As Large Language Models (LLMs) are deployed more widely, customization with respect to vocabulary, style and character becomes more important. In this work we introduce model arithmetic, a novel inference framework for composing and biasing LLMs without the need for model (re)training or highly specific datasets. In addition, the framework allows for more precise control of generated text than direct prompting and prior controlled text generation (CTG) techniques. Using model arithmetic, we can express prior CTG techniques as simple formulas and naturally extend them to new and more effective formulations. Further, we show that speculative sampling, a technique for efficient LLM sampling, extends to our setting. This enables highly efficient text generation with multiple composed models with only marginal overhead over a single model. Our empirical evaluation demonstrates that model arithmetic allows fine-grained control of generated text while outperforming state-of-the-art on the task of toxicity reduction.

This paper provides an efficient training-free painterly image harmonization (PIH) method, dubbed FreePIH, that leverages only a pre-trained diffusion model to achieve state-of-the-art harmonization results. Unlike existing methods that require either training auxiliary networks or fine-tuning a large pre-trained backbone, or both, to harmonize a foreground object with a painterly-style background image, our FreePIH tames the denoising process as a plug-in module for foreground image style transfer. Specifically, we find that the very last few steps of the denoising (i.e., generation) process strongly correspond to the stylistic information of images, and based on this, we propose to augment the latent features of both the foreground and background images with Gaussians for a direct denoising-based harmonization. To guarantee the fidelity of the harmonized image, we make use of multi-scale features to enforce the consistency of the content and stability of the foreground objects in the latent space, and meanwhile, aligning both fore-/back-grounds with the same style. Moreover, to accommodate the generation with more structural and textural details, we further integrate text prompts to attend to the latent features, hence improving the generation quality. Quantitative and qualitative evaluations on COCO and LAION 5B datasets demonstrate that our method can surpass representative baselines by large margins.

Support Vector Machine (SVM) algorithm requires a high computational cost (both in memory and time) to solve a complex quadratic programming (QP) optimization problem during the training process. Consequently, SVM necessitates high computing hardware capabilities. The central processing unit (CPU) clock frequency cannot be increased due to physical limitations in the miniaturization process. However, the potential of parallel multi-architecture, available in both multi-core CPUs and highly scalable GPUs, emerges as a promising solution to enhance algorithm performance. Therefore, there is an opportunity to reduce the high computational time required by SVM for solving the QP optimization problem. This paper presents a comparative study that implements the SVM algorithm on different parallel architecture frameworks. The experimental results show that SVM MPI-CUDA implementation achieves a speedup over SVM TensorFlow implementation on different datasets. Moreover, SVM TensorFlow implementation provides a cross-platform solution that can be migrated to alternative hardware components, which will reduces the development time.

This document describes an approach used in the Multi-Machine Disruption Prediction Challenge for Fusion Energy by ITU, a data science competition which ran from September to November 2023, on the online platform Zindi. The competition involved data from three fusion devices - C-Mod, HL-2A, and J-TEXT - with most of the training data coming from the last two, and the test data coming from the first one. Each device has multiple diagnostics and signals, and it turns out that a critical issue in this competition was to identify which signals, and especially which features from those signals, were most relevant to achieve accurate predictions. The approach described here is based on extracting features from signals, and then applying logistic regression on top of those features. Each signal is treated as a separate predictor and, in the end, a combination of such predictors achieved the first place on the leaderboard.

The importance of systems that can extract structured information from textual data becomes increasingly pronounced given the ever-increasing volume of text produced on a daily basis. Having a system that can effectively extract such information in an interoperable manner would be an asset for several domains, be it finance, health, or legal. Recent developments in natural language processing led to the production of powerful language models that can, to some degree, mimic human intelligence. Such effectiveness raises a pertinent question: Can these models be leveraged for the extraction of structured information? In this work, we address this question by evaluating the capabilities of two state-of-the-art language models -- GPT-3 and GPT-3.5, commonly known as ChatGPT -- in the extraction of narrative entities, namely events, participants, and temporal expressions. This study is conducted on the Text2Story Lusa dataset, a collection of 119 Portuguese news articles whose annotation framework includes a set of entity structures along with several tags and attribute values. We first select the best prompt template through an ablation study over prompt components that provide varying degrees of information on a subset of documents of the dataset. Subsequently, we use the best templates to evaluate the effectiveness of the models on the remaining documents. The results obtained indicate that GPT models are competitive with out-of-the-box baseline systems, presenting an all-in-one alternative for practitioners with limited resources. By studying the strengths and limitations of these models in the context of information extraction, we offer insights that can guide future improvements and avenues to explore in this field.

Machine translation (MT) for low-resource languages such as Ge'ez, an ancient language that is no longer spoken in daily life, faces challenges such as out-of-vocabulary words, domain mismatches, and lack of sufficient labeled training data. In this work, we explore various methods to improve Ge'ez MT, including transfer-learning from related languages, optimizing shared vocabulary and token segmentation approaches, finetuning large pre-trained models, and using large language models (LLMs) for few-shot translation with fuzzy matches. We develop a multilingual neural machine translation (MNMT) model based on languages relatedness, which brings an average performance improvement of about 4 BLEU compared to standard bilingual models. We also attempt to finetune the NLLB-200 model, one of the most advanced translation models available today, but find that it performs poorly with only 4k training samples for Ge'ez. Furthermore, we experiment with using GPT-3.5, a state-of-the-art LLM, for few-shot translation with fuzzy matches, which leverages embedding similarity-based retrieval to find context examples from a parallel corpus. We observe that GPT-3.5 achieves a remarkable BLEU score of 9.2 with no initial knowledge of Ge'ez, but still lower than the MNMT baseline of 15.2. Our work provides insights into the potential and limitations of different approaches for low-resource and ancient language MT.

We perform a comparative analysis of transformer-based models designed for modeling tabular data, specifically on an industry-scale dataset. While earlier studies demonstrated promising outcomes on smaller public or synthetic datasets, the effectiveness did not extend to larger industry-scale datasets. The challenges identified include handling high-dimensional data, the necessity for efficient pre-processing of categorical and numerical features, and addressing substantial computational requirements. To overcome the identified challenges, the study conducts an extensive examination of various transformer-based models using both synthetic datasets and the default prediction Kaggle dataset (2022) from American Express. The paper presents crucial insights into optimal data pre-processing, compares pre-training and direct supervised learning methods, discusses strategies for managing categorical and numerical features, and highlights trade-offs between computational resources and performance. Focusing on temporal financial data modeling, the research aims to facilitate the systematic development and deployment of transformer-based models in real-world scenarios, emphasizing scalability.

In this paper, we contend that a natural objective of representation learning is to compress and transform the distribution of the data, say sets of tokens, towards a low-dimensional Gaussian mixture supported on incoherent subspaces. The goodness of such a representation can be evaluated by a principled measure, called sparse rate reduction, that simultaneously maximizes the intrinsic information gain and extrinsic sparsity of the learned representation. From this perspective, popular deep network architectures, including transformers, can be viewed as realizing iterative schemes to optimize this measure. Particularly, we derive a transformer block from alternating optimization on parts of this objective: the multi-head self-attention operator compresses the representation by implementing an approximate gradient descent step on the coding rate of the features, and the subsequent multi-layer perceptron sparsifies the features. This leads to a family of white-box transformer-like deep network architectures, named CRATE, which are mathematically fully interpretable. We show, by way of a novel connection between denoising and compression, that the inverse to the aforementioned compressive encoding can be realized by the same class of CRATE architectures. Thus, the so-derived white-box architectures are universal to both encoders and decoders. Experiments show that these networks, despite their simplicity, indeed learn to compress and sparsify representations of large-scale real-world image and text datasets, and achieve performance very close to highly engineered transformer-based models: ViT, MAE, DINO, BERT, and GPT2. We believe the proposed computational framework demonstrates great potential in bridging the gap between theory and practice of deep learning, from a unified perspective of data compression. Code is available at: https://ma-lab-berkeley.github.io/CRATE .

Despite efforts to align large language models to produce harmless responses, they are still vulnerable to jailbreak prompts that elicit unrestricted behaviour. In this work, we investigate persona modulation as a black-box jailbreaking method to steer a target model to take on personalities that are willing to comply with harmful instructions. Rather than manually crafting prompts for each persona, we automate the generation of jailbreaks using a language model assistant. We demonstrate a range of harmful completions made possible by persona modulation, including detailed instructions for synthesising methamphetamine, building a bomb, and laundering money. These automated attacks achieve a harmful completion rate of 42.5% in GPT-4, which is 185 times larger than before modulation (0.23%). These prompts also transfer to Claude 2 and Vicuna with harmful completion rates of 61.0% and 35.9%, respectively. Our work reveals yet another vulnerability in commercial large language models and highlights the need for more comprehensive safeguards.

Artificial Intelligence (AI) has made its way into various scientific fields, providing astonishing improvements over existing algorithms for a wide variety of tasks. In recent years, there have been severe concerns over the trustworthiness of AI technologies. The scientific community has focused on the development of trustworthy AI algorithms. However, machine and deep learning algorithms, popular in the AI community today, depend heavily on the data used during their development. These learning algorithms identify patterns in the data, learning the behavioral objective. Any flaws in the data have the potential to translate directly into algorithms. In this study, we discuss the importance of Responsible Machine Learning Datasets and propose a framework to evaluate the datasets through a responsible rubric. While existing work focuses on the post-hoc evaluation of algorithms for their trustworthiness, we provide a framework that considers the data component separately to understand its role in the algorithm. We discuss responsible datasets through the lens of fairness, privacy, and regulatory compliance and provide recommendations for constructing future datasets. After surveying over 100 datasets, we use 60 datasets for analysis and demonstrate that none of these datasets is immune to issues of fairness, privacy preservation, and regulatory compliance. We provide modifications to the ``datasheets for datasets" with important additions for improved dataset documentation. With governments around the world regularizing data protection laws, the method for the creation of datasets in the scientific community requires revision. We believe this study is timely and relevant in today's era of AI.