Scientific research faces high costs and inefficiencies with traditional methods, but the rise of deep learning and large language models (LLMs) offers innovative solutions. This survey reviews LLM applications across scientific fields such as biology, medicine, chemistry, and meteorology, underscoring their role in advancing research. However, the continuous expansion of model size has led to significant memory demands, hindering further development and application of LLMs for science. To address this, we review memory-efficient training techniques for LLMs based on the transformer architecture, including distributed training, mixed precision training, and gradient checkpointing. Using AlphaFold 2 as an example, we demonstrate how tailored memory optimization methods can reduce storage needs while preserving prediction accuracy. We also discuss the challenges of memory optimization in practice and potential future directions, hoping to provide valuable insights for researchers and engineers.

Temporal knowledge graph question answering (TKGQA) poses a significant challenge task, due to the temporal constraints hidden in questions and the answers sought from dynamic structured knowledge. Although large language models (LLMs) have made considerable progress in their reasoning ability over structured data, their application to the TKGQA task is a relatively unexplored area. This paper first proposes a novel generative temporal knowledge graph question answering framework, GenTKGQA, which guides LLMs to answer temporal questions through two phases: Subgraph Retrieval and Answer Generation. First, we exploit LLM's intrinsic knowledge to mine temporal constraints and structural links in the questions without extra training, thus narrowing down the subgraph search space in both temporal and structural dimensions. Next, we design virtual knowledge indicators to fuse the graph neural network signals of the subgraph and the text representations of the LLM in a non-shallow way, which helps the open-source LLM deeply understand the temporal order and structural dependencies among the retrieved facts through instruction tuning. Experimental results demonstrate that our model outperforms state-of-the-art baselines, even achieving 100\% on the metrics for the simple question type.

Convolutional neural network (CNN) has achieved impressive success in computer vision during the past few decades. The image convolution operation helps CNNs to get good performance on image-related tasks. However, it also has high computation complexity and hard to be parallelized. This paper proposes a novel Element-wise Multiplication Layer (EML) to replace convolution layers, which can be trained in the frequency domain. Theoretical analyses show that EMLs lower the computation complexity and easier to be parallelized. Moreover, we introduce a Weight Fixation mechanism to alleviate the problem of over-fitting, and analyze the working behavior of Batch Normalization and Dropout in the frequency domain. To get the balance between the computation complexity and memory usage, we propose a new network structure, namely Time-Frequency Domain Mixture Network (TFDMNet), which combines the advantages of both convolution layers and EMLs. Experimental results imply that TFDMNet achieves good performance on MNIST, CIFAR-10 and ImageNet databases with less number of operations comparing with corresponding CNNs.

Foundation models are becoming the dominant deep learning technologies. Pretraining a foundation model is always time-consumed due to the large scale of both the model parameter and training dataset. Besides being computing-intensive, the training process is extremely memory-intensive and communication-intensive. These features make it necessary to apply 3D parallelism, which integrates data parallelism, pipeline model parallelism and tensor model parallelism, to achieve high training efficiency. To achieve this goal, some custom software frameworks such as Megatron-LM and DeepSpeed are developed. However, current 3D parallelism frameworks still meet two issues: i) they are not transparent to model developers, which need to manually modify the model to parallelize training. ii) their utilization of computation, GPU memory and network bandwidth are not sufficient. We propose Merak, an automated 3D parallelism deep learning training framework with high resource utilization. Merak automatically deploys with an automatic model partitioner, which uses a graph sharding algorithm on a proxy representation of the model. Merak also presents the non-intrusive API for scaling out foundation model training with minimal code modification. In addition, we design a high-performance 3D parallel runtime engine in Merak. It uses several techniques to exploit available training resources, including shifted critical path pipeline schedule that brings a higher computation utilization, stage-aware recomputation that makes use of idle worker memory, and sub-pipelined tensor model parallelism that overlaps communication and computation. Experiments on 64 GPUs show Merak can speedup the training performance over the state-of-the-art 3D parallelism frameworks of models with 1.5, 2.5, 8.3, and 20 billion parameters by up to 1.42X, 1.39X, 1.43X, and 1.61X, respectively.

It is hard to train Recurrent Neural Network (RNN) with stable convergence and avoid gradient vanishing and exploding, as the weights in the recurrent unit are repeated from iteration to iteration. Moreover, RNN is sensitive to the initialization of weights and bias, which brings difficulty in the training phase. With the gradient-free feature and immunity to poor conditions, the Alternating Direction Method of Multipliers (ADMM) has become a promising algorithm to train neural networks beyond traditional stochastic gradient algorithms. However, ADMM could not be applied to train RNN directly since the state in the recurrent unit is repetitively updated over timesteps. Therefore, this work builds a new framework named ADMMiRNN upon the unfolded form of RNN to address the above challenges simultaneously and provides novel update rules and theoretical convergence analysis. We explicitly specify key update rules in the iterations of ADMMiRNN with deliberately constructed approximation techniques and solutions to each subproblem instead of vanilla ADMM. Numerical experiments are conducted on MNIST and text classification tasks, where ADMMiRNN achieves convergent results and outperforms compared baselines. Furthermore, ADMMiRNN trains RNN in a more stable way without gradient vanishing or exploding compared to the stochastic gradient algorithms. Source code has been available at https://github.com/TonyTangYu/ADMMiRNN.

Support vector machines (SVMs) with sparsity-inducing nonconvex penalties have received considerable attentions for the characteristics of automatic classification and variable selection. However, it is quite challenging to solve the nonconvex penalized SVMs due to their nondifferentiability, nonsmoothness and nonconvexity. In this paper, we propose an efficient ADMM-based algorithm to the nonconvex penalized SVMs. The proposed algorithm covers a large class of commonly used nonconvex regularization terms including the smooth clipped absolute deviation (SCAD) penalty, minimax concave penalty (MCP), log-sum penalty (LSP) and capped-$\ell_1$ penalty. The computational complexity analysis shows that the proposed algorithm enjoys low computational cost. Moreover, the convergence of the proposed algorithm is guaranteed. Extensive experimental evaluations on five benchmark datasets demonstrate the superior performance of the proposed algorithm to other three state-of-the-art approaches.

In this paper, we propose a stochastic Primal-Dual Hybrid Gradient (PDHG) approach for solving a wide spectrum of regularized stochastic minimization problems, where the regularization term is composite with a linear function. It has been recognized that solving this kind of problem is challenging since the closed-form solution of the proximal mapping associated with the regularization term is not available due to the imposed linear composition, and the per-iteration cost of computing the full gradient of the expected objective function is extremely high when the number of input data samples is considerably large. Our new approach overcomes these issues by exploring the special structure of the regularization term and sampling a few data points at each iteration. Rather than analyzing the convergence in expectation, we provide the detailed iteration complexity analysis for the cases of both uniformly and non-uniformly averaged iterates with high probability. This strongly supports the good practical performance of the proposed approach. Numerical experiments demonstrate that the efficiency of stochastic PDHG, which outperforms other competing algorithms, as expected by the high-probability convergence analysis.

We consider a wide range of regularized stochastic minimization problems with two regularization terms, one of which is composed with a linear function. This optimization model abstracts a number of important applications in artificial intelligence and machine learning, such as fused Lasso, fused logistic regression, and a class of graph-guided regularized minimization. The computational challenges of this model are in two folds. On one hand, the closed-form solution of the proximal mapping associated with the composed regularization term or the expected objective function is not available. On the other hand, the calculation of the full gradient of the expectation in the objective is very expensive when the number of input data samples is considerably large. To address these issues, we propose a stochastic variant of extra-gradient type methods, namely \textsf{Stochastic Primal-Dual Proximal ExtraGradient descent (SPDPEG)}, and analyze its convergence property for both convex and strongly convex objectives. For general convex objectives, the uniformly average iterates generated by \textsf{SPDPEG} converge in expectation with $O(1/\sqrt{t})$ rate. While for strongly convex objectives, the uniformly and non-uniformly average iterates generated by \textsf{SPDPEG} converge with $O(\log(t)/t)$ and $O(1/t)$ rates, respectively. The order of the rate of the proposed algorithm is known to match the best convergence rate for first-order stochastic algorithms. Experiments on fused logistic regression and graph-guided regularized logistic regression problems show that the proposed algorithm performs very efficiently and consistently outperforms other competing algorithms.