State Space Models (SSMs) have recently emerged as an alternative to Vision Transformers (ViTs) due to their unique ability of modeling global relationships with linear complexity. SSMs are specifically designed to capture spatially proximate relationships of image patches. However, they fail to identify relationships between conceptually related yet not adjacent patches. This limitation arises from the non-causal nature of image data, which lacks inherent directional relationships. Additionally, current vision-based SSMs are highly sensitive to transformations such as rotation. Their predefined scanning directions depend on the original image orientation, which can cause the model to produce inconsistent patch-processing sequences after rotation. To address these limitations, we introduce Spectral VMamba, a novel approach that effectively captures the global structure within an image by leveraging spectral information derived from the graph Laplacian of image patches. Through spectral decomposition, our approach encodes patch relationships independently of image orientation, achieving patch traversal rotation invariance with our Rotational Feature Normalizer (RFN) module. Our experiments on classification tasks show that Spectral VMamba outperforms the leading SSM models in vision, such as VMamba, while maintaining invariance to rotations and a providing a similar runtime efficiency. The implementation is available at: https://github.com/Sahardastani/spectral_vmamba.git.

Rapid advancements in GPU computational power has outpaced memory capacity and bandwidth growth, creating bottlenecks in Large Language Model (LLM) inference. Post-training quantization is the leading method for addressing memory-related bottlenecks in LLM inference, but it suffers from significant performance degradation below 4-bit precision. This paper addresses these challenges by investigating the pretraining of low-bitwidth models specifically Ternary Language Models (TriLMs) as an alternative to traditional floating-point models (FloatLMs) and their post-training quantized versions (QuantLMs). We present Spectra LLM suite, the first open suite of LLMs spanning multiple bit-widths, including FloatLMs, QuantLMs, and TriLMs, ranging from 99M to 3.9B parameters trained on 300B tokens. Our comprehensive evaluation demonstrates that TriLMs offer superior scaling behavior in terms of model size (in bits). Surprisingly, at scales exceeding one billion parameters, TriLMs consistently outperform their QuantLM and FloatLM counterparts for a given bit size across various benchmarks. Notably, the 3.9B parameter TriLM matches the performance of the FloatLM 3.9B across all benchmarks, despite having fewer bits than FloatLM 830M. Overall, this research provides valuable insights into the feasibility and scalability of low-bitwidth language models, paving the way for the development of more efficient LLMs.

Scalable Vector Graphics (SVG) offer a powerful format for representing visual designs as interpretable code. Recent advances in vision-language models (VLMs) have enabled high-quality SVG generation by framing the problem as a code generation task and leveraging large-scale pretraining. VLMs are particularly suitable for this task as they capture both global semantics and fine-grained visual patterns, while transferring knowledge across vision, natural language, and code domains. However, existing VLM approaches often struggle to produce faithful and efficient SVGs because they never observe the rendered images during training. Although differentiable rendering for autoregressive SVG code generation remains unavailable, rendered outputs can still be compared to original inputs, enabling evaluative feedback suitable for reinforcement learning (RL). We introduce Reinforcement Learning from Rendering Feedback, an RL method that enhances SVG generation in autoregressive VLMs by leveraging feedback from rendered SVG outputs. Given an input image, the model generates SVG roll-outs that are rendered and compared to the original image to compute a reward. This visual fidelity feedback guides the model toward producing more accurate, efficient, and semantically coherent SVGs. \method significantly outperforms supervised fine-tuning, addressing common failure modes and enabling precise, high-quality SVG generation with strong structural understanding and generalization.

Understanding human motion from video is essential for a range of applications, including pose estimation, mesh recovery and action recognition. While state-of-the-art methods predominantly rely on transformer-based architectures, these approaches have limitations in practical scenarios. Transformers are slower when sequentially predicting on a continuous stream of frames in real-time, and do not generalize to new frame rates. In light of these constraints, we propose a novel attention-free spatiotemporal model for human motion understanding building upon recent advancements in state space models. Our model not only matches the performance of transformer-based models in various motion understanding tasks but also brings added benefits like adaptability to different video frame rates and enhanced training speed when working with longer sequences of keypoints. Moreover, the proposed model supports both offline and real-time applications. For real-time sequential prediction, our model is both memory efficient and several times faster than transformer-based approaches while maintaining their high accuracy.

The accurate modeling of dynamics in interactive environments is critical for successful long-range prediction. Such a capability could advance Reinforcement Learning (RL) and Planning algorithms, but achieving it is challenging. Inaccuracies in model estimates can compound, resulting in increased errors over long horizons. We approach this problem from the lens of Koopman theory, where the nonlinear dynamics of the environment can be linearized in a high-dimensional latent space. This allows us to efficiently parallelize the sequential problem of long-range prediction using convolution while accounting for the agent's action at every time step. Our approach also enables stability analysis and better control over gradients through time. Taken together, these advantages result in significant improvement over the existing approaches, both in the efficiency and the accuracy of modeling dynamics over extended horizons. We also show that this model can be easily incorporated into dynamics modeling for model-based planning and model-free RL and report promising experimental results.

Class imbalance is a common phenomenon in multiple application domains such as healthcare, where the sample occurrence of one or few class categories is more prevalent in the dataset than the rest. This work addresses the class-imbalance issue by proposing an over-sampling method for the minority classes in the latent space of a Regularized Auto-Encoder (RAE). Specifically, we construct a latent space by maximizing the conditional data likelihood using an Encoder-Decoder structure, such that oversampling through convex combinations of latent samples preserves the class identity. A jointly-trained linear classifier that separates convexly coupled latent vectors from different classes is used to impose this property on the AE’s latent space. Further, the aforesaid linear classifier is used for final classification without retraining. We theoretically show that our method can achieve a low variance risk estimate compared to naive oversampling methods and is robust to overfitting. We conduct several experiments on benchmark datasets and show that our method outperforms the existing oversampling techniques for handling class imbalance. The code of the proposed method is available at: https://github.com/arnabkmondal/oversamplingrae.

Harman Singh, Poorva Garg, Mohit Gupta, Kevin Shah, Ashish Goswami, Satyam Modi, Arnab Mondal, Dinesh Khandelwal, Dinesh Garg, Parag Singla. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Symmetry-based neural networks often constrain the architecture in order to achieve invariance or equivariance to a group of transformations. In this paper, we propose an alternative that avoids this architectural constraint by learning to produce canonical representations of the data. These canonicalization functions can readily be plugged into non-equivariant backbone architectures. We offer explicit ways to implement them for some groups of interest. We show that this approach enjoys universality while providing interpretable insights. Our main hypothesis, supported by our empirical results, is that learning a small neural network to perform canonicalization is better than using predefined heuristics. Our experiments show that learning the canonicalization function is competitive with existing techniques for learning equivariant functions across many tasks, including image classification, $N$-body dynamics prediction, point cloud classification and part segmentation, while being faster across the board.

Equivariant networks are specifically designed to ensure consistent behavior with respect to a set of input transformations, leading to higher sample efficiency and more accurate and robust predictions. However, redesigning each component of prevalent deep neural network architectures to achieve chosen equivariance is a difficult problem and can result in a computationally expensive network during both training and inference. A recently proposed alternative towards equivariance that removes the architectural constraints is to use a simple canonicalization network that transforms the input to a canonical form before feeding it to an unconstrained prediction network. We show here that this approach can effectively be used to make a large pretrained network equivariant. However, we observe that the produced canonical orientations can be misaligned with those of the training distribution, hindering performance. Using dataset-dependent priors to inform the canonicalization function, we are able to make large pretrained models equivariant while maintaining their performance. This significantly improves the robustness of these models to deterministic transformations of the data, such as rotations. We believe this equivariant adaptation of large pretrained models can help their domain-specific applications with known symmetry priors.

Single-cell RNA sequencing (scRNA-seq) is a revolutionary methodology that helps to analyze transcriptome or genome information from a single cell. However, high dimensionality and sparsity in data due to dropout events pose computational challenges for existing state-of-the-art scRNA-seq clustering methods. Learning efficient representations becomes even more challenging due to the presence of noise in scRNA-seq data. To overcome the effect of noise and learn effective representations, this paper proposes sc-INDC (Single-Cell Information Maximized Noise-Invariant Deep Clustering), a deep neural network that facilitates learning of informative and noise-invariant representations of scRNA-seq data. Furthermore, the time complexity of the proposed sc-INDC is significantly lower compared to state-of-the-art scRNA-seq clustering methods. Extensive experimentation on fourteen publicly available scRNA-seq datasets illustrates the efficacy of the proposed model. Additionally, visualizations of t-SNE plots and several ablation studies are also conducted to provide insights into the improved representation ability of sc-INDC. Code of the proposed sc-INDC will be available at: https://github.com/arnabkmondal/sc-INDC.

We study a variety of notions of equivariance as an inductive bias in Reinforcement Learning (RL). In particular, we propose new mechanisms for learning representations that are equivariant to both the agent’s action, as well as symmetry transformations of the state-action pairs. Whereas prior work on exploiting symmetries in deep RL can only incorporate predefined linear transformations, our approach allows non-linear symmetry transformations of state-action pairs to be learned from the data. This is achieved through 1) equivariant Lie algebraic parameterization of state and action encodings, 2) equivariant latent transition models, and 3) the incorporation of symmetry-based losses. We demonstrate the advantages of our method, which we call Equivariant representations for RL (EqR), for Atari games in a data-efficient setting limited to 100K steps of interactions with the environment.

Self-Supervised Learning (SSL) with large-scale unlabelled datasets enables learning useful representations for multiple downstream tasks. However, assessing the quality of such representations efficiently poses nontrivial challenges. Existing approaches train linear probes (with frozen features) to evaluate performance on a given task. This is expensive both computationally, since it requires retraining a new prediction head for each downstream task, and statistically, requires task-specific labels for multiple tasks. This poses a natural question, how do we efficiently determine the "goodness" of representations learned with SSL across a wide range of potential downstream tasks? In particular, a task-agnostic statistical measure of representation quality, that predicts generalization without explicit downstream task evaluation, would be highly desirable. In this work, we analyze characteristics of learned representations $\mathbf{f_\theta}$, in well-trained neural networks with canonical architectures \& across SSL objectives. We observe that the eigenspectrum of the empirical feature covariance $\mathrm{Cov}(\mathbf{f_\theta}$) can be well approximated with the family of power-law distribution. We analytically and empirically (using multiple datasets, e.g. CIFAR, STL10, MIT67, ImageNet) demonstrate that the decay coefficient $\alpha$ serves as a measure of representation quality for tasks that are solvable with a linear readout, i.e. there exist well-defined intervals for $\alpha$ where models exhibit excellent downstream generalization. Furthermore, our experiments suggest that key design parameters in SSL algorithms, such as BarlowTwins, implicitly modulate the decay coefficient of the eigenspectrum ($\alpha$). As $\alpha$ depends only on the features themselves, this measure for model selection with hyperparameter tuning for BarlowTwins enables search with less compute.

A finite set of invariants can identify many interesting transformation groups. For example, distances, inner products and angles are preserved by Euclidean, Orthogonal and Conformal transformations, respectively. In an equivariant representation, the group invariants should remain constant on the embedding as we transform the input. This gives a procedure for learning equivariant representations without knowing the possibly nonlinear action of the group in the input space. Rather than enforcing such hard invariance constraints on the latent space, we show how to use invariants for "symmetry regularization" of the latent, while guaranteeing equivariance through other means. We also show the feasibility of learning disentangled representations using this approach and provide favorable qualitative and quantitative results on downstream tasks, including world modeling and reinforcement learning.

Clustering single-cell RNA sequence (scRNA-seq) data poses statistical and computational challenges due to their high-dimensionality and data-sparsity, also known as 'dropout' events. Recently, Regularized Auto-Encoder (RAE) based deep neural network models have achieved remarkable success in learning robust low-dimensional representations. The basic idea in RAEs is to learn a non-linear mapping from the high-dimensional data space to a low-dimensional latent space and vice-versa, simultaneously imposing a distributional prior on the latent space, which brings in a regularization effect. This paper argues that RAEs suffer from the infamous problem of bias-variance trade-off in their naive formulation. While a simple AE wita latent regularization results in data over-fitting, a very strong prior leads to under-representation and thus bad clustering. To address the above issues, we propose a modified RAE framework (called the scRAE) for effective clustering of the single-cell RNA sequencing data. scRAE consists of deterministic AE with a flexibly learnable prior generator network, which is jointly trained with the AE. This facilitates scRAE to trade-off better between the bias and variance in the latent space. We demonstrate the efficacy of the proposed method through extensive experimentation on several real-world single-cell Gene expression datasets. The code for our work is available at https://github.com/arnabkmondal/scRAE.

Auto-Encoder (AE) based neural generative frameworks model the joint-distribution between the data and the latent space using an Encoder-Decoder pair, with regularization imposed in terms of a prior over the latent space. Despite their advantages, such as stability in training, efficient inference, the performance of AE based models has not reached the superior standards of the other generative models such as Generative Adversarial Networks (GANs). Motivated by this, we examine the effect of the latent prior on the generation quality of deterministic AE models in this paper. Specifically, we consider the class of Generative AE models with deterministic Encoder-Decoder pair (such as Wasserstein Auto-Encoder (WAE), Adversarial Auto-Encoder (AAE)), and show that having a fixed prior distribution, a priori, oblivious to the dimensionality of the ‘true’ latent space, will lead to the infeasibility of the optimization problem considered. As a remedy to the issue mentioned above, we introduce an additional state space in the form of flexibly learnable latent priors, in the optimization objective of WAE/AAE. Additionally, we employ a latent-space interpolation based smoothing scheme to address the non-smoothness that may arise from highly flexible priors. We show the efficacy of our proposed models, called FlexAE and FlexAE-SR, through several experiments on multiple datasets, and demonstrate that FlexAE-SR is the new state-of-the-art for the AE based generative models in terms of generation quality as measured by several metrics such as Fr\’echet Inception Distance, Precision/Recall score.

Estimation of information theoretic quantities such as mutual information and its conditional variant has drawn interest in recent times owing to their multifaceted applications. Newly proposed neural estimators for these quantities have overcome severe drawbacks of classical $k$NN-based estimators in high dimensions. In this work, we focus on conditional mutual information (CMI) estimation by utilizing its formulation as a \textit{minmax} optimization problem. Such a formulation leads to a joint training procedure similar to that of generative adversarial networks. We find that our proposed estimator provides better estimates than the existing approaches on a variety of simulated datasets comprising linear and non-linear relations between variables. As an application of CMI estimation, we deploy our estimator for conditional independence (CI) testing on real data and obtain better results than state-of-the-art CI testers.

The field of neural generative models is dominated by the highly successful Generative Adversarial Networks (GANs) despite their challenges, such as training instability and mode collapse. Auto-Encoders (AE) with regularized latent space provide an alternative framework for generative models, albeit their performance levels have not reached that of GANs. In this work, we hypothesise that the dimensionality of the AE model’s latent space has a critical effect on the quality of generated data. Under the assumption that nature generates data by sampling from a “true" generative latent space followed by a deterministic function, we show that the optimal performance is obtained when the dimensionality of the latent space of the AE-model matches with that of the “true" generative latent space. Further, we propose an algorithm called the Mask Adversarial Auto-Encoder (MaskAAE), in which the dimensionality of the latent space of an adversarial auto encoder is brought closer to that of the “true" generative latent space, via a procedure to mask the spurious latent dimensions. We demonstrate through experiments on synthetic and several real-world datasets that the proposed formulation yields betterment in the generation quality.

This paper presents a new Multiobjective Evolutionary Algorithm (MOEA) based on decomposition, with fuzzy dominance (MOEA/DFD). The algorithm introduces a fuzzy Pareto dominance concept to compare two solutions and uses the scalar decomposition method only when one of the solutions fails to dominate the other in terms of a fuzzy dominance level. The diversity is maintained through the uniformly distributed weight vectors. In addition, Dynamic Resource Allocation (DRA) is used to distribute the computational effort based on the utilities of the individuals. To assess the performance of the proposed algorithm, experiments were conducted on two general benchmarks and ten unconstrained benchmark problems taken from the competition on real parameter MOEAs held under the 2009 IEEE Congress on Evolutionary Computation (CEC). As per the IGD metric, MOEA/DFD outperforms other major MOEAs in most cases.