Have you ever had one of those days where it feels like you time-traveled into the future? That’s exactly how I felt when I attended the Data Science Salon’s conference on “Using AI & Machine Learning in the Enterprise” at Google’s San Francisco headquarters.
Being new to the city, my day started with a thrilling experience when I took my first ride in a Waymo self-driving car, and I can’t even begin to explain my feelings. It was like being in a sci-fi movie. Waymo navigated the hustling and bustling streets of San Francisco on a busy Thursday morning, and I sat there in awe, thinking, “This is the future, and I am living it.”
First things first, a huge shoutout to Anna from the Data Science Salon team who orchestrated the conference, turning an ordinary day into an extraordinary experience of learning.
Walking into the venue, I was immediately immersed in a sea of innovators, thinkers, and enthusiasts who were all gung-ho on what AI could be. The energy was contagious.
The conference kicked off with Alexis Baudron, a Senior AI Researcher at Sony, who introduced us to AITRIOS, Sony’s edge AI platform.
His talk on synthetic visual data generation was mind-blowing, especially since I don’t have a lot of experience with deep learning and text-to-image models. Alexis delved into the challenges of deploying AI on edge devices — limited processing power, software constraints, and fixed locations. Collecting data in these scenarios is labor-intensive and highly specific, posing significant hurdles for training effective AI models.
To overcome these challenges, Alexis and his team started generating their own synthetic data. However, creating ideal synthetic data that is realistic, diverse, and high-quality is no easy feat. They experimented with 3D rendering, which, while useful, often resulted in images lacking realism — more akin to video game graphics than real-world images. This led them to explore text-to-image diffusion models like Stable Diffusion, Flux, and Midjourney. These models offered new possibilities but had issues with lacking high-frequency details crucial for edge AI applications.
They tackled this by using techniques like LoRA (Low-Rank Adaptation) for fine-tuning models with less computational overhead and ControlNet to add more control to image generation. By feeding input images and prompts into a Stable Diffusion model equipped with ControlNet, they achieved more realistic and detailed synthetic data. Alexis emphasized that while challenges remain, modern techniques are starting to address the issues of realism, scalability, and manual effort in synthetic data generation. Listening to him felt like peering into the future of AI development.
Later, Aditi Godbole, a Senior Data Scientist at SAP, expanded on the theme of synthetic data in her talk about transforming enterprise data strategies.
She highlighted how synthetic data can help businesses overcome data scarcity and privacy concerns, demonstrating through examples on how realistic and diverse datasets can significantly improve machine learning models. It felt like discovering a treasure trove of untapped potential.
The importance of evaluating AI models became clear during Ari Heljakka’s session on EvalOps.
As the Founder and CEO of Root Signals, Ari introduced us to EvalOps, a framework distinct from MLOps and LLMOps, focusing specifically on the evaluation operations of AI models. This was a new arena for me, and I was fascinated by how LLM-as-a-Judge is used to assess response quality and output performance of large language models.
He explained that evaluating whether AI-generated texts comply with specific policies requires semantically rich analysis, and LLMs serve as judges to determine this compliance. This is crucial for organizations to ensure their AI outputs align with legal standards and company ethics. Intrigued, I later delved into an arXiv paper from Meta titled “Agent-as-a-Judge: Evaluate Agents with Agents,” which showed that using agents as judges for code generation aligns with human evaluator consensus about 90% of the time and is significantly more cost-effective.
Following Ari’s session, Vyoma Gajjar, an AI Technical Solution Architect at IBM, delved into the ethical dimensions of AI in her talk on shaping the future of AI governance.
She tackled tough questions about bias, transparency, and accountability, emphasizing that ethical considerations are essential for building trust and ensuring compliance in real-world applications. Vyoma’s practical examples from finance and healthcare underscored the necessity of integrating ethical frameworks into AI initiatives from the ground up.
The conference also showcased how AI is breaking new ground in understanding and processing diverse types of data. Raghavan Muthuregunathan, a Senior Engineering Manager at LinkedIn, presented a brilliant session on “Translation Augmented Generation.” He demonstrated how leveraging large language models to translate and enhance prompts can significantly improve the cultural relevance and quality of AI-generated images for non-English prompts. The simplicity of using clever prompt engineering to bridge language and cultural semantics gaps was a revelation, highlighting the power of prompt engineering in global applications.
Taking it a step further, Swagata Ashwani, Principal Data Scientist at Boomi, took us on an exciting exploration of multimodal AI.
In her talk, she discussed how AI systems are evolving to process not just text, but also images, audio, and sensor data — a development particularly impactful in healthcare. Swagata captivated us with live demos using Gemini 1.5 Flash.
She showed how AI could accurately identify fruits in a picture, calculate quantities and prices, and even help reset the clock on an oven by interpreting complex images. She also demonstrated AI’s ability to analyze videos to extract content information, identify actions, and generate relevant tags. Her presentation smartly highlighted how simple prompt engineering can bridge gaps across different data types, offering a fascinating glimpse into the future of AI.
As the sessions wound down, I glanced at my watch and felt a pang of regret. Stepping out of the vibrant atmosphere of the conference into the cool San Francisco evening, I was filled with a mix of exhilaration and anticipation.
The day had been a whirlwind of innovation, inspiration, and invaluable learning. I left with my mind buzzing with new ideas and a notebook filled with scribbles and sketches of concepts to explore further. I’m incredibly grateful to Anna, Tyler and the Data Science Salon team for the opportunity to be part of such an enriching experience.
As I made my way home, I couldn’t help but reflect on the day’s experiences — especially that unforgettable Waymo ride. It wasn’t just about the novelty of a self-driving car; it was a tangible manifestation of the technological advancements we had been discussing all day.
The insights shared at the conference painted a promising picture of where AI and machine learning are headed. Knowing that there are brilliant minds dedicated not just to advancing technology but also to ensuring it’s used ethically and responsibly gives me great hope.
I can’t wait to dive deeper into these topics and see how I can apply what I’ve learned. If this conference was any indication, we’re on the cusp of some truly exciting developments in the world of AI. Here’s to embracing the future, one innovation at a time!