DSS SF | Schedule

Collecting real visual data in new environments poses significant challenges, including privacy concerns, limited diversity, and the considerable effort required for data collection and annotation. With advancements in generative AI, it is now possible to create synthetic datasets that address these issues, though this approach also comes with its own challenges. This talk will cover various techniques for synthetic visual data generation, focusing on generative models. We will explore methods such as 3D rendering, procedural generation, and the fine-tuning of pre-trained models to produce high-fidelity training datasets. Attendees will gain insights into how synthetic data can be integrated into AI training pipelines to enhance model performance on downstream visual tasks, as well as the limitations and potential of these methods.

AI is probabilistic, so it inherently carries risk of mistakes. Ethical AI principles and best practices can help by empowering product builders to better uncover, assess, and mitigate risks and build AI safely. I’ll share design patterns, research methods, risk assessment tools and how they work in concert in the launch of generative AI features. We hope to inspire you to adapt the tools to your work as you responsibly design AI.

Using a Large Language Model (LLM) as a “judge” is a powerful way to assess and control the output of other AI pipelines, but principled approaches to judge management are still lacking. Judges are used to assess increasingly abstract semantic properties like hallucinations, relevance, policy compliance, and complex reasoning. However, by layering one opaque system over another, this methodology introduces new challenges, potentially compounding errors and biases. In this talk, I will examine how to avoid relying on the judge model as the ultimate arbiter of truth, and demonstrate a pragmatic framework to manage, measure, and optimize these models in production systems for the long-term. ‘EvalOps’ emerges as a principled methodology distinct from MLOps and LLMOps, for using judge models to align AI-driven processes with human norms, organizational policies, and KPIs. Attendees will gain actionable insights from practical examples and the latest research to robustly implement judge-controlled LLM automations in production environments.

This talk addresses the challenges faced by non-English users when interacting with text-to-image diffusion models and introduces a novel approach called Translation Augmented Generation. The predominance of English in training data has resulted in suboptimal performance for non-English prompts in image generation tasks, often producing culturally incongruent or irrelevant images. The authors present a prompt engineering technique that utilizes existing Large Language Models (LLMs) to detect language, translate, and augment non-English prompts with additional metadata. This method aims to enhance the quality and cultural relevance of generated images without requiring expensive model fine-tuning or building language-specific diffusion models from scratch. The article demonstrates the effectiveness of Translation Augmented Generation in improving image outputs for prompts in low-resource languages, showing how it captures both semantic meaning and cultural nuances. While acknowledging some limitations, such as increased token processing, the authors propose this technique as a promising solution to bridge the language gap in text-to-image generation, potentially democratizing access to AI-generated imagery across diverse linguistic communities.

The AI landscape is evolving rapidly, moving beyond the dominance of large language models (LLMs) like GPT-4, Gemini,Llama, Mistral towards a new frontier: multimodal AI. This roundtable discussion will explore the implications of this shift, where AI systems can process and generate not just text, but also images, audio, and video. We will delve into the state-of-the-art advancements in multimodal AI, examining how industries such as healthcare, autonomous vehicles, and creative arts are beginning to adopt these technologies. The conversation will cover the challenges of data collection, ethical considerations, and the technical hurdles associated with deploying multimodal systems at scale. Additionally, the discussion will highlight the interdisciplinary collaboration required to advance multimodal AI and consider the broader societal impact of these technologies. By focusing on the future directions of AI, this session aims to provide actionable insights and foster a deeper understanding of how AI will continue to transform industries and society in the years to come.

In this session, Jon will discuss how organizations build and scale AI operations while maintaining quality. Topics will include design considerations aligned to current trends, such as compute speed surpassing network performance, best practices for embedding, incorporating structured and unstructured data, model selection and deployment, efficient resource utilization (including expensive GPUs), workload management, and the role of integrated platforms in simplifying development, deployment and management of large-scale AI applications.

Businesses face privacy concerns, limited data availability, or unbalanced datasets, which can hinder the development of robust machine learning models. In this session I’ll examine promising solutions in in synthetic data and augmentation methods. This will include techniques for generating realistic, diverse datasets that mimic real-world data, to help enterprises overcome data scarcity while protecting sensitive information. I’ll also use case studies to explore ways to enhance existing data to improve machine learning models.

With AI systems increasingly influencing critical sectors, ensuring responsible governance is paramount. This talk dives into the technical process of converting ethical AI principles into enforceable governance frameworks. I’ll outline how to integrate bias mitigation techniques, transparency mechanisms, and accountability structures within machine learning workflows. Specifically, the session will cover the use of fairness metrics, model interpretability methods like SHAP and LIME, and automated tools for ensuring model compliance with legal and ethical standards throughout the development and post-deployment phases. Drawing from real-world examples in finance and healthcare, I will demonstrate how ethical AI policies can be operationalized, with a focus on monitoring models for fairness, enhancing auditability with explainable AI (XAI) tools, and aligning systems with industry regulations. Attendees will leave with practical strategies for engaging cross-functional teams—from data scientists to legal and compliance teams—in building robust AI governance frameworks that ensure trust, accountability, and compliance.

Abstract coming soon.

Discover how Pinterest kick-started its GenAI initiatives, democratized access across the organization for productivity use cases, and empowered teams to build and ship GenAI-powered experiences.

In the push for technological advancement, it’s crucial to remember that true progress lies in addressing the actual needs of individuals, especially those in underrepresented communities. This lightning talk explores the role of AI in enhancing accessibility, for instance through the development of improved audio descriptions for blind and low vision users. The discussion will focus on the importance of designing solutions that genuinely meet people’s needs, rather than simply advancing technology for its own sake.

This presentation explores the transformative potential of AI in User Experience Research (UXR), enabling greater efficiency and deeper insights. It provides practical guidance on leveraging AI for tasks like data analysis and user segmentation, while addressing ethical considerations such as data security, privacy, and bias mitigation. It encourages UXRs to embrace these tools for more effective and insightful research practices.

You wouldn’t dream of deploying software without monitoring or unit testing. However, this is what regularly happens with LLM applications, even though their fickleness and fragility are well known problems. While testing is often seen as a drag on a software team’s productivity, especially in fast moving organisations, the exact opposite is true for LLMs and AI agents. Great monitoring and eval can supercharge your development velocity. A good eval system can help you find the best ROI items to work on, and put together positive feedback loops in your iteration cycle, and help you find the most impactful items to work on- in the next week, month, half or year. Join me as I talk about how to put a world class eval system to work towards your LLM application.

This talk explores how multi-agent systems are pushing enterprise AI beyond simple language model interfaces. Unlike traditional single-agent approaches, multi-agent architectures offer enhanced problem-solving capabilities crucial for complex business environments The presentation will demonstrate why and how multiple AI agents, working in concert, can dynamically plan and execute actions. For instance, in healthcare, one agent might analyze patient data while another coordinates with specialists, collectively managing comprehensive care plans. Attendees will learn techniques for implementing goal-following behaviors, enabling agents to adapt to changing objectives in real-time. In education, this could mean a system persistently working towards ensuring a student fully grasps a concept, adjusting explanations, providing varied examples, and offering targeted practice until comprehension is achieved.

Incorporating Large Language Models (LLMs) into business systems presents challenges, particularly around maintaining accuracy and preventing hallucinations, especially with vast product catalogs and proprietary data. This presentation delves into the application of Retrieval-Augmented Generation (RAG) to enhance LLM outputs by leveraging vector embeddings for precise search and relevance. Attendees will explore how RAG integrates with existing architectures, improves the accuracy of LLM responses, and optimizes knowledge retrieval to deliver reliable, context-aware results in real-world applications.

Machine learning, deep learning, and generative AI will reduce redundant tasks, automate work, and, in many cases, replace the human workforce. According to the “The State of AI in 2023” report by McKinsey and Company, 40 percent of C-suite executives anticipate boosting their AI investments due to advancements in generative AI. Accenture’s 2024 article highlights a significant uptick in AI disruption within companies, with a 33% increase year over year. As a leader in AI, how do you deliver at pace? How do you ensure the products you build reach customers and drive revenue? In today’s talk, I will elaborate on my journey as a leader in AI for the past ten years, continuously delivering and reshaping the AI culture at organizations. I’ll share insights on key experiences, including (1) fostering alignment among executives and roadmaps, (2) strategic investment and outsourcing in AI infrastructure components, (3) capitalizing on the trends of Generative AI, and (4) refining the organizational structure of AI teams.

Most of the digital world has already incorporated AI/ML algorithms, websites and apps regularly feature chatbots, recommender engines, and an army of algorithms controlling every aspect of the user experience. In contrast, much of the work done with physical infrastructure and human-driven tasks is ripe for the AI automation revolution. However, the lessons learned on the digital world may not translate easily to optimizing processes and activities by putting technology in the hands of workers. AI/ML models “out in the wild” with humans interacting with them often don’t generalize as handily as models deployed into digital environments. The iterations and feedback loops are much higher than most software and product teams anticipate. In this talk, we’ll cover some of the challenges and strategies for mitigating the risks of automation out in the wild.