Xueqing Peng, Lingfei Qian, Yan Wang, Ruoyu Xiang, Yueru He, Yang Ren, Mingyang Jiang, Vincent Jim Zhang, Yuqing Guo, Jeff Zhao, Huan He, Yi Han, Yun Feng, Yuechen Jiang, Yupeng Cao, Haohang Li, Yangyang Yu, Xiaoyu Wang, Penglei Gao, Shengyuan Lin, Keyi Wang, Shanshan Yang, Yilun Zhao, Zhiwei Liu, Peng Lu, Jerry Huang, Suyuchen Wang, Triantafillos Papadopoulos, Polydoros Giannouris, Efstathia Soufleri, Nuo Chen, Zhiyang Deng, Heming Fu, Yijia Zhao, Mingquan Lin, Meikang Qiu, Kaleb E Smith, Arman Cohan, Xiao-Yang Liu, Jimin Huang, Guojun Xiong, Alejandro Lopez-Lira, Xi Chen, Junichi Tsujii, Jian-Yun Nie, Sophia Ananiadou, Qianqian Xie. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Training robust deep video representations has proven to be computationally challenging due to substantial decoding overheads, the enormous size of raw video streams, and their inherent high temporal redundancy. Different from existing schemes, operating exclusively in the compressed video domain and exploiting all freely available modalities, i.e., I-frames, and P-frames (motion vectors and residuals) offers a compute-efficient alternative. Existing methods approach this task as a naive multi-modality problem, ignoring the temporal correlation and implicit sparsity across P-frames for modeling stronger shared representations for videos of the same action, making training and generalization easier. By revisiting the high-level design of dominant video understanding backbones, we increase inference speed by a factor of 56 while retaining similar performance. For this, we propose a hybrid end-to-end framework that factorizes learning across three key concepts to reduce inference cost by 113× versus prior art: First, a specially designed dual-encoder scheme with efficient Spiking Temporal Modulators to minimize latency while retaining cross-domain feature aggregation. Second, a unified transformer model to capture inter-modal dependencies using global self-attention to enhance I-frame – P-frame contextual interactions. Third, a Multi-Modal Mixer Block to model rich hierarchical representations from the joint spatiotemporal token embeddings. Experiments show that our method results in a lightweight architecture achieving state-of-the-art video recognition performance on the popular UCF-101, HMDB-51, K-400, K-600 and SS-v2 benchmarks with favorable costs (0.73J/V) and fast inference (16V/s). Further, our observations bring new insights into practical design choices for efficient next-generation spatiotemporal learners. Code will be released upon publication.

Xueqing Peng, Triantafillos Papadopoulos, Efstathia Soufleri, Polydoros Giannouris, Ruoyu Xiang, Yan Wang, Lingfei Qian, Jimin Huang, Qianqian Xie, Sophia Ananiadou. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Deep learning has become the de facto approach in nearly all learning tasks. It has been observed that deep models tend to memorize and sometimes overfit data, which can lead to compromises in performance, privacy, and other critical metrics. In this paper, we explore the theoretical foundations that connect memorization to sample loss, focusing on learning dynamics to understand what and how deep models memorize. To this end, we introduce a novel proxy for memorization: Cumulative Sample Loss (CSL). CSL represents the accumulated loss of a sample throughout the training process. CSL exhibits remarkable similarity to stability-based memorization, as evidenced by considerably high cosine similarity scores. We delve into the theory behind these results, demonstrating that low CSL leads to nontrivial bounds on the extent of stability-based memorization and learning time. The proposed proxy, CSL, is four orders of magnitude less computationally expensive than the stability-based method and can be obtained with zero additional overhead during training. We demonstrate the practical utility of the proposed proxy in identifying mislabeled samples and detecting duplicates where our metric achieves state-of-the-art performance.

Deep Neural Nets (DNNs) have become a pervasive tool for solving many emerging problems. However, they tend to overfit to and memorize the training set. Memorization is of keen interest since it is closely related to several concepts such as generalization, noisy learning, and privacy. To study memorization, Feldman (2019) proposed a formal score, however its computational requirements limit its practical use. Recent research has shown empirical evidence linking input loss curvature (measured by the trace of the loss Hessian w.r.t inputs) and memorization. It was shown to be $\sim3$ orders of magnitude more efficient than calculating the memorization score. However, there is a lack of theoretical understanding linking memorization with input loss curvature. In this paper, we not only investigate this connection but also extend our analysis to establish theoretical links between differential privacy, memorization, and input loss curvature. First, we derive an upper bound on memorization characterized by both differential privacy and input loss curvature. Secondly, we present a novel insight showing that input loss curvature is upper-bounded by the differential privacy parameter. Our theoretical findings are further validated using deep models on CIFAR and ImageNet datasets, showing a strong correlation between our theoretical predictions and results observed in practice.

In this paper, we explore the properties of loss curvature with respect to input data in deep neural networks. Curvature of loss with respect to input (termed input loss curvature) is the trace of the Hessian of the loss with respect to the input. We investigate how input loss curvature varies between train and test sets, and its implications for train-test distinguishability. We develop a theoretical framework that derives an upper bound on the train-test distinguishability based on privacy and the size of the training set. This novel insight fuels the development of a new black box membership inference attack utilizing input loss curvature. We validate our theoretical findings through experiments in computer vision classification tasks, demonstrating that input loss curvature surpasses existing methods in membership inference effectiveness. Our analysis highlights how the performance of membership inference attack (MIA) methods varies with the size of the training set, showing that curvature-based MIA outperforms other methods on sufficiently large datasets. This condition is often met by real datasets, as demonstrated by our results on CIFAR10, CIFAR100, and ImageNet. These findings not only advance our understanding of deep neural network behavior but also improve the ability to test privacy-preserving techniques in machine learning.

Memristive crossbars based on Non-volatile Memory (NVM) technologies such as RRAM, have recently shown great promise for accelerating Deep Neural Networks (DNNs). They achieve this by performing efficient Matrix-Vector-Multiplications (MVMs) while offering dense on-chip storage and minimal off-chip data movement. However, their analog nature of computing introduces functional errors due to non-ideal RRAM devices, significantly degrading the application accuracy. Further, RRAMs suffer from low endurance and high write costs, hindering on-chip trainability. To alleviate these limitations, we propose HyperX, a hybrid RRAM-SRAM system that leverages the complementary benefits of NVM and CMOS technologies. Our proposed system consists of a fixed RRAM block offering area and energy-efficient MVMs and an SRAM block enabling on-chip training to recover the accuracy drop due to the RRAM non-idealities. The improvements are reported in terms of energy and product of latency and area${\left(ms\,\times \,mm^{2}\right)}$, termed as area-normalized latency. Our experiments on CIFAR datasets using ResNet-20 show up to 2.88 × and 10.1 × improvements in inference energy and area-normalized latency, respectively. In addition, for a transfer learning task from ImageNet to CIFAR datasets using ResNet-18, we observe up to 1.58 × and 4.48 × improvements in energy and area-normalized latency, respectively. These improvements are with respect to an all-SRAM baseline.

We analyze the stability of recurrent networks, specifically, reservoir computing models during training by evaluating the eigenvalue spectra of the reservoir dynamics. To circumvent the instability arising in examining a closed loop reservoir system with feedback, we propose to break the closed loop system. Essentially, we unroll the reservoir dynamics over time while incorporating the feedback effects that preserve the overall temporal integrity of the system. We evaluate our methodology for fixed point and time varying targets with least squares regression and FORCE training [6], respectively. Our analysis establishes eigenvalue spectra (which is, shrinking of spectral circle as training progresses) as a valid and effective metric to gauge the convergence of training as well as the convergence of the chaotic activity of the reservoir toward stable states.