Arijit Sehanobish

Krzysztof Choromanski, Arijit Sehanobish, Han Lin et al. · cambridge

We present two new classes of algorithms for efficient field integration on graphs encoding point clouds. The first class, SeparatorFactorization(SF), leverages the bounded genus of point cloud mesh graphs, while the second class, RFDiffusion(RFD), uses popular epsilon-nearest-neighbor graph representations for point clouds. Both can be viewed as providing the functionality of Fast Multipole Methods (FMMs), which have had a tremendous impact on efficient integration, but for non-Euclidean spaces. We focus on geometries induced by distributions of walk lengths between points (e.g., shortest-path distance). We provide an extensive theoretical analysis of our algorithms, obtaining new results in structural graph theory as a byproduct. We also perform exhaustive empirical evaluation, including on-surface interpolation for rigid and deformable objects (particularly for mesh-dynamics modeling), Wasserstein distance computations for point clouds, and the Gromov-Wasserstein variant.

Alessandro Manenti, Avinava Dubey, Arijit Sehanobish et al.

We propose SWING: Space Walks for Implicit Network Graphs, a new class of algorithms for computations involving Graph Random Features on graphs given by implicit representations (i-graphs), where edge-weights are defined as bi-variate functions of feature vectors in the corresponding nodes. Those classes of graphs include several prominent examples, such as: $ε$-neighborhood graphs, used on regular basis in machine learning. Rather than conducting walks on graphs' nodes, those methods rely on walks in continuous spaces, in which those graphs are embedded. To accurately and efficiently approximate original combinatorial calculations, SWING applies customized Gumbel-softmax sampling mechanism with linearized kernels, obtained via random features coupled with importance sampling techniques. This algorithm is of its own interest. SWING relies on the deep connection between implicitly defined graphs and Fourier analysis, presented in this paper. SWING is accelerator-friendly and does not require input graph materialization. We provide detailed analysis of SWING and complement it with thorough experiments on different classes of i-graphs.

Arijit Sehanobish, Krzysztof Choromanski, Yunfan Zhao et al.

We introduce the concept of scalable neural network kernels (SNNKs), the replacements of regular feedforward layers (FFLs), capable of approximating the latter, but with favorable computational properties. SNNKs effectively disentangle the inputs from the parameters of the neural network in the FFL, only to connect them in the final computation via the dot-product kernel. They are also strictly more expressive, as allowing to model complicated relationships beyond the functions of the dot-products of parameter-input vectors. We also introduce the neural network bundling process that applies SNNKs to compactify deep neural network architectures, resulting in additional compression gains. In its extreme version, it leads to the fully bundled network whose optimal parameters can be expressed via explicit formulae for several loss functions (e.g. mean squared error), opening a possibility to bypass backpropagation. As a by-product of our analysis, we introduce the mechanism of the universal random features (or URFs), applied to instantiate several SNNK variants, and interesting on its own in the context of scalable kernel methods. We provide rigorous theoretical analysis of all these concepts as well as an extensive empirical evaluation, ranging from point-wise kernel estimation to Transformers' fine-tuning with novel adapter layers inspired by SNNKs. Our mechanism provides up to 5x reduction in the number of trainable parameters, while maintaining competitive accuracy.

Sang Min Kim, Byeongchan Kim, Arijit Sehanobish et al.

Efficient neural networks are essential for scaling machine learning models to real-time applications and resource-constrained environments. Fully-connected feedforward layers (FFLs) introduce computation and parameter count bottlenecks within neural network architectures. To address this challenge, in this work, we propose a new class of dense layers that generalize standard fully-connected feedforward layers, \textbf{E}fficient, \textbf{U}nified and \textbf{Gen}eral dense layers (EUGens). EUGens leverage random features to approximate standard FFLs and go beyond them by incorporating a direct dependence on the input norms in their computations. The proposed layers unify existing efficient FFL extensions and improve efficiency by reducing inference complexity from quadratic to linear time. They also lead to \textbf{the first} unbiased algorithms approximating FFLs with arbitrary polynomial activation functions. Furthermore, EuGens reduce the parameter count and computational overhead while preserving the expressive power and adaptability of FFLs. We also present a layer-wise knowledge transfer technique that bypasses backpropagation, enabling efficient adaptation of EUGens to pre-trained models. Empirically, we observe that integrating EUGens into Transformers and MLPs yields substantial improvements in inference speed (up to \textbf{27}\%) and memory efficiency (up to \textbf{30}\%) across a range of tasks, including image classification, language model pre-training, and 3D scene reconstruction. Overall, our results highlight the potential of EUGens for the scalable deployment of large-scale neural networks in real-world scenarios.

Arijit Sehanobish, McCullen Sandora, Nabila Abraham et al.

Pretrained Transformer based models finetuned on domain specific corpora have changed the landscape of NLP. However, training or fine-tuning these models for individual tasks can be time consuming and resource intensive. Thus, a lot of current research is focused on using transformers for multi-task learning (Raffel et al.,2020) and how to group the tasks to help a multi-task model to learn effective representations that can be shared across tasks (Standley et al., 2020; Fifty et al., 2021). In this work, we show that a single multi-tasking model can match the performance of task specific models when the task specific models show similar representations across all of their hidden layers and their gradients are aligned, i.e. their gradients follow the same direction. We hypothesize that the above observations explain the effectiveness of multi-task learning. We validate our observations on our internal radiologist-annotated datasets on the cervical and lumbar spine. Our method is simple and intuitive, and can be used in a wide range of NLP problems.

Arijit Sehanobish, Kawshik Kannan, Nabila Abraham et al.

Large pretrained Transformer-based language models like BERT and GPT have changed the landscape of Natural Language Processing (NLP). However, fine tuning such models still requires a large number of training examples for each target task, thus annotating multiple datasets and training these models on various downstream tasks becomes time consuming and expensive. In this work, we propose a simple extension of the Prototypical Networks for few-shot text classification. Our main idea is to replace the class prototypes by Gaussians and introduce a regularization term that encourages the examples to be clustered near the appropriate class centroids. Experimental results show that our method outperforms various strong baselines on 13 public and 4 internal datasets. Furthermore, we use the class distributions as a tool for detecting potential out-of-distribution (OOD) data points during deployment.

Arijit Sehanobish, Nathaniel Brown, Ishita Daga et al.

Pretrained Transformer based models finetuned on domain specific corpora have changed the landscape of NLP. Generally, if one has multiple tasks on a given dataset, one may finetune different models or use task specific adapters. In this work, we show that a multi-task model can beat or achieve the performance of multiple BERT-based models finetuned on various tasks and various task specific adapter augmented BERT-based models. We validate our method on our internal radiologist's report dataset on cervical spine. We hypothesize that the tasks are semantically close and related and thus multitask learners are powerful classifiers. Our work opens the scope of using our method to radiologist's reports on various body parts.

Arijit Sehanobish, Charles Lovering

We study the \textit{parameter placement problem}: given a fixed budget of $k$ trainable entries within the B matrix of a LoRA adapter (A frozen), does the choice of which $k$ matter? Under supervised fine-tuning, random and informed subsets achieve comparable performance. Under GRPO on base models, random placement fails to improve over the base model, while gradient-informed placement recovers standard LoRA accuracy. This regime dependence traces to gradient structure: SFT gradients are low-rank and directionally stable, so any subset accumulates coherent updates; GRPO gradients are high-rank and near-orthogonal across steps, so only elements with consistently signed gradients retain the learning signal. Our scoring procedure identifies these critical parameters in under 10 seconds at less than 0.5% of training cost. Selected parameters concentrate on residual-stream-writing projections (V, O, Down), stable across model families and scales (1.5B - 8B).

Byeongchan Kim, Arijit Sehanobish, Avinava Dubey et al.

We present a new class of efficient attention mechanisms applying universal 3D Relative Positional Encoding (RPE) methods given by arbitrary integrable modulation functions $f$. They lead to the new class of 3D-Transformer models, called \textit{RelFlexformers}, flexibly integrating those RPEs, and characterized by the $O(L \log L)$ time complexity of the attention computation for the $L$-length input sequences. RelFlexformers builds on the theory of the Non-Uniform Fourier Transform (NU-FFT), naturally generalizing several existing efficient RPE-attention methods from structured settings with tokens homogeneously embedded in unweighted grids into general non-structured heterogeneous scenarios, where tokens' positions are arbitrarily distributed in the corresponding 3D spaces. As such, RelFlexformers can be applied in particular to model point clouds. Our extensive empirical evaluation on a large portfolio of 3D datasets confirms quality improvements provided by the NU-FFT-driven attention modulation techniques in the RelFlexformers.

Krzysztof Choromanski, Isaac Reid, Arijit Sehanobish et al.

We present the first linear time complexity randomized algorithms for unbiased approximation of the celebrated family of general random walk kernels (RWKs) for sparse graphs. This includes both labelled and unlabelled instances. The previous fastest methods for general RWKs were of cubic time complexity and not applicable to labelled graphs. Our method samples dependent random walks to compute novel graph embeddings in $\mathbb{R}^d$ whose dot product is equal to the true RWK in expectation. It does so without instantiating the direct product graph in memory, meaning we can scale to massive datasets that cannot be stored on a single machine. We derive exponential concentration bounds to prove that our estimator is sharp, and show that the ability to approximate general RWKs (rather than just special cases) unlocks efficient implicit graph kernel learning. Our method is up to $\mathbf{27\times}$ faster than its counterparts for efficient computation on large graphs and scales to graphs $\mathbf{128 \times}$ bigger than largest examples amenable to brute-force computation.

Krzysztof Choromanski, Avinava Dubey, Arijit Sehanobish et al.

We propose refined GRFs (GRFs++), a new class of Graph Random Features (GRFs) for efficient and accurate computations involving kernels defined on the nodes of a graph. GRFs++ resolve some of the long-standing limitations of regular GRFs, including difficulty modeling relationships between more distant nodes. They reduce dependence on sampling long graph random walks via a novel walk-stitching technique, concatenating several shorter walks without breaking unbiasedness. By applying these techniques, GRFs++ inherit the approximation quality provided by longer walks but with greater efficiency, trading sequential, inefficient sampling of a long walk for parallel computation of short walks and matrix-matrix multiplication. Furthermore, GRFs++ extend the simplistic GRFs walk termination mechanism (Bernoulli schemes with fixed halting probabilities) to a broader class of strategies, applying general distributions on the walks' lengths. This improves the approximation accuracy of graph kernels, without incurring extra computational cost. We provide empirical evaluations to showcase all our claims and complement our results with theoretical analysis.

Isaac Reid, Arijit Sehanobish, Cederik Höfs et al.

We introduce WIRE: Wavelet-Induced Rotary Encodings. WIRE extends Rotary Position Encodings (RoPE), a popular algorithm in LLMs and ViTs, to graph-structured data. We demonstrate that WIRE is more general than RoPE, recovering the latter in the special case of grid graphs. WIRE also enjoys a host of desirable theoretical properties, including equivariance under node ordering permutation, compatibility with linear attention, and (under select assumptions) asymptotic dependence on graph resistive distance. We test WIRE on a range of synthetic and real-world tasks, including identifying monochromatic subgraphs, semantic segmentation of point clouds, and more standard graph benchmarks. We find it to be effective in settings where the underlying graph structure is important.

Neal G. Ravindra, Arijit Sehanobish

High-quality, medically validated labels exist for clinical actigraphy data but not for ubiquitous consumer wearables like the Apple Watch. Manually labeling wearables data is expensive and doesn't scale. This paper offers a novel framework that transfers valuable labels from a source domain (e.g., actigraphy) to a target domain (e.g., Apple Watch) without requiring paired data. Instead of working with raw time-series signals, we project both domains into a shared latent embedding space using time-series foundation models (TSFMs) and develop a new framework to align the cross-device representations. Our method, Adversarial Alignment of TSFM Embeddings forces the distributions of source and target embeddings to align within this space, facilitating label transfer across device type.

Sang Min Kim, Byeongchan Kim, Arijit Sehanobish et al. · deepmind

Improving the efficiency and performance of implicit neural representations in 3D, particularly Neural Radiance Fields (NeRF) and Signed Distance Fields (SDF) is crucial for enabling their use in real-time applications. These models, while capable of generating photo-realistic novel views and detailed 3D reconstructions, often suffer from high computational costs and slow inference times. To address this, we introduce a novel neural network layer called the "magnituder", designed to reduce the number of training parameters in these models without sacrificing their expressive power. By integrating magnituders into standard feed-forward layer stacks, we achieve improved inference speed and adaptability. Furthermore, our approach enables a zero-shot performance boost in trained implicit neural representation models through layer-wise knowledge transfer without backpropagation, leading to more efficient scene reconstruction in dynamic environments.

Arijit Sehanobish, Avinava Dubey, Krzysztof Choromanski et al.

Recent efforts to scale Transformer models have demonstrated rapid progress across a wide range of tasks (Wei et al., 2022). However, fine-tuning these models for downstream tasks is expensive due to their large parameter counts. Parameter-efficient fine-tuning (PEFT) approaches have emerged as a viable alternative by allowing us to fine-tune models by updating only a small number of parameters. In this work, we propose a general framework for parameter efficient fine-tuning (PEFT), based on structured unrestricted-rank matrices (SURM) which can serve as a drop-in replacement for popular approaches such as Adapters and LoRA. Unlike other methods like LoRA, SURMs provides more flexibility in finding the right balance between compactness and expressiveness. This is achieved by using low displacement rank matrices (LDRMs), which hasn't been used in this context before. SURMs remain competitive with baselines, often providing significant quality improvements while using a smaller parameter budget. SURMs achieve 5-7% accuracy gains on various image classification tasks while replacing low-rank matrices in LoRA. It also results in up to 12x reduction of the number of parameters in adapters (with virtually no loss in quality) on the GLUE benchmark.

Somnath Basu Roy Chowdhury, Krzysztof Choromanski, Arijit Sehanobish et al.

Machine unlearning is the process of efficiently removing the influence of a training data instance from a trained machine learning model without retraining it from scratch. A popular subclass of unlearning approaches is exact machine unlearning, which focuses on techniques that explicitly guarantee the removal of the influence of a data instance from a model. Exact unlearning approaches use a machine learning model in which individual components are trained on disjoint subsets of the data. During deletion, exact unlearning approaches only retrain the affected components rather than the entire model. While existing approaches reduce retraining costs, it can still be expensive for an organization to retrain a model component as it requires halting a system in production, which leads to service failure and adversely impacts customers. To address these challenges, we introduce an exact unlearning framework -- Sequence-aware Sharded Sliced Training (S3T), which is designed to enhance the deletion capabilities of an exact unlearning system while minimizing the impact on model's performance. At the core of S3T, we utilize a lightweight parameter-efficient fine-tuning approach that enables parameter isolation by sequentially training layers with disjoint data slices. This enables efficient unlearning by simply deactivating the layers affected by data deletion. Furthermore, to reduce the retraining cost and improve model performance, we train the model on multiple data sequences, which allows S3T to handle an increased number of deletion requests. Both theoretically and empirically, we demonstrate that S3T attains superior deletion capabilities and enhanced performance compared to baselines across a wide range of settings.

Krzysztof Choromanski, Arijit Sehanobish, Somnath Basu Roy Chowdhury et al.

We present a new class of fast polylog-linear algorithms based on the theory of structured matrices (in particular low displacement rank) for integrating tensor fields defined on weighted trees. Several applications of the resulting fast tree-field integrators (FTFIs) are presented, including (a) approximation of graph metrics with tree metrics, (b) graph classification, (c) modeling on meshes, and finally (d) Topological Transformers (TTs) (Choromanski et al., 2022) for images. For Topological Transformers, we propose new relative position encoding (RPE) masking mechanisms with as few as three extra learnable parameters per Transformer layer, leading to 1.0-1.5%+ accuracy gains. Importantly, most of FTFIs are exact methods, thus numerically equivalent to their brute-force counterparts. When applied to graphs with thousands of nodes, those exact algorithms provide 5.7-13x speedups. We also provide an extensive theoretical analysis of our methods.

Krzysztof Choromanski, Haoxian Chen, Han Lin et al.

We propose a new class of random feature methods for linearizing softmax and Gaussian kernels called hybrid random features (HRFs) that automatically adapt the quality of kernel estimation to provide most accurate approximation in the defined regions of interest. Special instantiations of HRFs lead to well-known methods such as trigonometric (Rahimi and Recht, 2007) or (recently introduced in the context of linear-attention Transformers) positive random features (Choromanski et al., 2021). By generalizing Bochner's Theorem for softmax/Gaussian kernels and leveraging random features for compositional kernels, the HRF-mechanism provides strong theoretical guarantees - unbiased approximation and strictly smaller worst-case relative errors than its counterparts. We conduct exhaustive empirical evaluation of HRF ranging from pointwise kernel estimation experiments, through tests on data admitting clustering structure to benchmarking implicit-attention Transformers (also for downstream Robotics applications), demonstrating its quality in a wide spectrum of machine learning problems.

Onur Kara, Arijit Sehanobish, Hector H Corzo

Transformers are state-of-the-art deep learning models that are composed of stacked attention and point-wise, fully connected layers designed for handling sequential data. Transformers are not only ubiquitous throughout Natural Language Processing (NLP), but, recently, they have inspired a new wave of Computer Vision (CV) applications research. In this work, a Vision Transformer (ViT) is applied to predict the state variables of 2-dimensional Ising model simulations. Our experiments show that ViT outperform state-of-the-art Convolutional Neural Networks (CNN) when using a small number of microstate images from the Ising model corresponding to various boundary conditions and temperatures. This work opens the possibility of applying ViT to other simulations, and raises interesting research directions on how attention maps can learn about the underlying physics governing different phenomena.

Krzysztof Choromanski, Han Lin, Haoxian Chen et al.

In this paper we provide, to the best of our knowledge, the first comprehensive approach for incorporating various masking mechanisms into Transformers architectures in a scalable way. We show that recent results on linear causal attention (Choromanski et al., 2021) and log-linear RPE-attention (Luo et al., 2021) are special cases of this general mechanism. However by casting the problem as a topological (graph-based) modulation of unmasked attention, we obtain several results unknown before, including efficient d-dimensional RPE-masking and graph-kernel masking. We leverage many mathematical techniques ranging from spectral analysis through dynamic programming and random walks to new algorithms for solving Markov processes on graphs. We provide a corresponding empirical evaluation.

Hector H. Corzo, Arijit Sehanobish, Onur Kara

In this report, we present a deep learning framework termed the Electron Correlation Potential Neural Network (eCPNN) that can learn succinct and compact potential functions. These functions can effectively describe the complex instantaneous spatial correlations among electrons in many--electron atoms. The eCPNN was trained in an unsupervised manner with limited information from Full Configuration Interaction (FCI) one--electron density functions within predefined limits of accuracy. Using the effective correlation potential functions generated by eCPNN, we can predict the total energies of each of the studied atomic systems with a remarkable accuracy when compared to FCI energies.

Arijit Sehanobish, Neal Ravindra, David van Dijk

Understanding the space of probability measures on a metric space equipped with a Wasserstein distance is one of the fundamental questions in mathematical analysis. The Wasserstein metric has received a lot of attention in the machine learning community especially for its principled way of comparing distributions. In this work, we use a permutation invariant network to map samples from probability measures into a low-dimensional space such that the Euclidean distance between the encoded samples reflects the Wasserstein distance between probability measures. We show that our network can generalize to correctly compute distances between unseen densities. We also show that these networks can learn the first and the second moments of probability distributions.

Arijit Sehanobish, Neal G. Ravindra, David van Dijk

Graph Neural Networks (GNN) have been extensively used to extract meaningful representations from graph structured data and to perform predictive tasks such as node classification and link prediction. In recent years, there has been a lot of work incorporating edge features along with node features for prediction tasks. One of the main difficulties in using edge features is that they are often handcrafted, hard to get, specific to a particular domain, and may contain redundant information. In this work, we present a framework for creating new edge features, applicable to any domain, via a combination of self-supervised and unsupervised learning. In addition to this, we use Forman-Ricci curvature as an additional edge feature to encapsulate the local geometry of the graph. We then encode our edge features via a Set Transformer and combine them with node features extracted from popular GNN architectures for node classification in an end-to-end training scheme. We validate our work on three biological datasets comprising of single-cell RNA sequencing data of neurological disease, \textit{in vitro} SARS-CoV-2 infection, and human COVID-19 patients. We demonstrate that our method achieves better performance on node classification tasks over baseline Graph Attention Network (GAT) and Graph Convolutional Network (GCN) models. Furthermore, given the attention mechanism on edge and node features, we are able to interpret the cell types and genes that determine the course and severity of COVID-19, contributing to a growing list of potential disease biomarkers and therapeutic targets.

Arijit Sehanobish, Hector H. Corzo, Onur Kara et al.

Attempts to apply Neural Networks (NN) to a wide range of research problems have been ubiquitous and plentiful in recent literature. Particularly, the use of deep NNs for understanding complex physical and chemical phenomena has opened a new niche of science where the analysis tools from Machine Learning (ML) are combined with the computational concepts of the natural sciences. Reports from this unification of ML have presented evidence that NNs can learn classical Hamiltonian mechanics. This application of NNs to classical physics and its results motivate the following question: Can NNs be endowed with inductive biases through observation as means to provide insights into quantum phenomena? In this work, this question is addressed by investigating possible approximations for reconstructing the Hamiltonian of a quantum system in an unsupervised manner by using only limited information obtained from the system's probability distribution.

Arijit Sehanobish, Neal G. Ravindra, David van Dijk

A molecular and cellular understanding of how SARS-CoV-2 variably infects and causes severe COVID-19 remains a bottleneck in developing interventions to end the pandemic. We sought to use deep learning to study the biology of SARS-CoV-2 infection and COVID-19 severity by identifying transcriptomic patterns and cell types associated with SARS-CoV-2 infection and COVID-19 severity. To do this, we developed a new approach to generating self-supervised edge features. We propose a model that builds on Graph Attention Networks (GAT), creates edge features using self-supervised learning, and ingests these edge features via a Set Transformer. This model achieves significant improvements in predicting the disease state of individual cells, given their transcriptome. We apply our model to single-cell RNA sequencing datasets of SARS-CoV-2 infected lung organoids and bronchoalveolar lavage fluid samples of patients with COVID-19, achieving state-of-the-art performance on both datasets with our model. We then borrow from the field of explainable AI (XAI) to identify the features (genes) and cell types that discriminate bystander vs. infected cells across time and moderate vs. severe COVID-19 disease. To the best of our knowledge, this represents the first application of deep learning to identifying the molecular and cellular determinants of SARS-CoV-2 infection and COVID-19 severity using single-cell omics data.

Neal G. Ravindra, Arijit Sehanobish, Jenna L. Pappalardo et al.

Single-cell RNA sequencing (scRNA-seq) has revolutionized biological discovery, providing an unbiased picture of cellular heterogeneity in tissues. While scRNA-seq has been used extensively to provide insight into both healthy systems and diseases, it has not been used for disease prediction or diagnostics. Graph Attention Networks (GAT) have proven to be versatile for a wide range of tasks by learning from both original features and graph structures. Here we present a graph attention model for predicting disease state from single-cell data on a large dataset of Multiple Sclerosis (MS) patients. MS is a disease of the central nervous system that can be difficult to diagnose. We train our model on single-cell data obtained from blood and cerebrospinal fluid (CSF) for a cohort of seven MS patients and six healthy adults (HA), resulting in 66,667 individual cells. We achieve 92 % accuracy in predicting MS, outperforming other state-of-the-art methods such as a graph convolutional network and a random forest classifier. Further, we use the learned graph attention model to get insight into the features (cell types and genes) that are important for this prediction. The graph attention model also allow us to infer a new feature space for the cells that emphasizes the differences between the two conditions. Finally we use the attention weights to learn a new low-dimensional embedding that can be visualized. To the best of our knowledge, this is the first effort to use graph attention, and deep learning in general, to predict disease state from single-cell data. We envision applying this method to single-cell data for other diseases.

Arijit Sehanobish, Chan Hee Song

Most Named Entity Recognition (NER) systems use additional features like part-of-speech (POS) tags, shallow parsing, gazetteers, etc. Such kind of information requires external knowledge like unlabeled texts and trained taggers. Adding these features to NER systems have been shown to have a positive impact. However, sometimes creating gazetteers or taggers can take a lot of time and may require extensive data cleaning. In this paper for Chinese NER systems, we do not use these traditional features but we use lexicographic features of Chinese characters. Chinese characters are composed of graphical components called radicals and these components often have some semantic indicators. We propose CNN based models that incorporate this semantic information and use them for NER. Our models show an improvement over the baseline BERT-BiLSTM-CRF model. We set a new baseline score for Chinese OntoNotes v5.0 and show an improvement of +.64 F1 score. We present a state-of-the-art F1 score on Weibo dataset of 71.81 and show a competitive improvement of +0.72 over baseline on ResumeNER dataset.