Language models (LMs) proficiency in handling deterministic symbolic reasoning and rule-based tasks remains limited due to their dependency implicit learning on textual data. To enable fully rule comprehension ability, we explore how to incorporate compiled neural networks (CoNNs) which weight is specially designed into the architecture of LMs, to achieve high accuracy and robust performance. CoNNs are transformer-based neural networks that execute rules through artificially generated attention weights. Our method, which call "Neural Comprehension", by incorporating CoNN modules into the LM, the framework effectively tackles rule-intensive challenges. Our experiments on symbolic reasoning tasks and real-world arithmetic reasoning tasks demonstrate the superior performance of our method compared to existing techniques. Furthermore, our LM achieves flawless execution on symbolic operations tasks, highlighting the potential of our method in enabling LMs to possess true symbolic comprehension capabilities. Our code is publicly available at: https://github.com/WENGSYX/Neural-Comprehension.

This paper proposes an actor-critic algorithm for controlling the temperature of a battery pack using a cooling fluid. This is modeled by a coupled 1D partial differential equation (PDE) with a controlled advection term that determines the speed of the cooling fluid. The Hamilton-Jacobi-Bellman (HJB) equation is a PDE that evaluates the optimality of the value function and determines an optimal controller. We propose an algorithm that treats the value network as a Physics-Informed Neural Network (PINN) to solve for the continuous-time HJB equation rather than a discrete-time Bellman optimality equation, and we derive an optimal controller for the environment that we exploit to achieve optimal control. Our experiments show that a hybrid-policy method that updates the value network using the HJB equation and updates the policy network identically to PPO achieves the best results in the control of this PDE system.

A solar sail is one of the most promising space exploration system because of its theoretically infinite specific impulse using solar radiation pressure (SRP). Recently, some researchers proposed "transformable spacecrafts" that can actively reconfigure their body configurations with actuatable joints. The transformable spacecrafts are expected to greatly enhance orbit and attitude control capability due to its high redundancy in control degree of freedom if they are used like solar sails. However, its large number of input poses difficulties in control, and therefore, previous researchers imposed strong constraints to limit its potential control capabilities. This paper addresses novel attitude control techniques for the transformable spacecrafts under SRP. The authors have constructed two proposed methods; one of those is a joint angle optimization to acquire arbitrary SRP force and torque, and the other is a momentum damping control driven by joint angle actuation. Our proposed methods are formulated in general forms and applicable to any transformable spacecraft that has front faces that can dominantly receive SRP on each body. Validity of the proposed methods are confirmed by numerical simulations. This paper contributes to making most of the high control redundancy of transformable spacecrafts without consuming any expendable propellants, which is expected to greatly enhance orbit and attitude control capability.

Energy consumption from selecting, training and deploying deep learning models has continued to increase over the past few years. Our goal in this work is to support the design of energy-efficient deep learning models that are easier to train with lower compute resources, practical to deploy in real-world edge/mobile computing settings and environmentally sustainable. Tabular benchmarks for neural architecture search (NAS) allow the evaluation of NAS strategies at lower computational cost by providing pre-computed performance statistics. In this work, we suggest including energy efficiency as an additional performance criterion to NAS and present an updated tabular benchmark by including information on energy consumption and carbon footprint for different architectures. The benchmark called EC-NAS is made available open-source to support energy consumption-aware NAS research. EC-NAS also includes a surrogate model for predicting energy consumption, and helps us reduce the overall energy cost of creating this dataset. We demonstrate the usefulness of EC-NAS by applying multi-objective optimisation algorithms that reveal the trade-off between energy consumption and accuracy, showing that it is possible to discover energy-efficient architectures with little to no loss in performance.

Typical generative diffusion models rely on a Gaussian diffusion process for training the backward transformations, which can then be used to generate samples from Gaussian noise. However, real world data often takes place in discrete-state spaces, including many scientific applications. Here, we develop a theoretical formulation for arbitrary discrete-state Markov processes in the forward diffusion process using exact (as opposed to variational) analysis. We relate the theory to the existing continuous-state Gaussian diffusion as well as other approaches to discrete diffusion, and identify the corresponding reverse-time stochastic process and score function in the continuous-time setting, and the reverse-time mapping in the discrete-time setting. As an example of this framework, we introduce ``Blackout Diffusion'', which learns to produce samples from an empty image instead of from noise. Numerical experiments on the CIFAR-10, Binarized MNIST, and CelebA datasets confirm the feasibility of our approach. Generalizing from specific (Gaussian) forward processes to discrete-state processes without a variational approximation sheds light on how to interpret diffusion models, which we discuss.

Long-term time series forecasting has gained significant attention in recent years. While there are various specialized designs for capturing temporal dependency, previous studies have demonstrated that a single linear layer can achieve competitive forecasting performance compared to other complex architectures. In this paper, we thoroughly investigate the intrinsic effectiveness of recent approaches and make three key observations: 1) linear mapping is critical to prior long-term time series forecasting efforts; 2) RevIN (reversible normalization) and CI (Channel Independent) play a vital role in improving overall forecasting performance; and 3) linear mapping can effectively capture periodic features in time series and has robustness for different periods across channels when increasing the input horizon. We provide theoretical and experimental explanations to support our findings and also discuss the limitations and future works. Our framework's code is available at https://github.com/plumprc/RTSF.

We present the first controller for quasistatic robotic planar pushing with single-point contact using only force feedback. We consider a mobile robot equipped with a force-torque sensor to measure the force at the contact point with the pushed object (the "slider"). The parameters of the slider are not known to the controller, nor is feedback on the slider's pose. We assume that the global position of the contact point is always known and that the approximate initial position of the slider is provided. We focus specifically on the case when it is desired to push the slider along a straight line. Simulations and real-world experiments show that our controller yields stable pushes that are robust to a wide range of slider parameters and state perturbations.

Large language models are trained in two stages: (1) unsupervised pretraining from raw text, to learn general-purpose representations, and (2) large scale instruction tuning and reinforcement learning, to better align to end tasks and user preferences. We measure the relative importance of these two stages by training LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong performance, learning to follow specific response formats from only a handful of examples in the training data, including complex queries that range from planning trip itineraries to speculating about alternate history. Moreover, the model tends to generalize well to unseen tasks that did not appear in the training data. In a controlled human study, responses from LIMA are either equivalent or strictly preferred to GPT-4 in 43% of cases; this statistic is as high as 58% when compared to Bard and 65% versus DaVinci003, which was trained with human feedback. Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.

The accurate but fast calculation of molecular excited states is still a very challenging topic. For many applications, detailed knowledge of the energy funnel in larger molecular aggregates is of key importance requiring highly accurate excited state energies. To this end, machine learning techniques can be an extremely useful tool though the cost of generating highly accurate training datasets still remains a severe challenge. To overcome this hurdle, this work proposes the use of multi-fidelity machine learning where very little training data from high accuracies is combined with cheaper and less accurate data to achieve the accuracy of the costlier level. In the present study, the approach is employed to predict the first excited state energies for three molecules of increasing size, namely, benzene, naphthalene, and anthracene. The energies are trained and tested for conformations stemming from classical molecular dynamics simulations and from real-time density functional tight-binding calculations. It can be shown that the multi-fidelity machine learning model can achieve the same accuracy as a machine learning model built only on high cost training data while having a much lower computational effort to generate the data. The numerical gain observed in these benchmark test calculations was over a factor of 30 but certainly can be much higher for high accuracy data.

Data-driven industrial health prognostics require rich training data to develop accurate and reliable predictive models. However, stringent data privacy laws and the abundance of edge industrial data necessitate decentralized data utilization. Thus, the industrial health prognostics field is well suited to significantly benefit from federated learning (FL), a decentralized and privacy-preserving learning technique. However, FL-based health prognostics tasks have hardly been investigated due to the complexities of meaningfully aggregating model parameters trained from heterogeneous data to form a high performing federated model. Specifically, data heterogeneity among edge devices, stemming from dissimilar degradation mechanisms and unequal dataset sizes, poses a critical statistical challenge for developing accurate federated models. We propose a pioneering FL-based health prognostic model with a feature similarity-matched parameter aggregation algorithm to discriminatingly learn from heterogeneous edge data. The algorithm searches across the heterogeneous locally trained models and matches neurons with probabilistically similar feature extraction functions first, before selectively averaging them to form the federated model parameters. As the algorithm only averages similar neurons, as opposed to conventional naive averaging of coordinate-wise neurons, the distinct feature extractors of local models are carried over with less dilution to the resultant federated model. Using both cyclic degradation data of Li-ion batteries and non-cyclic data of turbofan engines, we demonstrate that the proposed method yields accuracy improvements as high as 44.5\% and 39.3\% for state-of-health estimation and remaining useful life estimation, respectively.