Toufiq Parag

Susmit Agrawal, Prabhat Kumar, Siddharth Seth et al.

Recent algorithms for image manipulation detection almost exclusively use deep network models. These approaches require either dense pixelwise groundtruth masks, camera ids, or image metadata to train the networks. On one hand, constructing a training set to represent the countless tampering possibilities is impractical. On the other hand, social media platforms or commercial applications are often constrained to remove camera ids as well as metadata from images. A self-supervised algorithm for training manipulation detection models without dense groundtruth or camera/image metadata would be extremely useful for many forensics applications. In this paper, we propose self-supervised approach for training splicing detection/localization models from frequency transforms of images. To identify the spliced regions, our deep network learns a representation to capture an image specific signature by enforcing (image) self consistency . We experimentally demonstrate that our proposed model can yield similar or better performances of multiple existing methods on standard datasets without relying on labels or metadata.

Shaunak Halbe, Bhagyashree Puranik, Jayakrishnan Unnikrishnan et al.

Modern video retrieval systems are expected to handle diverse tasks ranging from corpus-level retrieval and fine-grained moment localization to flexible multimodal querying. Specialized architectures achieve strong retrieval performance by training modality-specific encoders on massive datasets, but they lack the ability to process composed multimodal queries. In contrast, multimodal LLM (MLLM)-based methods support rich multimodal search but their retrieval performance remains well below that of specialized systems. We present VIRTUE, an MLLM-based versatile video retrieval framework that integrates corpus and moment-level retrieval capabilities while accommodating composed multimodal queries within a single architecture. We use contrastive alignment of visual and textual embeddings generated using a shared MLLM backbone to facilitate efficient embedding-based candidate search. Our embedding model, trained efficiently using low-rank adaptation (LoRA) on 700K paired visual-text data samples, surpasses other MLLM-based methods on zero-shot video retrieval tasks. Additionally, we demonstrate that the same model can be adapted without further training to achieve competitive results on zero-shot moment retrieval, and state of the art results for zero-shot composed video retrieval. With additional training for reranking candidates identified in the embedding-based search, our model substantially outperforms existing MLLM-based retrieval systems and achieves retrieval performance comparable to state of the art specialized models which are trained on orders of magnitude larger data.

Yifan Jiang, Yueying Wang, Rui Zhao et al.

Reinforcement fine-tuning (RFT), a two-stage framework consisting of supervised fine-tuning (SFT) and reinforcement learning (RL) has shown promising results on improving reasoning ability of large language models (LLMs). Yet extending RFT to large video language models (LVLMs) remains challenging. We propose VideoP2R, a novel process-aware video RFT framework that enhances video reasoning by modeling perception and reasoning as distinct processes. In the SFT stage, we develop a three-step pipeline to generate VideoP2R-CoT-162K, a high-quality, process-aware chain-of-thought (CoT) dataset for perception and reasoning. In the RL stage, we introduce a novel process-aware group relative policy optimization (PA-GRPO) algorithm that supplies separate rewards for perception and reasoning. Extensive experiments show that VideoP2R achieves state-of-the-art (SotA) performance on six out of seven video reasoning and understanding benchmarks. Ablation studies further confirm the effectiveness of our process-aware modeling and PA-GRPO and demonstrate that model's perception output is information-sufficient for downstream reasoning.

Toufiq Parag, Ahmed Elgammal

The long-context capability of recent large transformer models can be surmised to rely on techniques such as attention/model parallelism, as well as hardware-level optimizations. While these strategies allow input lengths to scale to millions of tokens, they do not fundamentally mitigate the quadratic computational and memory complexity of the core attention mechanism. In this paper, we address the challenge of long-context processing on a single device using generic implementations by reducing the quadratic memory footprint and inference cost. Existing approaches to extend the context length for generic single device implementations -- such as token merging and modified attentions -- are often modality specific and attain a suboptimal tradeoff between accuracy and efficiency. To overcome these limitations, we propose MAELRE (Modality Agnostic Efficient Long Range Encoder), a unified and efficient transformer architecture designed for long-range encoding across diverse modalities. MAELRE integrates token merging with attention approximation, progressively merging tokens at different stages of internal computational blocks. It employs a lightweight attention approximation when the number of tokens is large, and switches to standard dot-product attention as the sequence becomes shorter through successive aggregation. We demonstrate that MAELRE achieves superior accuracy while reducing computational cost compared to existing long-context models on classification tasks spanning multiple modalities, including text, time series, audio, and vision.

Soumya Banerjee, Vinay Kumar Verma, Toufiq Parag et al.

A wide variety of methods have been developed to enable lifelong learning in conventional deep neural networks. However, to succeed, these methods require a `batch' of samples to be available and visited multiple times during training. While this works well in a static setting, these methods continue to suffer in a more realistic situation where data arrives in \emph{online streaming manner}. We empirically demonstrate that the performance of current approaches degrades if the input is obtained as a stream of data with the following restrictions: $(i)$ each instance comes one at a time and can be seen only once, and $(ii)$ the input data violates the i.i.d assumption, i.e., there can be a class-based correlation. We propose a novel approach (CIOSL) for the class-incremental learning in an \emph{online streaming setting} to address these challenges. The proposed approach leverages implicit and explicit dual weight regularization and experience replay. The implicit regularization is leveraged via the knowledge distillation, while the explicit regularization incorporates a novel approach for parameter regularization by learning the joint distribution of the buffer replay and the current sample. Also, we propose an efficient online memory replay and replacement buffer strategy that significantly boosts the model's performance. Extensive experiments and ablation on challenging datasets show the efficacy of the proposed method.

Toufiq Parag, Hongcheng Wang

Classification is a pivotal function for many computer vision tasks such as object classification, detection, scene segmentation. Multinomial logistic regression with a single final layer of dense connections has become the ubiquitous technique for CNN-based classification. While these classifiers project a mapping between the input and a set of output category classes, they do not typically yield a comprehensive description of the category. In particular, when a CNN based image classifier correctly identifies the image of a Chimpanzee, its output does not clarify that Chimpanzee is a member of Primate, Mammal, Chordate families and a living thing. We propose a multilayer dense connectivity for concurrent prediction of category and its conceptual superclasses in hierarchical order by the same CNN. We experimentally demonstrate that our proposed network can simultaneously predict both the coarse superclasses and finer categories better than several existing algorithms in multiple datasets.

Longlong Jing, Toufiq Parag, Zhe Wu et al.

We propose a semi-supervised learning approach for video classification, VideoSSL, using convolutional neural networks (CNN). Like other computer vision tasks, existing supervised video classification methods demand a large amount of labeled data to attain good performance. However, annotation of a large dataset is expensive and time consuming. To minimize the dependence on a large annotated dataset, our proposed semi-supervised method trains from a small number of labeled examples and exploits two regulatory signals from unlabeled data. The first signal is the pseudo-labels of unlabeled examples computed from the confidences of the CNN being trained. The other is the normalized probabilities, as predicted by an image classifier CNN, that captures the information about appearances of the interesting objects in the video. We show that, under the supervision of these guiding signals from unlabeled examples, a video classification CNN can achieve impressive performances utilizing a small fraction of annotated examples on three publicly available datasets: UCF101, HMDB51 and Kinetics.

Felix Gonda, Donglai Wei, Toufiq Parag et al.

For video and volumetric data understanding, 3D convolution layers are widely used in deep learning, however, at the cost of increasing computation and training time. Recent works seek to replace the 3D convolution layer with convolution blocks, e.g. structured combinations of 2D and 1D convolution layers. In this paper, we propose a novel convolution block, Parallel Separable 3D Convolution (PmSCn), which applies m parallel streams of n 2D and one 1D convolution layers along different dimensions. We first mathematically justify the need of parallel streams (Pm) to replace a single 3D convolution layer through tensor decomposition. Then we jointly replace consecutive 3D convolution layers, common in modern network architectures, with the multiple 2D convolution layers (Cn). Lastly, we empirically show that PmSCn is applicable to different backbone architectures, such as ResNet, DenseNet, and UNet, for different applications, such as video action recognition, MRI brain segmentation, and electron microscopy segmentation. In all three applications, we replace the 3D convolution layers in state-of-the art models with PmSCn and achieve around 14% improvement in test performance and 40% reduction in model size and on average.

Toufiq Parag, Daniel Berger, Lee Kamentsky et al.

Synaptic connectivity detection is a critical task for neural reconstruction from Electron Microscopy (EM) data. Most of the existing algorithms for synapse detection do not identify the cleft location and direction of connectivity simultaneously. The few methods that computes direction along with contact location have only been demonstrated to work on either dyadic (most common in vertebrate brain) or polyadic (found in fruit fly brain) synapses, but not on both types. In this paper, we present an algorithm to automatically predict the location as well as the direction of both dyadic and polyadic synapses. The proposed algorithm first generates candidate synaptic connections from voxelwise predictions of signed proximity generated by a 3D U-net. A second 3D CNN then prunes the set of candidates to produce the final detection of cleft and connectivity orientation. Experimental results demonstrate that the proposed method outperforms the existing methods for determining synapses in both rodent and fruit fly brain.

Toufiq Parag, Fabian Tschopp, William Grisaitis et al.

The field of connectomics has recently produced neuron wiring diagrams from relatively large brain regions from multiple animals. Most of these neural reconstructions were computed from isotropic (e.g., FIBSEM) or near isotropic (e.g., SBEM) data. In spite of the remarkable progress on algorithms in recent years, automatic dense reconstruction from anisotropic data remains a challenge for the connectomics community. One significant hurdle in the segmentation of anisotropic data is the difficulty in generating a suitable initial over-segmentation. In this study, we present a segmentation method for anisotropic EM data that agglomerates a 3D over-segmentation computed from the 3D affinity prediction. A 3D U-net is trained to predict 3D affinities by the MALIS approach. Experiments on multiple datasets demonstrates the strength and robustness of the proposed method for anisotropic EM segmentation.

David Rolnick, Yaron Meirovitch, Toufiq Parag et al.

Deep learning algorithms for connectomics rely upon localized classification, rather than overall morphology. This leads to a high incidence of erroneously merged objects. Humans, by contrast, can easily detect such errors by acquiring intuition for the correct morphology of objects. Biological neurons have complicated and variable shapes, which are challenging to learn, and merge errors take a multitude of different forms. We present an algorithm, MergeNet, that shows 3D ConvNets can, in fact, detect merge errors from high-level neuronal morphology. MergeNet follows unsupervised training and operates across datasets. We demonstrate the performance of MergeNet both on a variety of connectomics data and on a dataset created from merged MNIST images.

Felix Gonda, Verena Kaynig, Ray Thouis et al.

We present an interactive approach to train a deep neural network pixel classifier for the segmentation of neuronal structures. An interactive training scheme reduces the extremely tedious manual annotation task that is typically required for deep networks to perform well on image segmentation problems. Our proposed method employs a feedback loop that captures sparse annotations using a graphical user interface, trains a deep neural network based on recent and past annotations, and displays the prediction output to users in almost real-time. Our implementation of the algorithm also allows multiple users to provide annotations in parallel and receive feedback from the same classifier. Quick feedback on classifier performance in an interactive setting enables users to identify and label examples that are more important than others for segmentation purposes. Our experiments show that an interactively-trained pixel classifier produces better region segmentation results on Electron Microscopy (EM) images than those generated by a network of the same architecture trained offline on exhaustive ground-truth labels.

Toufiq Parag

The prospect of neural reconstruction from Electron Microscopy (EM) images has been elucidated by the automatic segmentation algorithms. Although segmentation algorithms eliminate the necessity of tracing the neurons by hand, significant manual effort is still essential for correcting the mistakes they make. A considerable amount of human labor is also required for annotating groundtruth volumes for training the classifiers of a segmentation framework. It is critically important to diminish the dependence on human interaction in the overall reconstruction system. This study proposes a novel classifier training algorithm for EM segmentation aimed to reduce the amount of manual effort demanded by the groundtruth annotation and error refinement tasks. Instead of using an exhaustive pixel level groundtruth, an active learning algorithm is proposed for sparse labeling of pixel and boundaries of superpixels. Because over-segmentation errors are in general more tolerable and easier to correct than the under-segmentation errors, our algorithm is designed to prioritize minimization of false-merges over false-split mistakes. Our experiments on both 2D and 3D data suggest that the proposed method yields segmentation outputs that are more amenable to neural reconstruction than those of existing methods.

Toufiq Parag

This study formulates the IR target detection as a binary classification problem of each pixel. Each pixel is associated with a label which indicates whether it is a target or background pixel. The optimal label set for all the pixels of an image maximizes aposteriori distribution of label configuration given the pixel intensities. The posterior probability is factored into (or proportional to) a conditional likelihood of the intensity values and a prior probability of label configuration. Each of these two probabilities are computed assuming a Markov Random Field (MRF) on both pixel intensities and their labels. In particular, this study enforces neighborhood dependency on both intensity values, by a Simultaneous Auto Regressive (SAR) model, and on labels, by an Auto-Logistic model. The parameters of these MRF models are learned from labeled examples. During testing, an MRF inference technique, namely Iterated Conditional Mode (ICM), produces the optimal label for each pixel. The detection performance is further improved by incorporating temporal information through background subtraction. High performances on benchmark datasets demonstrate effectiveness of this method for IR target detection.

Stephen M. Plaza, Toufiq Parag, Gary B. Huang et al.

Reconstructing neuronal circuits at the level of synapses is a central problem in neuroscience and becoming a focus of the emerging field of connectomics. To date, electron microscopy (EM) is the most proven technique for identifying and quantifying synaptic connections. As advances in EM make acquiring larger datasets possible, subsequent manual synapse identification ({\em i.e.}, proofreading) for deciphering a connectome becomes a major time bottleneck. Here we introduce a large-scale, high-throughput, and semi-automated methodology to efficiently identify synapses. We successfully applied our methodology to the Drosophila medulla optic lobe, annotating many more synapses than previous connectome efforts. Our approaches are extensible and will make the often complicated process of synapse identification accessible to a wider-community of potential proofreaders.

Toufiq Parag, Stephen Plaza, Louis Scheffer

Pixel and superpixel classifiers have become essential tools for EM segmentation algorithms. Training these classifiers remains a major bottleneck primarily due to the requirement of completely annotating the dataset which is tedious, error-prone and costly. In this paper, we propose an interactive learning scheme for the superpixel classifier for EM segmentation. Our algorithm is "active semi-supervised" because it requests the labels of a small number of examples from user and applies label propagation technique to generate these queries. Using only a small set ($<20\%$) of all datapoints, the proposed algorithm consistently generates a classifier almost as accurate as that estimated from a complete groundtruth. We provide segmentation results on multiple datasets to show the strength of these classifiers.

Toufiq Parag, Anirban Chakraborty, Stephen Plaza et al.

Electron Microscopy (EM) image (or volume) segmentation has become significantly important in recent years as an instrument for connectomics. This paper proposes a novel agglomerative framework for EM segmentation. In particular, given an over-segmented image or volume, we propose a novel framework for accurately clustering regions of the same neuron. Unlike existing agglomerative methods, the proposed context-aware algorithm divides superpixels (over-segmented regions) of different biological entities into different subsets and agglomerates them separately. In addition, this paper describes a "delayed" scheme for agglomerative clustering that postpones some of the merge decisions, pertaining to newly formed bodies, in order to generate a more confident boundary prediction. We report significant improvements attained by the proposed approach in segmentation accuracy over existing standard methods on 2D and 3D datasets.

Toufiq Parag

A set label disagreement function is defined over the number of variables that deviates from the dominant label. The dominant label is the value assumed by the largest number of variables within a set of binary variables. The submodularity of a certain family of set label disagreement function is discussed in this manuscript. Such disagreement function could be utilized as a cost function in combinatorial optimization approaches for problems defined over hypergraphs.

Juan Nunez-Iglesias, Ryan Kennedy, Toufiq Parag et al.

We aim to improve segmentation through the use of machine learning tools during region agglomeration. We propose an active learning approach for performing hierarchical agglomerative segmentation from superpixels. Our method combines multiple features at all scales of the agglomerative process, works for data with an arbitrary number of dimensions, and scales to very large datasets. We advocate the use of variation of information to measure segmentation accuracy, particularly in 3D electron microscopy (EM) images of neural tissue, and using this metric demonstrate an improvement over competing algorithms in EM and natural images.