.

Quantization of deep neural networks is a promising approach that reduces the inference cost, making it feasible to run deep networks on resource-restricted devices. Inspired by existing methods, we propose a new framework to learn the quantization intervals (discrete values) using the knowledge of the network's weight and activation distributions, i.e., standard deviation. Furthermore, we propose a novel base-2 logarithmic quantization scheme to quantize weights to power-of-two discrete values. Our proposed scheme allows us to replace resource-hungry high-precision multipliers with simple shift-add operations. According to our evaluations, our method outperforms existing work on CIFAR10 and ImageNet datasets and even achieves better accuracy performance with 3-bit weights and activations when compared to the full-precision models. Moreover, our scheme simultaneously prunes the network's parameters and allows us to flexibly adjust the pruning ratio during the quantization process.

Samrudhdhi B. Rangrej, James J. Clark

A visual hard attention model actively selects and observes a sequence of subregions in an image to make a prediction. The majority of hard attention models determine the attention-worthy regions by first analyzing a complete image. However, it may be the case that the entire image is not available initially but instead sensed gradually through a series of partial observations. In this paper, we design an efficient hard attention model for classifying such sequentially observed scenes. The presented model never observes an image completely. To select informative regions under partial observability, the model uses Bayesian Optimal Experiment Design. First, it synthesizes the features of the unobserved regions based on the already observed regions. Then, it uses the predicted features to estimate the expected information gain (EIG) attained, should various regions be attended. Finally, the model attends to the actual content on the location where the EIG mentioned above is maximum. The model uses a) a recurrent feature aggregator to maintain a recurrent state, b) a linear classifier to predict the class label, c) a Partial variational autoencoder to predict the features of unobserved regions. We use normalizing flows in Partial VAE to handle multi-modality in the feature-synthesis problem. We train our model using a differentiable objective and test it on five datasets. Our model gains 2-10% higher accuracy than the baseline models when both have seen only a couple of glimpses.

Ali Edalati, Marzieh Tahaei, Ahmad Rashid et al.

GPT is an auto-regressive Transformer-based pre-trained language model which has attracted a lot of attention in the natural language processing (NLP) domain due to its state-of-the-art performance in several downstream tasks. The success of GPT is mostly attributed to its pre-training on huge amount of data and its large number of parameters (from ~100M to billions of parameters). Despite the superior performance of GPT (especially in few-shot or zero-shot setup), this overparameterized nature of GPT can be very prohibitive for deploying this model on devices with limited computational power or memory. This problem can be mitigated using model compression techniques; however, compressing GPT models has not been investigated much in the literature. In this work, we use Kronecker decomposition to compress the linear mappings of the GPT-22 model. Our Kronecker GPT-2 model (KnGPT2) is initialized based on the Kronecker decomposed version of the GPT-2 model and then is undergone a very light pre-training on only a small portion of the training data with intermediate layer knowledge distillation (ILKD). Finally, our KnGPT2 is fine-tuned on down-stream tasks using ILKD as well. We evaluate our model on both language modeling and General Language Understanding Evaluation benchmark tasks and show that with more efficient pre-training and similar number of parameters, our KnGPT2 outperforms the existing DistilGPT2 model significantly.

Samrudhdhi B. Rangrej, James J. Clark

We present a recurrent agent who perceives surroundings through a series of discrete fixations. At each timestep, the agent imagines a variety of plausible scenes consistent with the fixation history. The next fixation is planned using uncertainty in the content of the imagined scenes. As time progresses, the agent becomes more certain about the content of the surrounding, and the variety in the imagined scenes reduces. The agent is built using a variational autoencoder and normalizing flows, and trained in an unsupervised manner on a proxy task of scene-reconstruction. The latent representations of the imagined scenes are found to be useful for performing pixel-level and scene-level tasks by higher-order modules. The agent is tested on various 2D and 3D datasets.

Qing Tian, Tal Arbel, James J. Clark

Most of today's popular deep architectures are hand-engineered to be generalists. However, this design procedure usually leads to massive redundant, useless, or even harmful features for specific tasks. Unnecessarily high complexities render deep nets impractical for many real-world applications, especially those without powerful GPU support. In this paper, we attempt to derive task-dependent compact models from a deep discriminant analysis perspective. We propose an iterative and proactive approach for classification tasks which alternates between (1) a pushing step, with an objective to simultaneously maximize class separation, penalize co-variances, and push deep discriminants into alignment with a compact set of neurons, and (2) a pruning step, which discards less useful or even interfering neurons. Deconvolution is adopted to reverse 'unimportant' filters' effects and recover useful contributing sources. A simple network growing strategy based on the basic Inception module is proposed for challenging tasks requiring larger capacity than what the base net can offer. Experiments on the MNIST, CIFAR10, and ImageNet datasets demonstrate our approach's efficacy. On ImageNet, by pushing and pruning our grown Inception-88 model, we achieve more accurate models than Inception nets generated during growing, residual nets, and popular compact nets at similar sizes. We also show that our grown Inception nets (without hard-coded dimension alignment) clearly outperform residual nets of similar complexities.

Guanqing Hu, James J. Clark

Image compositing is a key step in film making and image editing that aims to segment a foreground object and combine it with a new background. Automatic image compositing can be done easily in a studio using chroma-keying when the background is pure blue or green. However, image compositing in natural scenes with complex backgrounds remains a tedious task, requiring experienced artists to hand-segment. In order to achieve automatic compositing in natural scenes, we propose a fully automated method that integrates instance segmentation and image matting processes to generate high-quality semantic mattes that can be used for image editing task. Our approach can be seen both as a refinement of existing instance segmentation algorithms and as a fully automated semantic image matting method. It extends automatic image compositing techniques such as chroma-keying to scenes with complex natural backgrounds without the need for any kind of user interaction. The output of our approach can be considered as both refined instance segmentations and alpha mattes with semantic meanings. We provide experimental results which show improved performance results as compared to existing approaches.

Qing Tian, Tal Arbel, James J. Clark

With deep learning's success, a limited number of popular deep nets have been widely adopted for various vision tasks. However, this usually results in unnecessarily high complexities and possibly many features of low task utility. In this paper, we address this problem by introducing a task-dependent deep pruning framework based on Fisher's Linear Discriminant Analysis (LDA). The approach can be applied to convolutional, fully-connected, and module-based deep network structures, in all cases leveraging the high decorrelation of neuron motifs found in the pre-decision space and cross-layer deconv dependency. Moreover, we examine our approach's potential in network architecture search for specific tasks and analyze the influence of our pruning on model robustness to noises and adversarial attacks. Experimental results on datasets of generic objects (ImageNet, CIFAR100) as well as domain specific tasks (Adience, and LFWA) illustrate our framework's superior performance over state-of-the-art pruning approaches and fixed compact nets (e.g. SqueezeNet, MobileNet). The proposed method successfully maintains comparable accuracies even after discarding most parameters (98%-99% for VGG16, up to 82% for the already compact InceptionNet) and with significant FLOP reductions (83% for VGG16, up to 64% for InceptionNet). Through pruning, we can also derive smaller, but more accurate and more robust models suitable for the task.

Bingqing Yu, James J. Clark

Most existing saliency models use low-level features or task descriptions when generating attention predictions. However, the link between observer characteristics and gaze patterns is rarely investigated. We present a novel saliency prediction technique which takes viewers' identities and personal traits into consideration when modeling human attention. Instead of only computing image salience for average observers, we consider the interpersonal variation in the viewing behaviors of observers with different personal traits and backgrounds. We present an enriched derivative of the GAN network, which is able to generate personalized saliency predictions when fed with image stimuli and specific information about the observer. Our model contains a generator which generates grayscale saliency heat maps based on the image and an observer label. The generator is paired with an adversarial discriminator which learns to distinguish generated salience from ground truth salience. The discriminator also has the observer label as an input, which contributes to the personalization ability of our approach. We evaluate the performance of our personalized salience model by comparison with a benchmark model along with other un-personalized predictions, and illustrate improvements in prediction accuracy for all tested observer groups.

Bingqing Yu, James J. Clark

We present a method for reconstructing images viewed by observers based only on their eye movements. By exploring the relationships between gaze patterns and image stimuli, the "What Are You Looking At?" (WAYLA) system learns to synthesize photo-realistic images that are similar to the original pictures being viewed. The WAYLA approach is based on the Conditional Generative Adversarial Network (Conditional GAN) image-to-image translation technique of Isola et al. We consider two specific applications - the first, of reconstructing newspaper images from gaze heat maps, and the second, of detailed reconstruction of images containing only text. The newspaper image reconstruction process is divided into two image-to-image translation operations, the first mapping gaze heat maps into image segmentations, and the second mapping the generated segmentation into a newspaper image. We validate the performance of our approach using various evaluation metrics, along with human visual inspection. All results confirm the ability of our network to perform image generation tasks using eye tracking data.

Qing Tian, Tal Arbel, James J. Clark

Many real-time tasks, such as human-computer interaction, require fast and efficient facial gender classification. Although deep CNN nets have been very effective for a multitude of classification tasks, their high space and time demands make them impractical for personal computers and mobile devices without a powerful GPU. In this paper, we develop a 16-layer, yet lightweight, neural network which boosts efficiency while maintaining high accuracy. Our net is pruned from the VGG-16 model starting from the last convolutional (conv) layer where we find neuron activations are highly uncorrelated given the gender. Through Fisher's Linear Discriminant Analysis (LDA), we show that this high decorrelation makes it safe to discard directly last conv layer neurons with high within-class variance and low between-class variance. Combined with either Support Vector Machines (SVM) or Bayesian classification, the reduced CNNs are capable of achieving comparable (or even higher) accuracies on the LFW and CelebA datasets than the original net with fully connected layers. On LFW, only four Conv5_3 neurons are able to maintain a comparably high recognition accuracy, which results in a reduction of total network size by a factor of 70X with a 11 fold speedup. Comparisons with a state-of-the-art pruning method as well as two smaller nets in terms of accuracy loss and convolutional layers pruning rate are also provided.

Siavash Gorji, James J. Clark

We present a novel visual attention tracking technique based on Shared Attention modeling. Our proposed method models the viewer as a participant in the activity occurring in the scene. We go beyond image salience and instead of only computing the power of an image region to pull attention to it, we also consider the strength with which other regions of the image push attention to the region in question. We use the term Attentional Push to refer to the power of image regions to direct and manipulate the attention allocation of the viewer. An attention model is presented that incorporates the Attentional Push cues with standard image salience-based attention modeling algorithms to improve the ability to predict where viewers will fixate. Experimental evaluation validates significant improvements in predicting viewers' fixations using the proposed methodology in both static and dynamic imagery.