We interpret a deep-learning semantic segmentation model used to classify coastline satellite images into land and water. This is to build trust in the model and gain new insight into the process of coastal water body extraction. Specifically, we seek to understand which spectral bands are important for predicting segmentation masks. This is done using a permutation importance approach. Results show that the NIR is the most important spectral band. Permuting this band lead to a decrease in accuracy of 38.12 percentage points. This is followed by Water Vapour, SWIR 1, and Blue bands with 2.58, 0.78 and 0.19 respectively. Water Vapour is not typically used in water indices and these results suggest it may be useful for water body extraction. Permuting, the Coastal Aerosol, Green, Red, RE1, RE2, RE3, RE4, and SWIR 2 bands did not decrease accuracy. This suggests they could be excluded from future model builds reducing complexity and computational requirements.
With over 200 million published academic documents and millions of new documents being written each year, academic researchers face the challenge of searching for information within this vast corpus. However, existing retrieval systems struggle to understand the semantics and domain knowledge present in academic papers. In this work, we demonstrate that by utilizing large language models, a document retrieval system can achieve advanced semantic understanding capabilities, significantly outperforming existing systems. Our approach involves training the retriever and reranker using domain-specific data generated by large language models. Additionally, we utilize large language models to identify candidates from the references of retrieved papers to further enhance the performance. We use a test set annotated by academic researchers in the fields of quantum physics and computer vision to evaluate our system's performance. The results show that DocReLM achieves a Top 10 accuracy of 44.12% in computer vision, compared to Google Scholar's 15.69%, and an increase to 36.21% in quantum physics, while that of Google Scholar is 12.96%.
Self-distillation methods using Siamese networks are popular for self-supervised pre-training. DINO is one such method based on a cross-entropy loss between $K$-dimensional probability vectors, obtained by applying a softmax function to the dot product between representations and learnt prototypes. Given the fact that the learned representations are $L^2$-normalized, we show that DINO and its derivatives, such as iBOT, can be interpreted as a mixture model of von Mises-Fisher components. With this interpretation, DINO assumes equal precision for all components when the prototypes are also $L^2$-normalized. Using this insight we propose DINO-vMF, that adds appropriate normalization constants when computing the cluster assignment probabilities. Unlike DINO, DINO-vMF is stable also for the larger ViT-Base model with unnormalized prototypes. We show that the added flexibility of the mixture model is beneficial in terms of better image representations. The DINO-vMF pre-trained model consistently performs better than DINO on a range of downstream tasks. We obtain similar improvements for iBOT-vMF vs iBOT and thereby show the relevance of our proposed modification also for other methods derived from DINO.
Optimization algorithms and large language models (LLMs) enhance decision-making in dynamic environments by integrating artificial intelligence with traditional techniques. LLMs, with extensive domain knowledge, facilitate intelligent modeling and strategic decision-making in optimization, while optimization algorithms refine LLM architectures and output quality. This synergy offers novel approaches for advancing general AI, addressing both the computational challenges of complex problems and the application of LLMs in practical scenarios. This review outlines the progress and potential of combining LLMs with optimization algorithms, providing insights for future research directions.
In this paper, we propose a predictive quantifier to estimate the retraining cost of a trained model in distribution shifts. The proposed Aggregated Representation Measure (ARM) quantifies the change in the model's representation from the old to new data distribution. It provides, before actually retraining the model, a single concise index of resources - epochs, energy, and carbon emissions - required for the retraining. This enables reuse of a model with a much lower cost than training a new model from scratch. The experimental results indicate that ARM reasonably predicts retraining costs for varying noise intensities and enables comparisons among multiple model architectures to determine the most cost-effective and sustainable option.
This paper presents a novel compact delay model of Ovonic Threshold Switch (OTS) devices that works efficiently for circuit simulations. The internal state variable of the two terminal devices is estimated using a delay system that uses a few electrical components related to a suggested equivalent circuit of the device. Finally, we tested the proposed model against measured data from devices fabricated by Western Digital Research.
Palm oil production has been identified as one of the major drivers of deforestation for tropical countries. To meet supply chain objectives, commodity producers and other stakeholders need timely information of land cover dynamics in their supply shed. However, such data are difficult to obtain from suppliers who may lack digital geographic representations of their supply sheds and production locations. Here we present a "community model," a machine learning model trained on pooled data sourced from many different stakeholders, to produce a map of palm probability at global scale. An advantage of this method is the inclusion of varied inputs, the ability to easily update the model as new training data becomes available and run the model on any year that input imagery is available. Inclusion of diverse data sources into one probability map can help establish a shared understanding across stakeholders on the presence and absence of a land cover or commodity (in this case oil palm). The model predictors are annual composites built from publicly available satellite imagery provided by Sentinel-1, Sentinel-2, and ALOS-2, and terrain data from Jaxa (AW3D30) and Copernicus (GLO-30). We provide map outputs as the probability of palm in a given pixel, to reflect the uncertainty of the underlying state (palm or not palm). This version of this model provides global accuracy estimated to be 92% (at 0.5 probability threshold) on an independent test set. This model, and resulting oil palm probability map products are useful for accurately identifying the geographic footprint of palm cultivation. Used in conjunction with timely deforestation information, this palm model is useful for understanding the risk of continued oil palm plantation expansion in sensitive forest areas.
Large language models (LLMs) have shown incredible proficiency in performing tasks that require semantic understanding of natural language instructions. Recently, many works have further expanded this capability to perceive multimodal audio and text inputs, but their capabilities are often limited to specific fine-tuned tasks such as automatic speech recognition and translation. We therefore develop SpeechVerse, a robust multi-task training and curriculum learning framework that combines pre-trained speech and text foundation models via a small set of learnable parameters, while keeping the pre-trained models frozen during training. The models are instruction finetuned using continuous latent representations extracted from the speech foundation model to achieve optimal zero-shot performance on a diverse range of speech processing tasks using natural language instructions. We perform extensive benchmarking that includes comparing our model performance against traditional baselines across several datasets and tasks. Furthermore, we evaluate the model's capability for generalized instruction following by testing on out-of-domain datasets, novel prompts, and unseen tasks. Our empirical experiments reveal that our multi-task SpeechVerse model is even superior to conventional task-specific baselines on 9 out of the 11 tasks.
Amazon Web Services (AWS) offers three important Model Deployment Services for model developers: SageMaker, Lambda, and Elastic Container Service (ECS). These services have critical advantages and disadvantages, influencing model developer's adoption decisions. This comparative analysis reviews the merits and drawbacks of these services. This analysis found that Lambda AWS service leads in efficiency, autoscaling aspects, and integration during model development. However, ECS was found to be outstanding in terms of flexibility, scalability, and infrastructure control; conversely, ECS is better suited when it comes to managing complex container environments during model development, as well as addressing budget concerns -- it is, therefore, the preferred option for model developers whose objective is to achieve complete freedom and framework flexibility with horizontal scaling. ECS is better suited to ensuring performance requirements align with project goals and constraints. The AWS service selection process considered factors that include but are not limited to load balance and cost-effectiveness. ECS is a better choice when model development begins from the abstract. It offers unique benefits, such as the ability to scale horizontally and vertically, making it the best preferable tool for model deployment.
In this paper, we introduce SUTRA, multilingual Large Language Model architecture capable of understanding, reasoning, and generating text in over 50 languages. SUTRA's design uniquely decouples core conceptual understanding from language-specific processing, which facilitates scalable and efficient multilingual alignment and learning. Employing a Mixture of Experts framework both in language and concept processing, SUTRA demonstrates both computational efficiency and responsiveness. Through extensive evaluations, SUTRA is demonstrated to surpass existing models like GPT-3.5, Llama2 by 20-30% on leading Massive Multitask Language Understanding (MMLU) benchmarks for multilingual tasks. SUTRA models are also online LLMs that can use knowledge from the internet to provide hallucination-free, factual and up-to-date responses while retaining their multilingual capabilities. Furthermore, we explore the broader implications of its architecture for the future of multilingual AI, highlighting its potential to democratize access to AI technology globally and to improve the equity and utility of AI in regions with predominantly non-English languages. Our findings suggest that SUTRA not only fills pivotal gaps in multilingual model capabilities but also establishes a new benchmark for operational efficiency and scalability in AI applications.
Human motion synthesis is an important task in computer graphics and computer vision. While focusing on various conditioning signals such as text, action class, or audio to guide the generation process, most existing methods utilize skeleton-based pose representation, requiring additional skinning to produce renderable meshes. Given that human motion is a complex interplay of bones, joints, and muscles, considering solely the skeleton for generation may neglect their inherent interdependency, which can limit the variability and precision of the generated results. To address this issue, we propose a Shape-conditioned Motion Diffusion model (SMD), which enables the generation of motion sequences directly in mesh format, conditioned on a specified target mesh. In SMD, the input meshes are transformed into spectral coefficients using graph Laplacian, to efficiently represent meshes. Subsequently, we propose a Spectral-Temporal Autoencoder (STAE) to leverage cross-temporal dependencies within the spectral domain. Extensive experimental evaluations show that SMD not only produces vivid and realistic motions but also achieves competitive performance in text-to-motion and action-to-motion tasks when compared to state-of-the-art methods.
This study explores the innovative use of Large Language Models (LLMs) as analytical tools for interpreting complex financial regulations. The primary objective is to design effective prompts that guide LLMs in distilling verbose and intricate regulatory texts, such as the Basel III capital requirement regulations, into a concise mathematical framework that can be subsequently translated into actionable code. This novel approach aims to streamline the implementation of regulatory mandates within the financial reporting and risk management systems of global banking institutions. A case study was conducted to assess the performance of various LLMs, demonstrating that GPT-4 outperforms other models in processing and collecting necessary information, as well as executing mathematical calculations. The case study utilized numerical simulations with asset holdings -- including fixed income, equities, currency pairs, and commodities -- to demonstrate how LLMs can effectively implement the Basel III capital adequacy requirements. Keywords: Large Language Models, Prompt Engineering, LLMs in Finance, Basel III, Minimum Capital Requirements, LLM Ethics
Over the past few years, as large language models have ushered in an era of intelligence emergence, there has been an intensified focus on scaling networks. Currently, many network architectures are designed manually, often resulting in sub-optimal configurations. Although Neural Architecture Search (NAS) methods have been proposed to automate this process, they suffer from low search efficiency. This study introduces Differentiable Model Scaling (DMS), increasing the efficiency for searching optimal width and depth in networks. DMS can model both width and depth in a direct and fully differentiable way, making it easy to optimize. We have evaluated our DMS across diverse tasks, ranging from vision tasks to NLP tasks and various network architectures, including CNNs and Transformers. Results consistently indicate that our DMS can find improved structures and outperforms state-of-the-art NAS methods. Specifically, for image classification on ImageNet, our DMS improves the top-1 accuracy of EfficientNet-B0 and Deit-Tiny by 1.4% and 0.6%, respectively, and outperforms the state-of-the-art zero-shot NAS method, ZiCo, by 1.3% while requiring only 0.4 GPU days for searching. For object detection on COCO, DMS improves the mAP of Yolo-v8-n by 2.0%. For language modeling, our pruned Llama-7B outperforms the prior method with lower perplexity and higher zero-shot classification accuracy. We will release our code in the future.
Electrical power conversions are common in a large variety of engineering applications. With reference to AC/DC and DC/AC power conversions, a strong research interest resides in multilevel converters, thanks to the many advantages they provide over standard two-level converters. In this paper, we first provide a power-oriented model of Modular Multilevel Converters (MMCs), followed by a detailed harmonic analysis. The model is given in the form of a block scheme that can be directly implemented in the Matlab/Simulink environment. The performed harmonic analysis gives a deep and exact understanding of the different terms affecting the evolution of the voltage trajectories in the upper and lower arms of the converter. Next, we propose a new model-based adaptive control scheme for MMCs. The proposed control allows to determine the optimal average capacitor voltages reference in real-time, thus allowing to properly track the desired load current while minimizing the harmonic content in the generated load current itself.
We develop a methodology for analyzing language model task performance at the individual example level based on training data density estimation. Experiments with paraphrasing as a controlled intervention on finetuning data demonstrate that increasing the support in the training distribution for specific test queries results in a measurable increase in density, which is also a significant predictor of the performance increase caused by the intervention. Experiments with pretraining data demonstrate that we can explain a significant fraction of the variance in model perplexity via density measurements. We conclude that our framework can provide statistical evidence of the dependence of a target model's predictions on subsets of its training data, and can more generally be used to characterize the support (or lack thereof) in the training data for a given test task.
Enhancing 3D and Z-axis positioning accuracy is crucial for effective rescue in indoor emergencies, ensuring safety for emergency responders and at-risk individuals. Additionally, reducing the dependence of a positioning system on fixed infrastructure is crucial, given its vulnerability to power failures and damage during emergencies. Further challenges from a signal propagation perspective include poor indoor signal coverage, multipath effects and the problem of Non-Line-OfSight (NLOS) measurement bias. In this study, we utilize the mobility provided by a rapidly deployable Uncrewed Aerial Vehicle (UAV) based wireless network to address these challenges. We recognize diffraction from window edges as a crucial signal propagation mechanism and employ the Geometrical Theory of Diffraction (GTD) to introduce a novel NLOS path length model. Using this path length model, we propose two different techniques to improve the indoor positioning performance for emergency scenarios.
Transformer-based models trained on large and general purpose datasets consisting of molecular strings have recently emerged as a powerful tool for successfully modeling various structure-property relations. Inspired by this success, we extend the paradigm of training chemical language transformers on large-scale chemical datasets to generative tasks in this work. Specifically, we propose GP-MoLFormer, an autoregressive molecular string generator that is trained on more than 1.1B (billion) chemical SMILES. GP-MoLFormer uses a 46.8M parameter transformer decoder model with linear attention and rotary positional encodings as the base architecture. GP-MoLFormer's utility is evaluated and compared with that of existing baselines on three different tasks: de novo generation, scaffold-constrained molecular decoration, and unconstrained property-guided optimization. While the first two are handled with no additional training, we propose a parameter-efficient fine-tuning method for the last task, which uses property-ordered molecular pairs as input. We call this new approach pair-tuning. Our results show GP-MoLFormer performs better or comparable with baselines across all three tasks, demonstrating its general utility for a variety of molecular generation tasks. We further report strong memorization of training data in GP-MoLFormer generations, which has so far remained unexplored for chemical language models. Our analyses reveal that training data memorization and novelty in generations are impacted by the quality and scale of the training data; duplication bias in training data can enhance memorization at the cost of lowering novelty. We further establish a scaling law relating inference compute and novelty in generations.
I propose a quantile-based nonadditive fixed effects panel model to study heterogeneous causal effects. Similar to standard fixed effects (FE) model, my model allows arbitrary dependence between regressors and unobserved heterogeneity, but it generalizes the additive separability of standard FE to allow the unobserved heterogeneity to enter nonseparably. Similar to structural quantile models, my model's random coefficient vector depends on an unobserved, scalar ''rank'' variable, in which outcomes (excluding an additive noise term) are monotonic at a particular value of the regressor vector, which is much weaker than the conventional monotonicity assumption that must hold at all possible values. This rank is assumed to be stable over time, which is often more economically plausible than the panel quantile studies that assume individual rank is iid over time. It uncovers the heterogeneous causal effects as functions of the rank variable. I provide identification and estimation results, establishing uniform consistency and uniform asymptotic normality of the heterogeneous causal effect function estimator. Simulations show reasonable finite-sample performance and show my model complements fixed effects quantile regression. Finally, I illustrate the proposed methods by examining the causal effect of a country's oil wealth on its military defense spending.
Model Inversion (MI) attacks aim to reconstruct private training data by abusing access to machine learning models. Contemporary MI attacks have achieved impressive attack performance, posing serious threats to privacy. Meanwhile, all existing MI defense methods rely on regularization that is in direct conflict with the training objective, resulting in noticeable degradation in model utility. In this work, we take a different perspective, and propose a novel and simple Transfer Learning-based Defense against Model Inversion (TL-DMI) to render MI-robust models. Particularly, by leveraging TL, we limit the number of layers encoding sensitive information from private training dataset, thereby degrading the performance of MI attack. We conduct an analysis using Fisher Information to justify our method. Our defense is remarkably simple to implement. Without bells and whistles, we show in extensive experiments that TL-DMI achieves state-of-the-art (SOTA) MI robustness. Our code, pre-trained models, demo and inverted data are available at: https://hosytuyen.github.io/projects/TL-DMI
With the recent success of large language models, particularly foundation models with generalization abilities, applying foundation models for recommendations becomes a new paradigm to improve existing recommendation systems. It becomes a new open challenge to enable the foundation model to capture user preference changes in a timely manner with reasonable communication and computation costs while preserving privacy. This paper proposes a novel federated adaptation mechanism to enhance the foundation model-based recommendation system in a privacy-preserving manner. Specifically, each client will learn a lightweight personalized adapter using its private data. The adapter then collaborates with pre-trained foundation models to provide recommendation service efficiently with fine-grained manners. Importantly, users' private behavioral data remains secure as it is not shared with the server. This data localization-based privacy preservation is embodied via the federated learning framework. The model can ensure that shared knowledge is incorporated into all adapters while simultaneously preserving each user's personal preferences. Experimental results on four benchmark datasets demonstrate our method's superior performance. Implementation code is available to ease reproducibility.