Amit Ranjan Trivedi

Nastaran Darabi, Dinithi Jayasuriya, Devashri Naik et al.

Adversarial attacks exploit vulnerabilities in a model's decision boundaries through small, carefully crafted perturbations that lead to significant mispredictions. In 3D vision, the high dimensionality and sparsity of data greatly expand the attack surface, making 3D vision particularly vulnerable for safety-critical robotics. To enhance 3D vision's adversarial robustness, we propose a training objective that simultaneously minimizes prediction loss and mutual information (MI) under adversarial perturbations to contain the upper bound of misprediction errors. This approach simplifies handling adversarial examples compared to conventional methods, which require explicit searching and training on adversarial samples. However, minimizing prediction loss conflicts with minimizing MI, leading to reduced robustness and catastrophic forgetting. To address this, we integrate curriculum advisors in the training setup that gradually introduce adversarial objectives to balance training and prevent models from being overwhelmed by difficult cases early in the process. The advisors also enhance robustness by encouraging training on diverse MI examples through entropy regularizers. We evaluated our method on ModelNet40 and KITTI using PointNet, DGCNN, SECOND, and PointTransformers, achieving 2-5% accuracy gains on ModelNet40 and a 5-10% mAP improvement in object detection. Our code is publicly available at https://github.com/nstrndrbi/Mine-N-Learn.

Divake Kumar, Sina Tayebati, Devashri Naik et al.

Vision-language models (VLMs) are increasingly used as automated judges for multimodal systems, yet their scores provide no indication of reliability. We study this problem through conformal prediction, a distribution-free framework that converts a judge's point score into a calibrated prediction interval using only score-token log-probabilities, with no retraining. We present the first systematic analysis of conformal prediction for VLM-as-a-Judge across 3 judges and 14 visual task categories. Our results show that evaluation uncertainty is strongly task-dependent: intervals cover ~40% of the score range for aesthetics and natural images but expand to ~70% for chart and mathematical reasoning, yielding a quantitative reliability map for multimodal evaluation. We further identify a failure mode not captured by standard evaluation metrics, ranking-scoring decoupling, where judges achieve high ranking correlation while producing wide, uninformative intervals, correctly ordering responses but failing to assign reliable absolute scores. Finally, we show that interval width is driven primarily by task difficulty and annotation quality, i.e., the same judge and method yield 4.5x narrower intervals on a clean, multi-annotator captioning benchmark. Code: https://github.com/divake/VLM-Judge-Uncertainty

Davide Giacomini, Maeesha Binte Hashem, Jeremiah Suarez et al.

The rapid advancement of deep neural networks has significantly improved various tasks, such as image and speech recognition. However, as the complexity of these models increases, so does the computational cost and the number of parameters, making it difficult to deploy them on resource-constrained devices. This paper proposes a novel memorization-based inference (MBI) that is compute free and only requires lookups. Specifically, our work capitalizes on the inference mechanism of the recurrent attention model (RAM), where only a small window of input domain (glimpse) is processed in a one time step, and the outputs from multiple glimpses are combined through a hidden vector to determine the overall classification output of the problem. By leveraging the low-dimensionality of glimpse, our inference procedure stores key value pairs comprising of glimpse location, patch vector, etc. in a table. The computations are obviated during inference by utilizing the table to read out key-value pairs and performing compute-free inference by memorization. By exploiting Bayesian optimization and clustering, the necessary lookups are reduced, and accuracy is improved. We also present in-memory computing circuits to quickly look up the matching key vector to an input query. Compared to competitive compute-in-memory (CIM) approaches, MBI improves energy efficiency by almost 2.7 times than multilayer perceptions (MLP)-CIM and by almost 83 times than ResNet20-CIM for MNIST character recognition.

Alex C. Stutts, Danilo Erricolo, Theja Tulabandhula et al.

Data-driven visual odometry (VO) is a critical subroutine for autonomous edge robotics, and recent progress in the field has produced highly accurate point predictions in complex environments. However, emerging autonomous edge robotics devices like insect-scale drones and surgical robots lack a computationally efficient framework to estimate VO's predictive uncertainties. Meanwhile, as edge robotics continue to proliferate into mission-critical application spaces, awareness of model's the predictive uncertainties has become crucial for risk-aware decision-making. This paper addresses this challenge by presenting a novel, lightweight, and statistically robust framework that leverages conformal inference (CI) to extract VO's uncertainty bands. Our approach represents the uncertainties using flexible, adaptable, and adjustable prediction intervals that, on average, guarantee the inclusion of the ground truth across all degrees of freedom (DOF) of pose estimation. We discuss the architectures of generative deep neural networks for estimating multivariate uncertainty bands along with point (mean) prediction. We also present techniques to improve the uncertainty estimation accuracy, such as leveraging Monte Carlo dropout (MC-dropout) for data augmentation. Finally, we propose a novel training loss function that combines interval scoring and calibration loss with traditional training metrics--mean-squared error and KL-divergence--to improve uncertainty-aware learning. Our simulation results demonstrate that the presented framework consistently captures true uncertainty in pose estimations across different datasets, estimation models, and applied noise types, indicating its wide applicability.

Nethmi Jayasinghe, Diana Gontero, Spencer T. Brown et al.

Robotic policies deployed in real-world environments often encounter post-training faults, where retraining, exploration, or system identification are impractical. We introduce an inference-time, cerebellar-inspired residual control framework that augments a frozen reinforcement learning policy with online corrective actions, enabling fault recovery without modifying base policy parameters. The framework instantiates core cerebellar principles, including high-dimensional pattern separation via fixed feature expansion, parallel microzone-style residual pathways, and local error-driven plasticity with excitatory and inhibitory eligibility traces operating at distinct time scales. These mechanisms enable fast, localized correction under post-training disturbances while avoiding destabilizing global policy updates. A conservative, performance-driven meta-adaptation regulates residual authority and plasticity, preserving nominal behavior and suppressing unnecessary intervention. Experiments on MuJoCo benchmarks under actuator, dynamic, and environmental perturbations show improvements of up to $+66\%$ on \texttt{HalfCheetah-v5} and $+53\%$ on \texttt{Humanoid-v5} under moderate faults, with graceful degradation under severe shifts and complementary robustness from consolidating persistent residual corrections into policy parameters.

Alex C. Stutts, Danilo Erricolo, Sathya Ravi et al.

In the expanding landscape of AI-enabled robotics, robust quantification of predictive uncertainties is of great importance. Three-dimensional (3D) object detection, a critical robotics operation, has seen significant advancements; however, the majority of current works focus only on accuracy and ignore uncertainty quantification. Addressing this gap, our novel study integrates the principles of conformal inference (CI) with information theoretic measures to perform lightweight, Monte Carlo-free uncertainty estimation within a multimodal framework. Through a multivariate Gaussian product of the latent variables in a Variational Autoencoder (VAE), features from RGB camera and LiDAR sensor data are fused to improve the prediction accuracy. Normalized mutual information (NMI) is leveraged as a modulator for calibrating uncertainty bounds derived from CI based on a weighted loss function. Our simulation results show an inverse correlation between inherent predictive uncertainty and NMI throughout the model's training. The framework demonstrates comparable or better performance in KITTI 3D object detection benchmarks to similar methods that are not uncertainty-aware, making it suitable for real-time edge robotics.

Shamma Nasrin, Maeesha Binte Hashem, Nastaran Darabi et al.

This work discusses memory-immersed collaborative digitization among compute-in-memory (CiM) arrays to minimize the area overheads of a conventional analog-to-digital converter (ADC) for deep learning inference. Thereby, using the proposed scheme, significantly more CiM arrays can be accommodated within limited footprint designs to improve parallelism and minimize external memory accesses. Under the digitization scheme, CiM arrays exploit their parasitic bit lines to form a within-memory capacitive digital-to-analog converter (DAC) that facilitates area-efficient successive approximation (SA) digitization. CiM arrays collaborate where a proximal array digitizes the analog-domain product-sums when an array computes the scalar product of input and weights. We discuss various networking configurations among CiM arrays where Flash, SA, and their hybrid digitization steps can be efficiently implemented using the proposed memory-immersed scheme. The results are demonstrated using a 65 nm CMOS test chip. Compared to a 40 nm-node 5-bit SAR ADC, our 65 nm design requires $\sim$25$\times$ less area and $\sim$1.4$\times$ less energy by leveraging in-memory computing structures. Compared to a 40 nm-node 5-bit Flash ADC, our design requires $\sim$51$\times$ less area and $\sim$13$\times$ less energy.

Domenico Parente, Nastaran Darabi, Alex C. Stutts et al.

This paper introduces a lightweight uncertainty estimator capable of predicting multimodal (disjoint) uncertainty bounds by integrating conformal prediction with a deep-learning regressor. We specifically discuss its application for visual odometry (VO), where environmental features such as flying domain symmetries and sensor measurements under ambiguities and occlusion can result in multimodal uncertainties. Our simulation results show that uncertainty estimates in our framework adapt sample-wise against challenging operating conditions such as pronounced noise, limited training data, and limited parametric size of the prediction model. We also develop a reasoning framework that leverages these robust uncertainty estimates and incorporates optical flow-based reasoning to improve prediction prediction accuracy. Thus, by appropriately accounting for predictive uncertainties of data-driven learning and closing their estimation loop via rule-based reasoning, our methodology consistently surpasses conventional deep learning approaches on all these challenging scenarios--pronounced noise, limited training data, and limited model size-reducing the prediction error by 2-3x.

Nastaran Darabi, Maeesha Binte Hashem, Hongyi Pan et al.

The edge processing of deep neural networks (DNNs) is becoming increasingly important due to its ability to extract valuable information directly at the data source to minimize latency and energy consumption. Frequency-domain model compression, such as with the Walsh-Hadamard transform (WHT), has been identified as an efficient alternative. However, the benefits of frequency-domain processing are often offset by the increased multiply-accumulate (MAC) operations required. This paper proposes a novel approach to an energy-efficient acceleration of frequency-domain neural networks by utilizing analog-domain frequency-based tensor transformations. Our approach offers unique opportunities to enhance computational efficiency, resulting in several high-level advantages, including array micro-architecture with parallelism, ADC/DAC-free analog computations, and increased output sparsity. Our approach achieves more compact cells by eliminating the need for trainable parameters in the transformation matrix. Moreover, our novel array micro-architecture enables adaptive stitching of cells column-wise and row-wise, thereby facilitating perfect parallelism in computations. Additionally, our scheme enables ADC/DAC-free computations by training against highly quantized matrix-vector products, leveraging the parameter-free nature of matrix multiplications. Another crucial aspect of our design is its ability to handle signed-bit processing for frequency-based transformations. This leads to increased output sparsity and reduced digitization workload. On a 16$\times$16 crossbars, for 8-bit input processing, the proposed approach achieves the energy efficiency of 1602 tera operations per second per Watt (TOPS/W) without early termination strategy and 5311 TOPS/W with early termination strategy at VDD = 0.8 V.

Patrick Poggi, Divake Kumar, Theja Tulabandhula et al.

Transformer-based single-object trackers achieve state-of-the-art accuracy but rely on fixed-depth inference, executing the full encoder--decoder stack for every frame regardless of visual complexity, thereby incurring unnecessary computational cost in long video sequences dominated by temporally coherent frames. We propose UncL-STARK, an architecture-preserving approach that enables dynamic, uncertainty-aware depth adaptation in transformer-based trackers without modifying the underlying network or adding auxiliary heads. The model is fine-tuned to retain predictive robustness at multiple intermediate depths using random-depth training with knowledge distillation, thus enabling safe inference-time truncation. At runtime, we derive a lightweight uncertainty estimate directly from the model's corner localization heatmaps and use it in a feedback-driven policy that selects the encoder and decoder depth for the next frame based on the prediction confidence by exploiting temporal coherence in video. Extensive experiments on GOT-10k and LaSOT demonstrate up to 12\% GFLOPs reduction, 8.9\% latency reduction, and 10.8\% energy savings while maintaining tracking accuracy within 0.2\% of the full-depth baseline across both short-term and long-term sequences.

Sina Tayebati, Divake Kumar, Nastaran Darabi et al.

Large Language and Vision-Language Models (LLMs/VLMs) are increasingly used in safety-critical applications, yet their opaque decision-making complicates risk assessment and reliability. Uncertainty quantification (UQ) helps assess prediction confidence and enables abstention when uncertainty is high. Conformal prediction (CP), a leading UQ method, provides statistical guarantees but relies on static thresholds, which fail to adapt to task complexity and evolving data distributions, leading to suboptimal trade-offs in accuracy, coverage, and informativeness. To address this, we propose learnable conformal abstention, integrating reinforcement learning (RL) with CP to optimize abstention thresholds dynamically. By treating CP thresholds as adaptive actions, our approach balances multiple objectives, minimizing prediction set size while maintaining reliable coverage. Extensive evaluations across diverse LLM/VLM benchmarks show our method outperforms Least Ambiguous Classifiers (LAC) and Adaptive Prediction Sets (APS), improving accuracy by up to 3.2%, boosting AUROC for hallucination detection by 22.19%, enhancing uncertainty-guided selective generation (AUARC) by 21.17%, and reducing calibration error by 70%-85%. These improvements hold across multiple models and datasets while consistently meeting the 90% coverage target, establishing our approach as a more effective and flexible solution for reliable decision-making in safety-critical applications. The code is available at: {https://github.com/sinatayebati/vlm-uncertainty}.

Nastaran Darabi, Amit Ranjan Trivedi

Vision language action (VLA) models enable generalist robotic agents but often exhibit language ignorance, relying on visual shortcuts and remaining insensitive to instruction changes. We present Prospective Grounding and Alignment VLA (ProGAL-VLA), which constructs a 3D entity-centric graph (GSM), uses a slow planner to produce symbolic sub-goals, and aligns them with grounded entities via a Grounding Alignment Contrastive (GAC) loss. All actions are conditioned on a verified goal embedding $g_t$, whose attention entropy provides an intrinsic ambiguity signal. On LIBERO-Plus, ProGAL-VLA increases robustness under robot perturbations from 30.3 to 71.5 percent, reduces language ignorance by 3x-4x, and improves entity retrieval from 0.41 to 0.71 Recall@1. On the Custom Ambiguity Benchmark, it reaches AUROC 0.81 (vs. 0.52), AUPR 0.79, and raises clarification on ambiguous inputs from 0.09 to 0.81 without harming unambiguous success. The verification bottleneck increases mutual information of language-actions, the GAC loss imposes an entity-level InfoNCE bound, and attention entropy yields calibrated selective prediction, indicating that explicit verified grounding is an effective path toward instruction-sensitive, ambiguity-aware agents.

Devashri Naik, Divake Kumar, Nastaran Darabi et al.

Robots operating in shared workspaces must maintain safe coordination with other agents whose behavior may change during task execution. When a collaborating agent switches strategy mid-episode, continuing under outdated assumptions can lead to unsafe actions and increased collision risk. Reliable detection of such behavioral regime changes is therefore critical. We study regime-switch detection under controlled non-stationarity in ManiSkill shared-workspace manipulation tasks. Across ten detection methods and five random seeds, enabling detection reduces post-switch collisions by 52%. However, average performance hides significant reliability differences: under a realistic tolerance of +-3 steps, detection ranges from 86% to 30%, while under +-5 steps all methods achieve 100%. We introduce UA-TOM, a lightweight belief-tracking module that augments frozen vision-language-action (VLA) control backbones using selective state-space dynamics, causal attention, and prediction-error signals. Across five seeds and 1200 episodes, UA-TOM achieves the highest detection rate among unassisted methods (85.7% at +-3) and the lowest close-range time (4.8 steps), outperforming an Oracle (5.3 steps). Analysis shows hidden-state update magnitude increases by 17x at regime switches and decays over roughly 10 timesteps, while the discretization step converges to a near-constant value (Delta_t approx 0.78), indicating sensitivity driven by learned dynamics rather than input-dependent gating. Cross-domain experiments in Overcooked show complementary roles of causal attention and prediction-error signals. UA-TOM introduces 7.4 ms inference overhead (14.8% of a 50 ms control budget), enabling reliable regime-switch detection without modifying the base policy.

Sina Tayebati, Divake Kumar, Nastaran Darabi et al.

Estimating uncertainty for AI agents in real-world multi-turn tool-using interaction with humans is difficult because failures are often triggered by sparse critical episodes (e.g., looping, incoherent tool use, or user-agent miscoordination) even when local generation appears confident. Existing uncertainty proxies focus on single-shot text generation and therefore miss these trajectory-level breakdown signals. We introduce TRACER, a trajectory-level uncertainty metric for dual-control Tool-Agent-User interaction. TRACER combines content-aware surprisal with situational-awareness signals, semantic and lexical repetition, and tool-grounded coherence gaps, and aggregates them using a tail-focused risk functional with a MAX-composite step risk to surface decisive anomalies. We evaluate TRACER on $τ^2$-bench by predicting task failure and selective task execution. To this end, TRACER improves AUROC by up to 37.1% and AUARC by up to 55% over baselines, enabling earlier and more accurate detection of uncertainty in complex conversational tool-use settings. Our code and benchmark are available at https://github.com/sinatayebati/agent-tracer.

Divake Kumar, Patrick Poggi, Sina Tayebati et al.

Most estimators collapse all uncertainty modes into a single confidence score, preventing reliable reasoning about when to allocate more compute or adjust inference. We introduce Uncertainty-Guided Inference-Time Selection, a lightweight inference time framework that disentangles aleatoric (data-driven) and epistemic (model-driven) uncertainty directly in deep feature space. Aleatoric uncertainty is estimated using a regularized global density model, while epistemic uncertainty is formed from three complementary components that capture local support deficiency, manifold spectral collapse, and cross-layer feature inconsistency. These components are empirically orthogonal and require no sampling, no ensembling, and no additional forward passes. We integrate the decomposed uncertainty into a distribution free conformal calibration procedure that yields significantly tighter prediction intervals at matched coverage. Using these components for uncertainty guided adaptive model selection reduces compute by approximately 60 percent on MOT17 with negligible accuracy loss, enabling practical self regulating visual inference. Additionally, our ablation results show that the proposed orthogonal uncertainty decomposition consistently yields higher computational savings across all MOT17 sequences, improving margins by 13.6 percentage points over the total-uncertainty baseline.

Amit Ranjan Trivedi, Sina Tayebati, Hemant Kumawat et al.

Autonomous edge computing in robotics, smart cities, and autonomous vehicles relies on the seamless integration of sensing, processing, and actuation for real-time decision-making in dynamic environments. At its core is the sensing-to-action loop, which iteratively aligns sensor inputs with computational models to drive adaptive control strategies. These loops can adapt to hyper-local conditions, enhancing resource efficiency and responsiveness, but also face challenges such as resource constraints, synchronization delays in multi-modal data fusion, and the risk of cascading errors in feedback loops. This article explores how proactive, context-aware sensing-to-action and action-to-sensing adaptations can enhance efficiency by dynamically adjusting sensing and computation based on task demands, such as sensing a very limited part of the environment and predicting the rest. By guiding sensing through control actions, action-to-sensing pathways can improve task relevance and resource use, but they also require robust monitoring to prevent cascading errors and maintain reliability. Multi-agent sensing-action loops further extend these capabilities through coordinated sensing and actions across distributed agents, optimizing resource use via collaboration. Additionally, neuromorphic computing, inspired by biological systems, provides an efficient framework for spike-based, event-driven processing that conserves energy, reduces latency, and supports hierarchical control--making it ideal for multi-agent optimization. This article highlights the importance of end-to-end co-design strategies that align algorithmic models with hardware and environmental dynamics and improve cross-layer interdependencies to improve throughput, precision, and adaptability for energy-efficient edge autonomy in complex environments.

Dinithi Jayasuriya, Sina Tayebati, Davide Ettori et al.

We propose SPARC, a lightweight continual learning framework for large language models (LLMs) that enables efficient task adaptation through prompt tuning in a lower-dimensional space. By leveraging principal component analysis (PCA), we identify a compact subspace of the training data. Optimizing prompts in this lower-dimensional space enhances training efficiency, as it focuses updates on the most relevant features while reducing computational overhead. Furthermore, since the model's internal structure remains unaltered, the extensive knowledge gained from pretraining is fully preserved, ensuring that previously learned information is not compromised during adaptation. Our method achieves high knowledge retention in both task-incremental and domain-incremental continual learning setups while fine-tuning only 0.04% of the model's parameters. Additionally, by integrating LoRA, we enhance adaptability to computational constraints, allowing for a tradeoff between accuracy and training cost. Experiments on the SuperGLUE benchmark demonstrate that our PCA-based prompt tuning combined with LoRA maintains full knowledge retention while improving accuracy, utilizing only 1% of the model's parameters. These results establish our approach as a scalable and resource-efficient solution for continual learning in LLMs.

Nastaran Darabi, Devashri Naik, Sina Tayebati et al.

Vision-Language Models (VLMs) inherit adversarial vulnerabilities of Large Language Models (LLMs), which are further exacerbated by their multimodal nature. Existing defenses, including adversarial training, input transformations, and heuristic detection, are computationally expensive, architecture-dependent, and fragile against adaptive attacks. We introduce EigenShield, an inference-time defense leveraging Random Matrix Theory to quantify adversarial disruptions in high-dimensional VLM representations. Unlike prior methods that rely on empirical heuristics, EigenShield employs the spiked covariance model to detect structured spectral deviations. Using a Robustness-based Nonconformity Score (RbNS) and quantile-based thresholding, it separates causal eigenvectors, which encode semantic information, from correlational eigenvectors that are susceptible to adversarial artifacts. By projecting embeddings onto the causal subspace, EigenShield filters adversarial noise without modifying model parameters or requiring adversarial training. This architecture-independent, attack-agnostic approach significantly reduces the attack success rate, establishing spectral analysis as a principled alternative to conventional defenses. Our results demonstrate that EigenShield consistently outperforms all existing defenses, including adversarial training, UNIGUARD, and CIDER.

Divake Kumar, Nastaran Darabi, Sina Tayebati et al.

Safety-critical perception systems require both reliable uncertainty quantification and principled abstention mechanisms to maintain safety under diverse operational conditions. We present a novel dual-threshold conformalization framework that provides statistically-guaranteed uncertainty estimates while enabling selective prediction in high-risk scenarios. Our approach uniquely combines a conformal threshold ensuring valid prediction sets with an abstention threshold optimized through ROC analysis, providing distribution-free coverage guarantees (>= 1 - alpha) while identifying unreliable predictions. Through comprehensive evaluation on CIFAR-100, ImageNet1K, and ModelNet40 datasets, we demonstrate superior robustness across camera and LiDAR modalities under varying environmental perturbations. The framework achieves exceptional detection performance (AUC: 0.993 to 0.995) under severe conditions while maintaining high coverage (>90.0%) and enabling adaptive abstention (13.5% to 63.4% +/- 0.5) as environmental severity increases. For LiDAR-based perception, our approach demonstrates particularly strong performance, maintaining robust coverage (>84.5%) while appropriately abstaining from unreliable predictions. Notably, the framework shows remarkable stability under heavy perturbations, with detection performance (AUC: 0.995 +/- 0.001) significantly outperforming existing methods across all modalities. Our unified approach bridges the gap between theoretical guarantees and practical deployment needs, offering a robust solution for safety-critical autonomous systems operating in challenging real-world conditions.

Nastaran Darabi, Divake Kumar, Sina Tayebati et al.

In this work, we present INTACT, a novel two-phase framework designed to enhance the robustness of deep neural networks (DNNs) against noisy LiDAR data in safety-critical perception tasks. INTACT combines meta-learning with adversarial curriculum training (ACT) to systematically address challenges posed by data corruption and sparsity in 3D point clouds. The meta-learning phase equips a teacher network with task-agnostic priors, enabling it to generate robust saliency maps that identify critical data regions. The ACT phase leverages these saliency maps to progressively expose a student network to increasingly complex noise patterns, ensuring targeted perturbation and improved noise resilience. INTACT's effectiveness is demonstrated through comprehensive evaluations on object detection, tracking, and classification benchmarks using diverse datasets, including KITTI, Argoverse, and ModelNet40. Results indicate that INTACT improves model robustness by up to 20% across all tasks, outperforming standard adversarial and curriculum training methods. This framework not only addresses the limitations of conventional training strategies but also offers a scalable and efficient solution for real-world deployment in resource-constrained safety-critical systems. INTACT's principled integration of meta-learning and adversarial training establishes a new paradigm for noise-tolerant 3D perception in safety-critical applications. INTACT improved KITTI Multiple Object Tracking Accuracy (MOTA) by 9.6% (64.1% -> 75.1%) and by 12.4% under Gaussian noise (52.5% -> 73.7%). Similarly, KITTI mean Average Precision (mAP) rose from 59.8% to 69.8% (50% point drop) and 49.3% to 70.9% (Gaussian noise), highlighting the framework's ability to enhance deep learning model resilience in safety-critical object tracking scenarios.

Dinithi Jayasuriya, Nastaran Darabi, Maeesha Binte Hashem et al.

We introduce a precision polarization scheme for DNN inference that utilizes only very low and very high precision levels, assigning low precision to the majority of network weights and activations while reserving high precision paths for targeted error compensation. This separation allows for distinct optimization of each precision level, thereby reducing memory and computation demands without compromising model accuracy. In the discussed approach, a floating-point model can be trained in the cloud and then downloaded to an edge device, where network weights and activations are directly quantized to meet the edge devices' desired level, such as NF4 or INT8. To address accuracy loss from quantization, surrogate paths are introduced, leveraging low-rank approximations on a layer-by-layer basis. These paths are trained with a sensitivity-based metric on minimal training data to recover accuracy loss under quantization as well as due to process variability, such as when the main prediction path is implemented using analog acceleration. Our simulation results show that neural precision polarization enables approximately 464 TOPS per Watt MAC efficiency and reliability by integrating rank-8 error recovery paths with highly efficient, though potentially unreliable, bit plane-wise compute-in-memory processing.

Divake Kumar, Sina Tayebati, Francesco Migliarba et al.

Deep learning models in robotics often output point estimates with poorly calibrated confidences, offering no native mechanism to quantify predictive reliability under novel, noisy, or out-of-distribution inputs. Conformal prediction (CP) addresses this gap by providing distribution-free coverage guarantees, yet its reliance on fixed nonconformity scores ignores context and can yield intervals that are overly conservative or unsafe. We address this with Learnable Conformal Prediction (LCP), which replaces fixed scores with a lightweight neural function that leverages geometric, semantic, and task-specific features to produce context-aware uncertainty sets. LCP maintains CP's theoretical guarantees while reducing prediction set sizes by 18% in classification, tightening detection intervals by 52%, and improving path planning safety from 72% to 91% success with minimal overhead. Across three robotic tasks on seven benchmarks, LCP consistently outperforms Standard CP and ensemble baselines. In classification on CIFAR-100 and ImageNet, it achieves smaller set sizes (4.7-9.9% reduction) at target coverage. For object detection on COCO, BDD100K, and Cityscapes, it produces 46-54% tighter bounding boxes. In path planning through cluttered environments, it improves success to 91.5% with only 4.5% path inflation, compared to 12.2% for Standard CP. The method is lightweight (approximately 4.8% runtime overhead, 42 KB memory) and supports online adaptation, making it well suited to resource-constrained autonomous systems. Hardware evaluation shows LCP adds less than 1% memory and 15.9% inference overhead, yet sustains 39 FPS on detection tasks while being 7.4 times more energy-efficient than ensembles.

Davide Ettori, Nastaran Darabi, Sina Tayebati et al.

Large language models (LLMs) offer broad utility but remain prone to hallucination and out-of-distribution (OOD) errors. We propose EigenTrack, an interpretable real-time detector that uses the spectral geometry of hidden activations, a compact global signature of model dynamics. By streaming covariance-spectrum statistics such as entropy, eigenvalue gaps, and KL divergence from random baselines into a lightweight recurrent classifier, EigenTrack tracks temporal shifts in representation structure that signal hallucination and OOD drift before surface errors appear. Unlike black- and grey-box methods, it needs only a single forward pass without resampling. Unlike existing white-box detectors, it preserves temporal context, aggregates global signals, and offers interpretable accuracy-latency trade-offs.

Davide Ettori, Nastaran Darabi, Sureshkumar Senthilkumar et al.

Large deep learning models such as BERT and ResNet achieve state-of-the-art performance but are costly to deploy at the edge due to their size and compute demands. We present RMT-KD, a compression method that leverages Random Matrix Theory (RMT) for knowledge distillation to iteratively reduce network size. Instead of pruning or heuristic rank selection, RMT-KD preserves only informative directions identified via the spectral properties of hidden representations. RMT-based causal reduction is applied layer by layer with self-distillation to maintain stability and accuracy. On GLUE, AG News, and CIFAR-10, RMT-KD achieves up to 80% parameter reduction with only 2% accuracy loss, delivering 2.8x faster inference and nearly halved power consumption. These results establish RMT-KD as a mathematically grounded approach to network distillation.

Divake Kumar, Sina Tayebati, Devashri Naik et al.

Most uncertainty-aware robotic systems collapse prediction uncertainty into a single scalar score and use it to trigger uniform corrective responses. This aggregation obscures whether uncertainty arises from corrupted observations or from mismatch between the learned model and the true system dynamics. As a result, corrective actions may be applied to the wrong component of the closed loop, degrading performance relative to leaving the policy unchanged. We introduce a lightweight post hoc framework that decomposes uncertainty into aleatoric and epistemic components and uses these signals to regulate system responses at inference time. Aleatoric uncertainty is estimated from deviations in the observation distribution using a Mahalanobis density model, while epistemic uncertainty is detected using a noise robust forward dynamics ensemble that isolates model mismatch from measurement corruption. The two signals remain empirically near orthogonal during closed loop execution and enable type specific responses. High aleatoric uncertainty triggers observation recovery, while high epistemic uncertainty moderates control actions. The same signals also regulate adaptive perception by guiding model capacity selection during tracking inference. Experiments demonstrate consistent improvements across both control and perception tasks. In robotic manipulation, the decomposed controller improves task success from 59.4% to 80.4% under compound perturbations and outperforms a combined uncertainty baseline by up to 21.0%. In adaptive tracking inference on MOT17, uncertainty-guided model selection reduces average compute by 58.2% relative to a fixed high capacity detector while preserving detection quality within 0.4%. Code and demo videos are available at https://divake.github.io/uncertainty-decomposition/.

Nethmi Jayasinghe, Diana Gontero, Francesco Migliarba et al.

Robotic systems operating in real-world environments inevitably encounter unobserved dynamics shifts during continuous execution, including changes in actuation, mass distribution, or contact conditions. When such shifts occur mid-episode, even locally stabilizing learned policies can experience substantial transient performance degradation. While input-to-state stability guarantees bounded state deviation, it does not ensure rapid restoration of task-level performance. We address inference-time recovery under frozen policy parameters by casting adaptation as constrained disturbance shaping around a nominal stabilizing controller. We propose a stability-aligned residual control architecture in which a reinforcement learning policy trained under nominal dynamics remains fixed at deployment, and adaptation occurs exclusively through a bounded additive residual channel. A Stability Alignment Gate (SAG) regulates corrective authority through magnitude constraints, directional coherence with the nominal action, performance-conditioned activation, and adaptive gain modulation. These mechanisms preserve the nominal closed-loop structure while enabling rapid compensation for unobserved dynamics shifts without retraining or privileged disturbance information. Across mid-episode perturbations including actuator degradation, mass variation, and contact changes, the proposed method consistently reduces recovery time relative to frozen and online-adaptation baselines while maintaining near-nominal steady-state performance. Recovery time is reduced by \textbf{87\%} on the Go1 quadruped, \textbf{48\%} on the Cassie biped, \textbf{30\%} on the H1 humanoid, and \textbf{20\%} on the Scout wheeled platform on average across evaluated conditions relative to a frozen SAC policy.

Alex Christopher Stutts, Danilo Erricolo, Theja Tulabandhula et al.

We present a novel statistical approach to incorporating uncertainty awareness in model-free distributional reinforcement learning involving quantile regression-based deep Q networks. The proposed algorithm, $\textit{Calibrated Evidential Quantile Regression in Deep Q Networks (CEQR-DQN)}$, aims to address key challenges associated with separately estimating aleatoric and epistemic uncertainty in stochastic environments. It combines deep evidential learning with quantile calibration based on principles of conformal inference to provide explicit, sample-free computations of $\textit{global}$ uncertainty as opposed to $\textit{local}$ estimates based on simple variance, overcoming limitations of traditional methods in computational and statistical efficiency and handling of out-of-distribution (OOD) observations. Tested on a suite of miniaturized Atari games (i.e., MinAtar), CEQR-DQN is shown to surpass similar existing frameworks in scores and learning speed. Its ability to rigorously evaluate uncertainty improves exploration strategies and can serve as a blueprint for other algorithms requiring uncertainty awareness.

Priyesh Shukla, Shamma Nasrin, Nastaran Darabi et al.

We propose MC-CIM, a compute-in-memory (CIM) framework for robust, yet low power, Bayesian edge intelligence. Deep neural networks (DNN) with deterministic weights cannot express their prediction uncertainties, thereby pose critical risks for applications where the consequences of mispredictions are fatal such as surgical robotics. To address this limitation, Bayesian inference of a DNN has gained attention. Using Bayesian inference, not only the prediction itself, but the prediction confidence can also be extracted for planning risk-aware actions. However, Bayesian inference of a DNN is computationally expensive, ill-suited for real-time and/or edge deployment. An approximation to Bayesian DNN using Monte Carlo Dropout (MC-Dropout) has shown high robustness along with low computational complexity. Enhancing the computational efficiency of the method, we discuss a novel CIM module that can perform in-memory probabilistic dropout in addition to in-memory weight-input scalar product to support the method. We also propose a compute-reuse reformulation of MC-Dropout where each successive instance can utilize the product-sum computations from the previous iteration. Even more, we discuss how the random instances can be optimally ordered to minimize the overall MC-Dropout workload by exploiting combinatorial optimization methods. Application of the proposed CIM-based MC-Dropout execution is discussed for MNIST character recognition and visual odometry (VO) of autonomous drones. The framework reliably gives prediction confidence amidst non-idealities imposed by MC-CIM to a good extent. Proposed MC-CIM with 16x31 SRAM array, 0.85 V supply, 16nm low-standby power (LSTP) technology consumes 27.8 pJ for 30 MC-Dropout instances of probabilistic inference in its most optimal computing and peripheral configuration, saving 43% energy compared to typical execution.

Shamma Nasrin, Ahish Shylendra, Yuti Kadakia et al.

This work proposes a novel Energy-Aware Network Operator Search (ENOS) approach to address the energy-accuracy trade-offs of a deep neural network (DNN) accelerator. In recent years, novel inference operators have been proposed to improve the computational efficiency of a DNN. Augmenting the operators, their corresponding novel computing modes have also been explored. However, simplification of DNN operators invariably comes at the cost of lower accuracy, especially on complex processing tasks. Our proposed ENOS framework allows an optimal layer-wise integration of inference operators and computing modes to achieve the desired balance of energy and accuracy. The search in ENOS is formulated as a continuous optimization problem, solvable using typical gradient descent methods, thereby scalable to larger DNNs with minimal increase in training cost. We characterize ENOS under two settings. In the first setting, for digital accelerators, we discuss ENOS on multiply-accumulate (MAC) cores that can be reconfigured to different operators. ENOS training methods with single and bi-level optimization objectives are discussed and compared. We also discuss a sequential operator assignment strategy in ENOS that only learns the assignment for one layer in one training step, enabling greater flexibility in converging towards the optimal operator allocations. Furthermore, following Bayesian principles, a sampling-based variational mode of ENOS is also presented. ENOS is characterized on popular DNNs ShuffleNet and SqueezeNet on CIFAR10 and CIFAR100.

Priyesh Shukla, Ankith Muralidhar, Nick Iliev et al.

We propose a novel compute-in-memory (CIM)-based ultra-low-power framework for probabilistic localization of insect-scale drones. The conventional probabilistic localization approaches rely on the three-dimensional (3D) Gaussian Mixture Model (GMM)-based representation of a 3D map. A GMM model with hundreds of mixture functions is typically needed to adequately learn and represent the intricacies of the map. Meanwhile, localization using complex GMM map models is computationally intensive. Since insect-scale drones operate under extremely limited area/power budget, continuous localization using GMM models entails much higher operating energy -- thereby, limiting flying duration and/or size of the drone due to a larger battery. Addressing the computational challenges of localization in an insect-scale drone using a CIM approach, we propose a novel framework of 3D map representation using a harmonic mean of "Gaussian-like" mixture (HMGM) model. The likelihood function useful for drone localization can be efficiently implemented by connecting many multi-input inverters in parallel, each programmed with the parameters of the 3D map model represented as HMGM. When the depth measurements are projected to the input of the implementation, the summed current of the inverters emulates the likelihood of the measurement. We have characterized our approach on an RGB-D indoor localization dataset. The average localization error in our approach is $\sim$0.1125 m which is only slightly degraded than software-based evaluation ($\sim$0.08 m). Meanwhile, our localization framework is ultra-low-power, consuming as little as $\sim$17 $μ$W power while processing a depth frame in 1.33 ms over hundred pose hypotheses in the particle-filtering (PF) algorithm used to localize the drone.

Nick Iliev, Amit Ranjan Trivedi

We present a novel low latency CMOS hardware accelerator for fully connected (FC) layers in deep neural networks (DNNs). The FC accelerator, FC-ACCL, is based on 128 8x8 or 16x16 processing elements (PEs) for matrix-vector multiplication, and 128 multiply-accumulate (MAC) units integrated with 128 High Bandwidth Memory (HBM) units for storing the pretrained weights. Micro-architectural details for CMOS ASIC implementations are presented and simulated performance is compared to recent hardware accelerators for DNNs for AlexNet and VGG 16. When comparing simulated processing latency for a 4096-1000 FC8 layer, our FC-ACCL is able to achieve 48.4 GOPS (with a 100 MHz clock) which improves on a recent FC8 layer accelerator quoted at 28.8 GOPS with a 150 MHz clock. We have achieved this considerable improvement by fully utilizing the HBM units for storing and reading out column-specific FClayer weights in 1 cycle with a novel colum-row-column schedule, and implementing a maximally parallel datapath for processing these weights with the corresponding MAC and PE units. When up-scaled to 128 16x16 PEs, for 16x16 tiles of weights, the design can reduce latency for the large FC6 layer by 60 % in AlexNet and by 3 % in VGG16 when compared to an alternative EIE solution which uses compression.

Priyesh Shukla, Ahish Shylendra, Theja Tulabandhula et al.

This work discusses the implementation of Markov Chain Monte Carlo (MCMC) sampling from an arbitrary Gaussian mixture model (GMM) within SRAM. We show a novel architecture of SRAM by embedding it with random number generators (RNGs), digital-to-analog converters (DACs), and analog-to-digital converters (ADCs) so that SRAM arrays can be used for high performance Metropolis-Hastings (MH) algorithm-based MCMC sampling. Most of the expensive computations are performed within the SRAM and can be parallelized for high speed sampling. Our iterative compute flow minimizes data movement during sampling. We characterize power-performance trade-off of our design by simulating on 45 nm CMOS technology. For a two-dimensional, two mixture GMM, the implementation consumes ~ 91 micro-Watts power per sampling iteration and produces 500 samples in 2000 clock cycles on an average at 1 GHz clock frequency. Our study highlights interesting insights on how low-level hardware non-idealities can affect high-level sampling characteristics, and recommends ways to optimally operate SRAM within area/power constraints for high performance sampling.

Shamma Nasrin, Srikanth Ramakrishna, Theja Tulabandhula et al.

In this work, we introduce bitcell array-based support parameters to improve the prediction accuracy of SRAM-based binarized neural network (SRAM-BNN). Our approach enhances the training weight space of SRAM-BNN while requiring minimal overheads to a typical design. More flexibility of the weight space leads to higher prediction accuracy in our design. We adapt row digital-to-analog (DAC) converter, and computing flow in SRAM-BNN for bitcell array-based weight supports. Using the discussed interventions, our scheme also allows a dynamic trade-off of accuracy against latency to address dynamic latency constraints in typical real-time applications. We specifically discuss results on two training cases: (i) learning of support parameters on a pre-trained BNN and (ii) simultaneous learning of supports and weight binarization. In the former case, our approach reduces classification error in MNIST by 35.71% (error rate decreases from 1.4% to 0.91%). In the latter case, the error is reduced by 27.65% (error rate decreases from 1.4% to 1.13%). To reduce the power overheads, we propose a dynamic drop out a part of the support parameters. Our architecture can drop out 52% of the bitcell array-based support parameters without losing accuracy. We also characterize our design under varying degrees of process variability in the transistors.