Aveedibya Dey

❝ How do you taste your Wine? With the right Glass, served from the right Bottle. How do you utilize your LLM? With the right Task, served from the right Compute environment. ❞ If the quote above caught your interest, check out my blog post that was featured on the Toyota Connected India Medium channel last week: Leveraging LLMs in Cloud - 🔗 https://lnkd.in/gfHsdevH --- I hope this blog adds a bit of amusement for the seasoned AI/ML engineers 🧠👨‍💻 with the carefully crafted images and perhaps even the silly wine analogies 😉. Whether you're a cloud engineer ☁️ 👨‍💻, a product owner of cloud-based apps, or someone new to Generative AI in the cloud, this blog offers a detailed, beginner-friendly look into the selection and deployment of LLMs in cloud environments. --- A heartfelt thanks 🙏 to my colleagues at Toyota Connected India for taking the time to review, shape this piece and even share a creative writeup like below in the original post 🫡 👇 🤔 Curious to dive in? Click the link below to explore the blog on Medium! 🌐 🔗 https://lnkd.in/gfHsdevH

2

Tour of Reinforcement Learning and Stochastic Optimization Chapter 9 – Modeling sequential decision problems Chapter 9 is available as a free download from the webpage https://lnkd.in/dB99tHtM   This is Part II of my presentation of Chapter 9: Advanced concepts    9.10 Base models and lookahead models – Easily one of the most overlooked concepts in stochastic models. The base model is how we test our policy, written as a sequence of state (what we know), decision (what we choose to do) and information (that is learned after we make a decision). Normally this is a simulator, but it might be the real world.   The lookahead model is an approximation of the base model, so I put tilde’s on all the variables, along with a double time index. This notation is critical in chapter 19.   9.11 A classification of problems – It is helpful to organize sequential decision problems along two dimensions (see table below):   o Whether the problem itself depends on dynamic information captured in the state variable, or not. The dynamic information may or may not be influenced by decisions.   o Whether the objective is to optimize the final reward (performance after a series of learning iterations) or cumulative reward.   This produces the four problem classes in the table below, which are labeled 1 to 4, 9.12 Policy evaluation – It is one thing to write out the table below, but it is an altogether different task to compute the objectives, even approximately. In this section I describe how to estimate the objectives using simulation. Pay special attention to class 4 – state-dependent problems, final reward.   9.13 Advanced probabilistic modeling concepts – My book demonstrates that you can do advanced, real-world modeling and computation without ever needing to know what a “sigma-algebra” is. But for the readers who would like to know what this means, this section explains it, with a numerical example!   9.14 This section provides a roadmap of the rest of the book.

80

1 Comment

Measuring the performance of new language models on domain-specific data remains a significant challenge, particularly due to the lack of annotated data for natural language generation tasks. Instead of getting caught up in the constant one-upmanship among tech giants, we should shift our focus and the goal should be to develop scalable generative AI solutions that are robust, and the prompts are LLM agnostic. By prioritizing the creation of versatile, model-agnostic prompts, one can enhance the utility and applicability of Gen AI across various domains.

2

A very useful checklist to follow when you have to put in production Transformer based models

5

Tour of Reinforcement Learning and Stochastic Optimization (RLSO)   I am going to do a chapter-by-chapter summary of my “big book”: Reinforcement Learning and Stochastic Optimization: A unified framework for sequential decisions Each post will cover one chapter. Brief summaries of each chapter are contained on the book webpage at https://lnkd.in/dB99tHtM (“tinyurl.com/” with “RlandSO”) so in these posts, I am going to offer my thoughts on what is significant and different in each chapter.   I will eventually make the entire sequence of chapter summaries available on the RLSO webpage.   Chapter 1 – Sequential Decision Problems I start by introducing “sequential decision problems” along with a list of examples spanning business, science, engineering, economics, health, finance, transportation, supply chain management, and even the design of algorithms.   I then introduce the universal modeling framework consisting of five elements (state variables, decision variables, exogenous information, transition function and objective function), after which I give the four classes of policies for making decisions – the first book to do so.    Next, I discuss the roles of statistical learning in the modeling and solution of sequential decision problems, and discuss the relationship between these two important and complementary fields, which are distinctly different.   I show how to make the transition from a deterministic model of a time-dependent problem to a fully sequential, stochastic, dynamic optimization problem. The key distinction is that deterministic problems optimize over “x,” while sequential decision problems optimize over functions called policies. This parallels machine learning where the (often unstated) optimization problem is also optimizing over functions (statistical models) to find the best fit.   As I close the chapter, I talk about two major classes of sequential decision problems:   o Pure learning problems, which come in two flavors: derivative-based stochastic search and derivative-free stochastic search. Pure learning problems feature a state variable that only captures what we know about a function. The function itself is typically static. Pure learning problems are covered in chapters 5, 6 and 7.  o General problems that typically include either or both of a physical state (how much inventory, where something is located) and an information state (prices, weather, market information). The state variable may or may not include a belief state.

165

2 Comments

Exploring the Future of BVH Construction with H-PLOC "H-PLOC: Hierarchical Parallel Locally-Ordered Clustering for Bounding Volume Hierarchy Construction," recently published in the Proceedings of the ACM on Computer Graphics and Interactive Techniques. 🚀 Authored by a talented team of researchers, this paper introduces a groundbreaking GPU-oriented approach for binary BVH construction using the parallel locally-ordered clustering (PLOC/PLOC++) algorithm. Here are some key highlights: - Enhanced BVH Quality: The novel algorithm ensures high-quality BVH with just a single kernel launch. - Reduced Construction Times: It achieves 1.1-3.6× lower times for entire BVH builds and 1.6-13× lower times for the binary BVH phase alone. - Efficient Conversion Algorithm: Presents a method for converting binary BVH to =-wide BVH in a single kernel launch, making it seamlessly integrable into modern rendering frameworks. Congratulations to the AMD's team for their remarkable contribution to the field: Carsten Benthin, Advanced Micro Devices, Inc., Germany Daniel Meister, Advanced Micro Devices, Inc., Japan Josh Barczak, Advanced Micro Devices, Inc., USA Rohan M., Advanced Micro Devices, Inc., USA John Tsakok , Advanced Micro Devices, Inc., USA Andrew Kensler , Advanced Micro Devices, Inc., USA

39

1 Comment