Les chercheurs de Vector présentent leurs travaux à ACL 2024
Code Hallucination
In this study, we introduce various code hallucinations created manually via large language models. We detail HallTrigger, a method for efficiently inducing arbitrary code hallucinations. It utilizes three dynamic aspects of LLMs to elicit hallucinations without accessing the model's architecture or parameters. Tests on well-known blackbox models indicate HallTrigger's effectiveness and its significant influence on software development.
Looking into Black Box Code Language Models
Recent studies on code language models (LMs) primarily focus on performance metrics, often treating LMs as inscrutable entities, while only a few explore the role of attention layers. This research delves into the feed-forward layers of code LMs, using Codegen-Mono and PolyCoder, and languages like Java, Go, and Python, to understand concept organization, editability, and layer roles in output generation. Findings reveal lower layers grasp syntax, upper layers grasp abstract concepts, and editing within feed-forward layers is feasible without affecting performance, offering insights for improved code LM understanding and development.
Les différentes philosophies d’implémentation en RL
Cet article a été par Ludovic Denoyer. Il est chercheur scientifique au FAIR, se concentrant principalement sur divers problèmes d’apprentissage automatique, en particulier sur l’apprentissage par renforcement et l’interaction homme-machine. Il était auparavant professeur à Sorbonne Universités. Rendre plus simple la programmation d’algorithmes d’apprentissage par renforcement à l’aide des principes de l’apprentissage supervisé : la librairie « RLStructures ».
ChartGemma: Visual Instruction-tuning for Chart Reasoning in the Wild
ChartGemma, a new model for chart understanding, overcomes the limitations of prior models by training directly on chart images, not just data tables. This approach captures detailed visual trends, leading to superior performance in summarization, question answering, and fact-checking across various benchmarks. ChartGemma's effectiveness is further proven through extensive real-world chart analysis. GitHub
Blog: A Lifecycle Management Approach toward Delivering Safe, Effective AI-enabled Health Care
in this article, we will focus on the potential of leveraging LCM to address the unique challenges of generative AI in health care, with practices to help ensure these systems meet real-world needs while managing their inherent risks across the software lifecycle.
Dreaming is All You Need
This research introduces two novel deep learning models, SleepNet and DreamNet, to strike this balance. SleepNet seamlessly integrates supervised learning with unsupervised “sleep" stages using pre-trained encoder models. Dedicated neurons within SleepNet are embedded in these unsupervised features, forming intermittent “sleep" blocks that facilitate exploratory learning. Building upon the foundation of SleepNet, DreamNet employs full encoder-decoder frameworks to reconstruct the hidden states, mimicking the human "dreaming" process.
Query languages for neural networks
We introduce a database-inspired approach for interpreting neural network models using declarative query languages. We compare black-box query languages, which only access the network’s input-output function, with white-box languages that treat the network as a weighted graph. We show that the white-box approach can subsume the black-box approach, demonstrating this for feedforward neural networks with piecewise linear activation functions.
What is CyclOps? 🛠️
The toolkit is designed to help researchers and practitioners to adopt Vector Institute’s AI trust and safety principles. Specifically, the toolkit focuses on evaluation and monitoring of AI systems developed for clinical applications where robustness, safety, and fairness are critical.The primary goal of CyclOps is to improve transparency, accountability, and trust in AI systems that are developed for clinical applications.
