Siddhant Garg

Luca Di Liello, Siddhant Garg, Luca Soldaini et al. · allen-ai, amazon-science

Inference tasks such as answer sentence selection (AS2) or fact verification are typically solved by fine-tuning transformer-based models as individual sentence-pair classifiers. Recent studies show that these tasks benefit from modeling dependencies across multiple candidate sentences jointly. In this paper, we first show that popular pre-trained transformers perform poorly when used for fine-tuning on multi-candidate inference tasks. We then propose a new pre-training objective that models the paragraph-level semantics across multiple input sentences. Our evaluation on three AS2 and one fact verification datasets demonstrates the superiority of our pre-training technique over the traditional ones for transformers used as joint models for multi-candidate inference tasks, as well as when used as cross-encoders for sentence-pair formulations of these tasks. Our code and pre-trained models are released at https://github.com/amazon-research/wqa-multi-sentence-inference .

Luca Di Liello, Siddhant Garg, Luca Soldaini et al. · allen-ai, amazon-science

An important task for designing QA systems is answer sentence selection (AS2): selecting the sentence containing (or constituting) the answer to a question from a set of retrieved relevant documents. In this paper, we propose three novel sentence-level transformer pre-training objectives that incorporate paragraph-level semantics within and across documents, to improve the performance of transformers for AS2, and mitigate the requirement of large labeled datasets. Specifically, the model is tasked to predict whether: (i) two sentences are extracted from the same paragraph, (ii) a given sentence is extracted from a given paragraph, and (iii) two paragraphs are extracted from the same document. Our experiments on three public and one industrial AS2 datasets demonstrate the empirical superiority of our pre-trained transformers over baseline models such as RoBERTa and ELECTRA for AS2.

Matteo Gabburo, Rik Koncel-Kedziorski, Siddhant Garg et al. · allen-ai, amazon-science

Recent studies show that Question Answering (QA) based on Answer Sentence Selection (AS2) can be improved by generating an improved answer from the top-k ranked answer sentences (termed GenQA). This allows for synthesizing the information from multiple candidates into a concise, natural-sounding answer. However, creating large-scale supervised training data for GenQA models is very challenging. In this paper, we propose to train a GenQA model by transferring knowledge from a trained AS2 model, to overcome the aforementioned issue. First, we use an AS2 model to produce a ranking over answer candidates for a set of questions. Then, we use the top ranked candidate as the generation target, and the next k top ranked candidates as context for training a GenQA model. We also propose to use the AS2 model prediction scores for loss weighting and score-conditioned input/output shaping, to aid the knowledge transfer. Our evaluation on three public and one large industrial datasets demonstrates the superiority of our approach over the AS2 baseline, and GenQA trained using supervised data.

Matteo Gabburo, Siddhant Garg, Rik Koncel Kedziorski et al. · amazon-science

Evaluation of QA systems is very challenging and expensive, with the most reliable approach being human annotations of correctness of answers for questions. Recent works (AVA, BEM) have shown that transformer LM encoder based similarity metrics transfer well for QA evaluation, but they are limited by the usage of a single correct reference answer. We propose a new evaluation metric: SQuArE (Sentence-level QUestion AnsweRing Evaluation), using multiple reference answers (combining multiple correct and incorrect references) for sentence-form QA. We evaluate SQuArE on both sentence-level extractive (Answer Selection) and generative (GenQA) QA systems, across multiple academic and industrial datasets, and show that it outperforms previous baselines and obtains the highest correlation with human annotations.

Siddhant Garg, Mudit Chaudhary · amazon-science

We present SeRP, a framework for Self-Supervised Learning of 3D point clouds. SeRP consists of encoder-decoder architecture that takes perturbed or corrupted point clouds as inputs and aims to reconstruct the original point cloud without corruption. The encoder learns the high-level latent representations of the points clouds in a low-dimensional subspace and recovers the original structure. In this work, we have used Transformers and PointNet-based Autoencoders. The proposed framework also addresses some of the limitations of Transformers-based Masked Autoencoders which are prone to leakage of location information and uneven information density. We trained our models on the complete ShapeNet dataset and evaluated them on ModelNet40 as a downstream classification task. We have shown that the pretrained models achieved 0.5-1% higher classification accuracies than the networks trained from scratch. Furthermore, we also proposed VASP: Vector-Quantized Autoencoder for Self-supervised Representation Learning for Point Clouds that employs Vector-Quantization for discrete representation learning for Transformer-based autoencoders.

Siddhant Garg, Lijun Zhang, Hui Guan · amazon-science

Although multi-task deep neural network (DNN) models have computation and storage benefits over individual single-task DNN models, they can be further optimized via model compression. Numerous structured pruning methods are already developed that can readily achieve speedups in single-task models, but the pruning of multi-task networks has not yet been extensively studied. In this work, we investigate the effectiveness of structured pruning on multi-task models. We use an existing single-task filter pruning criterion and also introduce an MTL-based filter pruning criterion for estimating the filter importance scores. We prune the model using an iterative pruning strategy with both pruning methods. We show that, with careful hyper-parameter tuning, architectures obtained from different pruning methods do not have significant differences in their performances across tasks when the number of parameters is similar. We also show that iterative structure pruning may not be the best way to achieve a well-performing pruned model because, at extreme pruning levels, there is a high drop in performance across all tasks. But when the same models are randomly initialized and re-trained, they show better results.

Siddhant Garg, Dhruval Jain · amazon-science

In this work, we have worked towards two major goals. Firstly, we have investigated the importance of Batch Normalisation (BN) layers in a non-contrastive representation learning framework called Bootstrap Your Own Latent (BYOL). We conducted several experiments to conclude that BN layers are not necessary for representation learning in BYOL. Moreover, BYOL only learns from the positive pairs of images but ignores other semantically similar images in the same input batch. For the second goal, we have introduced two new loss functions to determine the semantically similar pairs in the same input batch of images and reduce the distance between their representations. These loss functions are Cross-Cosine Similarity Loss (CCSL) and Cross-Sigmoid Similarity Loss (CSSL). Using the proposed loss functions, we are able to surpass the performance of Vanilla BYOL (71.04%) by training the BYOL framework using CCSL loss (76.87%) on the STL10 dataset. BYOL trained using CSSL loss performs comparably with Vanilla BYOL.

Mehmet F. Demirel, Shengchao Liu, Siddhant Garg et al.

Graph neural networks (GNNs) have been shown to possess strong representation power, which can be exploited for downstream prediction tasks on graph-structured data, such as molecules and social networks. They typically learn representations by aggregating information from the $K$-hop neighborhood of individual vertices or from the enumerated walks in the graph. Prior studies have demonstrated the effectiveness of incorporating weighting schemes into GNNs; however, this has been primarily limited to $K$-hop neighborhood GNNs so far. In this paper, we aim to design an algorithm incorporating weighting schemes into walk-aggregating GNNs and analyze their effect. We propose a novel GNN model, called AWARE, that aggregates information about the walks in the graph using attention schemes. This leads to an end-to-end supervised learning method for graph-level prediction tasks in the standard setting where the input is the adjacency and vertex information of a graph, and the output is a predicted label for the graph. We then perform theoretical, empirical, and interpretability analyses of AWARE. Our theoretical analysis in a simplified setting identifies successful conditions for provable guarantees, demonstrating how the graph information is encoded in the representation, and how the weighting schemes in AWARE affect the representation and learning performance. Our experiments demonstrate the strong performance of AWARE in graph-level prediction tasks in the standard setting in the domains of molecular property prediction and social networks. Lastly, our interpretation study illustrates that AWARE can successfully capture the important substructures of the input graph. The code is available on $\href{https://github.com/mehmetfdemirel/aware}{GitHub}$.

Hongda Jiang, Xinyuan Zhang, Siddhant Garg et al. · amazon-science

We introduce Memory-QA, a novel real-world task that involves answering recall questions about visual content from previously stored multimodal memories. This task poses unique challenges, including the creation of task-oriented memories, the effective utilization of temporal and location information within memories, and the ability to draw upon multiple memories to answer a recall question. To address these challenges, we propose a comprehensive pipeline, Pensieve, integrating memory-specific augmentation, time- and location-aware multi-signal retrieval, and multi-memory QA fine-tuning. We created a multimodal benchmark to illustrate various real challenges in this task, and show the superior performance of Pensieve over state-of-the-art solutions (up to 14% on QA accuracy).

Matteo Gabburo, Nicolaas Paul Jedema, Siddhant Garg et al.

In this paper, we investigate which questions are challenging for retrieval-based Question Answering (QA). We (i) propose retrieval complexity (RC), a novel metric conditioned on the completeness of retrieved documents, which measures the difficulty of answering questions, and (ii) propose an unsupervised pipeline to measure RC given an arbitrary retrieval system. Our proposed pipeline measures RC more accurately than alternative estimators, including LLMs, on six challenging QA benchmarks. Further investigation reveals that RC scores strongly correlate with both QA performance and expert judgment across five of the six studied benchmarks, indicating that RC is an effective measure of question difficulty. Subsequent categorization of high-RC questions shows that they span a broad set of question shapes, including multi-hop, compositional, and temporal QA, indicating that RC scores can categorize a new subset of complex questions. Our system can also have a major impact on retrieval-based systems by helping to identify more challenging questions on existing datasets.

Luca Di Liello, Siddhant Garg, Alessandro Moschitti

Answer Sentence Selection (AS2) is a core component for building an accurate Question Answering pipeline. AS2 models rank a set of candidate sentences based on how likely they answer a given question. The state of the art in AS2 exploits pre-trained transformers by transferring them on large annotated datasets, while using local contextual information around the candidate sentence. In this paper, we propose three pre-training objectives designed to mimic the downstream fine-tuning task of contextual AS2. This allows for specializing LMs when fine-tuning for contextual AS2. Our experiments on three public and two large-scale industrial datasets show that our pre-training approaches (applied to RoBERTa and ELECTRA) can improve baseline contextual AS2 accuracy by up to 8% on some datasets.

Matteo Gabburo, Siddhant Garg, Rik Koncel-Kedziorski et al.

Recent studies show that sentence-level extractive QA, i.e., based on Answer Sentence Selection (AS2), is outperformed by Generation-based QA (GenQA) models, which generate answers using the top-k answer sentences ranked by AS2 models (a la retrieval-augmented generation style). In this paper, we propose a novel training paradigm for GenQA using supervision from automatic QA evaluation models (GAVA). Specifically, we propose three strategies to transfer knowledge from these QA evaluation models to a GenQA model: (i) augmenting training data with answers generated by the GenQA model and labelled by GAVA (either statically, before training, or (ii) dynamically, at every training epoch); and (iii) using the GAVA score for weighting the generator loss during the learning of the GenQA model. We evaluate our proposed methods on two academic and one industrial dataset, obtaining a significant improvement in answering accuracy over the previous state of the art.

Siddhant Garg, Debi Prasanna Mohanty, Siva Prasad Thota et al.

The main contributions of our work are two-fold. First, we present a Self-Attention MobileNet, called SA-MobileNet Network that can model long-range dependencies between the image features instead of processing the local region as done by standard convolutional kernels. SA-MobileNet contains self-attention modules integrated with the inverted bottleneck blocks of the MobileNetV3 model which results in modeling of both channel-wise attention and spatial attention of the image features and at the same time introduce a novel self-attention architecture for low-resource devices. Secondly, we propose a novel training pipeline for the task of image tilt detection. We treat this problem in a multi-label scenario where we predict multiple angles for a tilted input image in a narrow interval of range 1-2 degrees, depending on the dataset used. This process induces an implicit correlation between labels without any computational overhead of the second or higher-order methods in multi-label learning. With the combination of our novel approach and the architecture, we present state-of-the-art results on detecting the image tilt angle on mobile devices as compared to the MobileNetV3 model. Finally, we establish that SA-MobileNet is more accurate than MobileNetV3 on SUN397, NYU-V1, and ADE20K datasets by 6.42%, 10.51%, and 9.09% points respectively, and faster by at least 4 milliseconds on Snapdragon 750 Octa-core.

Siddhant Garg, Alessandro Moschitti

In this paper we propose a novel approach towards improving the efficiency of Question Answering (QA) systems by filtering out questions that will not be answered by them. This is based on an interesting new finding: the answer confidence scores of state-of-the-art QA systems can be approximated well by models solely using the input question text. This enables preemptive filtering of questions that are not answered by the system due to their answer confidence scores being lower than the system threshold. Specifically, we learn Transformer-based question models by distilling Transformer-based answering models. Our experiments on three popular QA datasets and one industrial QA benchmark demonstrate the ability of our question models to approximate the Precision/Recall curves of the target QA system well. These question models, when used as filters, can effectively trade off lower computation cost of QA systems for lower Recall, e.g., reducing computation by ~60%, while only losing ~3-4% of Recall.

Siddhant Garg, Goutham Ramakrishnan, Varun Thumbe

Large datasets in NLP suffer from noisy labels, due to erroneous automatic and human annotation procedures. We study the problem of text classification with label noise, and aim to capture this noise through an auxiliary noise model over the classifier. We first assign a probability score to each training sample of having a noisy label, through a beta mixture model fitted on the losses at an early epoch of training. Then, we use this score to selectively guide the learning of the noise model and classifier. Our empirical evaluation on two text classification tasks shows that our approach can improve over the baseline accuracy, and prevent over-fitting to the noise.

Siddhant Garg, Yingyu Liang

Unsupervised and self-supervised learning approaches have become a crucial tool to learn representations for downstream prediction tasks. While these approaches are widely used in practice and achieve impressive empirical gains, their theoretical understanding largely lags behind. Towards bridging this gap, we present a unifying perspective where several such approaches can be viewed as imposing a regularization on the representation via a learnable function using unlabeled data. We propose a discriminative theoretical framework for analyzing the sample complexity of these approaches, which generalizes the framework of (Balcan and Blum, 2010) to allow learnable regularization functions. Our sample complexity bounds show that, with carefully chosen hypothesis classes to exploit the structure in the data, these learnable regularization functions can prune the hypothesis space, and help reduce the amount of labeled data needed. We then provide two concrete examples of functional regularization, one using auto-encoders and the other using masked self-supervision, and apply our framework to quantify the reduction in the sample complexity bound of labeled data. We also provide complementary empirical results to support our analysis.

Siddhant Garg, Adarsh Kumar, Vibhor Goel et al.

Adversarial machine learning has exposed several security hazards of neural models and has become an important research topic in recent times. Thus far, the concept of an "adversarial perturbation" has exclusively been used with reference to the input space referring to a small, imperceptible change which can cause a ML model to err. In this work we extend the idea of "adversarial perturbations" to the space of model weights, specifically to inject backdoors in trained DNNs, which exposes a security risk of using publicly available trained models. Here, injecting a backdoor refers to obtaining a desired outcome from the model when a trigger pattern is added to the input, while retaining the original model predictions on a non-triggered input. From the perspective of an adversary, we characterize these adversarial perturbations to be constrained within an $\ell_{\infty}$ norm around the original model weights. We introduce adversarial perturbations in the model weights using a composite loss on the predictions of the original model and the desired trigger through projected gradient descent. We empirically show that these adversarial weight perturbations exist universally across several computer vision and natural language processing tasks. Our results show that backdoors can be successfully injected with a very small average relative change in model weight values for several applications.

Siddhant Garg, Goutham Ramakrishnan

The last few decades have seen significant breakthroughs in the fields of deep learning and quantum computing. Research at the junction of the two fields has garnered an increasing amount of interest, which has led to the development of quantum deep learning and quantum-inspired deep learning techniques in recent times. In this work, we present an overview of advances in the intersection of quantum computing and deep learning by discussing the technical contributions, strengths and similarities of various research works in this domain. To this end, we review and summarise the different schemes proposed to model quantum neural networks (QNNs) and other variants like quantum convolutional networks (QCNNs). We also briefly describe the recent progress in quantum inspired classic deep learning algorithms and their applications to natural language processing.

Siddhant Garg, Rohit Kumar Sharma, Yingyu Liang

Fine-tuning (FT) pre-trained sentence embedding models on small datasets has been shown to have limitations. In this paper we show that concatenating the embeddings from the pre-trained model with those from a simple sentence embedding model trained only on the target data, can improve over the performance of FT for few-sample tasks. To this end, a linear classifier is trained on the combined embeddings, either by freezing the embedding model weights or training the classifier and embedding models end-to-end. We perform evaluation on seven small datasets from NLP tasks and show that our approach with end-to-end training outperforms FT with negligible computational overhead. Further, we also show that sophisticated combination techniques like CCA and KCCA do not work as well in practice as concatenation. We provide theoretical analysis to explain this empirical observation.

Siddhant Garg, Goutham Ramakrishnan

Modern text classification models are susceptible to adversarial examples, perturbed versions of the original text indiscernible by humans which get misclassified by the model. Recent works in NLP use rule-based synonym replacement strategies to generate adversarial examples. These strategies can lead to out-of-context and unnaturally complex token replacements, which are easily identifiable by humans. We present BAE, a black box attack for generating adversarial examples using contextual perturbations from a BERT masked language model. BAE replaces and inserts tokens in the original text by masking a portion of the text and leveraging the BERT-MLM to generate alternatives for the masked tokens. Through automatic and human evaluations, we show that BAE performs a stronger attack, in addition to generating adversarial examples with improved grammaticality and semantic coherence as compared to prior work.

Marios Loizou, Siddhant Garg, Dmitry Petrov et al.

We present a deep learning method that propagates point-wise feature representations across shapes within a collection for the purpose of 3D shape segmentation. We propose a cross-shape attention mechanism to enable interactions between a shape's point-wise features and those of other shapes. The mechanism assesses both the degree of interaction between points and also mediates feature propagation across shapes, improving the accuracy and consistency of the resulting point-wise feature representations for shape segmentation. Our method also proposes a shape retrieval measure to select suitable shapes for cross-shape attention operations for each test shape. Our experiments demonstrate that our approach yields state-of-the-art results in the popular PartNet dataset.

Siddhant Garg, Thuy Vu, Alessandro Moschitti

We propose TANDA, an effective technique for fine-tuning pre-trained Transformer models for natural language tasks. Specifically, we first transfer a pre-trained model into a model for a general task by fine-tuning it with a large and high-quality dataset. We then perform a second fine-tuning step to adapt the transferred model to the target domain. We demonstrate the benefits of our approach for answer sentence selection, which is a well-known inference task in Question Answering. We built a large scale dataset to enable the transfer step, exploiting the Natural Questions dataset. Our approach establishes the state of the art on two well-known benchmarks, WikiQA and TREC-QA, achieving MAP scores of 92% and 94.3%, respectively, which largely outperform the previous highest scores of 83.4% and 87.5%, obtained in very recent work. We empirically show that TANDA generates more stable and robust models reducing the effort required for selecting optimal hyper-parameters. Additionally, we show that the transfer step of TANDA makes the adaptation step more robust to noise. This enables a more effective use of noisy datasets for fine-tuning. Finally, we also confirm the positive impact of TANDA in an industrial setting, using domain specific datasets subject to different types of noise.

Siddhant Garg, Aditya Kumar Akash

We explore a novel setting of the Multi-Armed Bandit (MAB) problem inspired from real world applications which we call bandits with "stochastic delayed composite anonymous feedback (SDCAF)". In SDCAF, the rewards on pulling arms are stochastic with respect to time but spread over a fixed number of time steps in the future after pulling the arm. The complexity of this problem stems from the anonymous feedback to the player and the stochastic generation of the reward. Due to the aggregated nature of the rewards, the player is unable to associate the reward to a particular time step from the past. We present two algorithms for this more complicated setting of SDCAF using phase based extensions of the UCB algorithm. We perform regret analysis to show sub-linear theoretical guarantees on both the algorithms.

Siddhant Garg

Recent works show that ordering of the training data affects the model performance for Neural Machine Translation. Several approaches involving dynamic data ordering and data sharding based on curriculum learning have been analysed for the their performance gains and faster convergence. In this work we propose to empirically study several ordering approaches for the training data based on different metrics and evaluate their impact on the model performance. Results from our study show that pre-fixing the ordering of the training data based on perplexity scores from a pre-trained model performs the best and outperforms the default approach of randomly shuffling the training data every epoch.