A5103131863- Ethics and Social Impacts of AI
- Hate Speech and Cyberbullying Detection
- Artificial Intelligence in Healthcare and Education
- Multimodal Machine Learning Applications
- Natural Language Processing Techniques
- Semantic Web and Ontologies
- Software Engineering Research
- Health Education and Validation
- Radiomics and Machine Learning in Medical Imaging
- E-Learning and COVID-19
- Software Reliability and Analysis Research
- Ethics in Clinical Research
- Risk and Safety Analysis
- Domain Adaptation and Few-Shot Learning
- Information and Cyber Security
- Adversarial Robustness in Machine Learning
- Human Pose and Action Recognition
University of Oxford
2023
DeepMind (United Kingdom)
2022
This paper aims to help structure the risk landscape associated with large-scale Language Models (LMs). In order foster advances in responsible innovation, an in-depth understanding of potential risks posed by these models is needed. A wide range established and anticipated are analysed detail, drawing on multidisciplinary expertise literature from computer science, linguistics, social sciences. We outline six specific areas: I. Discrimination, Exclusion Toxicity, II. Information Hazards,...
Responsible innovation on large-scale Language Models (LMs) requires foresight into and in-depth understanding of the risks these models may pose. This paper develops a comprehensive taxonomy ethical social associated with LMs. We identify twenty-one risks, drawing expertise literature from computer science, linguistics, sciences. situate in our six risk areas: I. Discrimination, Hate speech Exclusion, II. Information Hazards, III. Misinformation Harms, IV. Malicious Uses, V. Human-Computer...
Current approaches to building general-purpose AI systems tend produce with both beneficial and harmful capabilities. Further progress in development could lead capabilities that pose extreme risks, such as offensive cyber or strong manipulation skills. We explain why model evaluation is critical for addressing risks. Developers must be able identify dangerous (through "dangerous capability evaluations") the propensity of models apply their harm "alignment evaluations"). These evaluations...
This paper focuses on the opportunities and ethical societal risks posed by advanced AI assistants. We define assistants as artificial agents with natural language interfaces, whose function is to plan execute sequences of actions behalf a user, across one or more domains, in line user's expectations. The starts considering technology itself, providing an overview assistants, their technical foundations potential range applications. It then explores questions around value alignment,...
The technical progression of artificial intelligence (AI) research has been built on breakthroughs in fields such as computer science, statistics, and mathematics. However, the past decade AI researchers have increasingly looked to social sciences, turning human interactions solve challenges model development. Paying crowdsourcing workers generate or curate data, data enrichment, become indispensable for many areas research, from natural language processing reinforcement learning feedback...
Safety and responsibility evaluations of advanced AI models are a critical but developing field research practice. In the development Google DeepMind's models, we innovated on applied broad set approaches to safety evaluation. this report, summarise share elements our evolving approach as well lessons learned for audience. Key include: First, theoretical underpinnings frameworks invaluable organise breadth risk domains, modalities, forms, metrics, goals. Second, theory practice evaluation...
We introduce Imagen 3, a latent diffusion model that generates high quality images from text prompts. describe our and responsibility evaluations. 3 is preferred over other state-of-the-art (SOTA) models at the time of evaluation. In addition, we discuss issues around safety representation, as well methods used to minimize potential harm models.
Fine-tuning language models has become increasingly popular following the proliferation of open and improvements in cost-effective parameter efficient fine-tuning. However, fine-tuning can influence model properties such as safety. We assess how impact different models' propensity to output toxic content. impacts Gemma, Llama, Phi on toxicity through three experiments. compare is reduced by developers during instruction-tuning. show that small amounts parameter-efficient developer-tuned via...
A shared goal of several machine learning communities like continual learning, meta-learning and transfer is to design algorithms models that efficiently robustly adapt unseen tasks. An even more ambitious build never stop adapting, become increasingly efficient through time by suitably transferring the accrued knowledge. Beyond study actual algorithm model architecture, there are hurdles towards our quest such models, as choice protocol, metric success data needed validate research...