Jiajun Zhou, Yifan Yang, Kai Zhen, Ziyue Liu, Yequan Zhao, Ershad Banijamali, Athanasios Mouchtaris, Ngai Wong, Zheng Zhang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Fine-tuning large language models (LLMs) has achieved remarkable performance across various natural language processing tasks, yet it demands more and more memory as model sizes keep growing.To address this issue, the recently proposed Memory-efficient Zerothorder (MeZO) methods attempt to fine-tune LLMs using only forward passes, thereby avoiding the need for a backpropagation graph.However, significant performance drops and a high risk of divergence have limited their widespread adoption.In this paper, we propose the Adaptive Zeroth-order Tensor-Train Adaption (AdaZeta) framework, specifically designed to improve the performance and convergence of the ZO methods.To enhance dimension-dependent ZO estimation accuracy, we introduce a fast-forward, low-parameter tensorized adapter.To tackle the frequently observed divergence issue in large-scale ZO finetuning tasks, we propose an adaptive query number schedule that guarantees convergence.Detailed theoretical analysis and extensive experimental results on Roberta-Large and Llama-2-7B models substantiate the efficacy of our AdaZeta framework in terms of accuracy, memory efficiency, and convergence speed. 1

Transformer-based models demonstrate state of the art results on several natural language understanding tasks. However, their deployment comes at the cost of increased footprint and inference latency, limiting their adoption to real-time applications. Early exit strategies are designed to speed-up the inference by routing out a subset of samples at the earlier layers of the model. Exiting early causes losing model accuracy. In order to optimize the trade-off between model accuracy and latency, we propose Pyramid Dynamic Inference (PDI), a scheme that encourages fast inference via boosting the performance of early exit heads. PDI allows for more confident early inference by injecting stronger classifiers at earlier layers. It also prevents a significant increase in the model footprint by gradually shrinking the classifiers as the semantic capacity of the deeper transformer layers increase. Experiment results show that PDI outperforms the baselines on both accuracy and latency on the GLUE benchmark.

While supervised learning is widely used for perception modules in conventional autonomous driving solutions, scalability is hindered by the huge amount of data labeling needed. In contrast, while end-to-end architectures do not require labeled data and are potentially more scalable, interpretability is sacrificed. We introduce a novel architecture that is trained in a fully self-supervised fashion for simultaneous multi-step prediction of space-time cost map and road dynamics. Our solution replaces the manually designed cost function for motion planning with a learned high dimensional cost map that is naturally interpretable and allows diverse contextual information to be integrated without manual data labeling. Experiments on real world driving data show that our solution leads to lower number of collisions and road violations in long planning horizons in comparison to baselines, demonstrating the feasibility of fully self-supervised prediction without sacrificing scalability.

In autonomous driving (AD), accurately predicting changes in the environment can effectively improve safety and comfort. Due to complex interactions among traffic participants, however, it is very hard to achieve accurate prediction for a long horizon. To address this challenge, we propose prediction by anticipation, which views interaction in terms of a latent probabilistic generative process wherein some vehicles move partly in response to the anticipated motion of other vehicles. Under this view, consecutive data frames can be factorized into sequential samples from an action-conditional distribution that effectively generalizes to a wider range of actions and driving situations. Our proposed prediction model, variational Bayesian in nature, is trained to maximize the evidence lower bound (ELBO) of the log-likelihood of this conditional distribution. Evaluations of our approach with prominent AD datasets NGSIM I-80 and Argoverse show significant improvement over current state-of-the-art in both accuracy and generalization.

We address the problem of finding an optimal policy in a Markov decision process (MDP) under a restricted policy class defined by the convex hull of a set of base policies. This problem is of great interest in applications in which a number of reasonably good (or safe) policies are already known and we are interested in optimizing in their convex hull. We first prove that solving this problem is NP-hard. We then propose an efficient algorithm that finds a policy whose performance is almost as good as that of the best convex combination of the base policies, under the assumption that the occupancy measures of the base policies have a large overlap. The running time of the proposed algorithm is linear in the number of states and polynomial in the number of base policies. A distinct advantage of the proposed algorithm is that, apart from the computation of the occupancy measures of the base policies, it does not need to interact with the environment during the optimization process. This is especially important (i) in problems that due to concerns such as safety, we are restricted in interacting with the environment only through the (safe) base policies, and (ii) in complex systems where estimating the value of a policy can be a time consuming process.

Embed-to-control (E2C) is a model for solving high-dimensional optimal control problems by combining variational auto-encoders with locally-optimal controllers. However, the E2C model suffers from two major drawbacks: 1) its objective function does not correspond to the likelihood of the data sequence and 2) the variational encoder used for embedding typically has large variational approximation error, especially when there is noise in the system dynamics. In this paper, we present a new model for learning robust locally-linear controllable embedding (RCE). Our model directly estimates the predictive conditional density of the future observation given the current one, while introducing the bottleneck between the current and future observations. Although the bottleneck provides a natural embedding candidate for control, our RCE model introduces additional specific structures in the generative graphical model so that the model dynamics can be robustly linearized. We also propose a principled variational approximation of the embedding posterior that takes the future observation into account, and thus, makes the variational approximation more robust against the noise. Experimental results show that RCE outperforms the E2C model, and does so significantly when the underlying dynamics is noisy.

A generative model based on training deep architectures is proposed. The model consists of K networks that are trained together to learn the underlying distribution of a given data set. The process starts with dividing the input data into K clusters and feeding each of them into a separate network. After few iterations of training networks separately, we use an EM-like algorithm to train the networks together and update the clusters of the data. We call this model Mixture of Networks. The provided model is a platform that can be used for any deep structure and be trained by any conventional objective function for distribution modeling. As the components of the model are neural networks, it has high capability in characterizing complicated data distributions as well as clustering data. We apply the algorithm on MNIST hand-written digits and Yale face datasets. We also demonstrate the clustering ability of the model using some real-world and toy examples.

In recent years, advances in medical imaging have led to the emergence of massive databases, containing images from a diverse range of modalities. This has significantly heightened the need for automated annotation of the images on one side, and fast and memory-efficient content-based image retrieval systems on the other side. Binary descriptors have recently gained more attention as a potential vehicle to achieve these goals. One of the recently introduced binary descriptors for tagging of medical images are Radon barcodes (RBCs) that are driven from Radon transform via local thresholding. Gabor transform is also a powerful transform to extract texture-based information. Gabor features have exhibited robustness against rotation, scale, and also photometric disturbances, such as illumination changes and image noise in many applications. This paper introduces Gabor Barcodes (GBCs), as a novel framework for the image annotation. To find the most discriminative GBC for a given query image, the effects of employing Gabor filters with different parameters, i.e., different sets of scales and orientations, are investigated, resulting in different barcode lengths and retrieval performances. The proposed method has been evaluated on the IRMA dataset with 193 classes comprising of 12,677 x-ray images for indexing, and 1,733 x-rays images for testing. A total error score as low as 351 (≈ 80% accuracy for the first hit) was achieved.

In recent years, with the explosion of digital images on the Web, content-based retrieval has emerged as a significant research area. Shapes, textures, edges and segments may play a key role in describing the content of an image. Radon and Gabor transforms are both powerful techniques that have been widely studied to extract shape-texture-based information. The combined Radon-Gabor features may be more robust against scale/rotation variations, presence of noise, and illumination changes. The objective of this paper is to harness the potentials of both Gabor and Radon transforms in order to introduce expressive binary features, called barcodes, for image annotation/tagging tasks. We propose two different techniques: Gabor-of-Radon-Image Barcodes (GRIBCs), and Guided-Radon-of-Gabor Barcodes (GRGBCs). For validation, we employ the IRMA x-ray dataset with 193 classes, containing 12,677 training images and 1,733 test images. A total error score as low as 322 and 330 were achieved for GRGBCs and GRIBCs, respectively. This corresponds to ≈ 81% retrieval accuracy for the first hit.

It is well-known that the so-called Shannon Achievable Region (SAR) in a collocated two-user Gaussian Two-Way Channel (GTWC) does not depend on the self-interference that is due to the leakage of the signal transmitted by each user at its own receiver. This is simply because each user can completely remove its self-interference. In this paper, we study a class of GTWCs where each user is unable to cancel the self interference due to random erasures at its receiver. The mixture of the intended signal for each user and its self-interference is erased independently from transmission slot to transmission slot. It is assumed that both users adopt PAM constellations for transmission purposes. Due to the fact that both users are unaware of the erasure pattern, the noise plus interference at each user is mixed Gaussian. To analyze this setup, a sequence of upper and lower bounds are developed on the differential entropy of a general mixed Gaussian random variable where it is shown that the upper and lower bounds meet as the sequence index increases. Utilizing such bounds, it is shown that the achievable rate for each user is monotonically increasing in terms of the level of self-interference and eventually saturates as self-interference grows to infinity. This saturation effect is justified analytically by showing that as self-interference increases, each user is enabled to extract the erasure pattern at its receiver. Treating the erasure pattern as side information, both users are able to cancel self-interference and decode the useful information at higher transmission rates.