Antonio Bibiano

When invited to the Super Data Science Podcast, Ritchie Vink, author of Polars, mentioned an upcoming Polars book set to launch in January 2025: "Python Polars: The Definitive Guide" by Jeroen Janssens and Thijs Nieuwdorp. The official Github repository already has a lot of code examples that worth checking out: https://lnkd.in/dJGbwQnM The Early Release version can also be read on O'Reilly Learning platform: https://lnkd.in/daTEf8Tv The book is offering guidance for you to:  - Process larger-than-memory datasets at record speed - Apply the eager, lazy, and streaming APIs of Polars and decide when to use which - Transition smoothly from Pandas or Spark to Polars - Integrate Polars into your existing codebase - Work with Arrow and Parquet to efficiently read and write data - Translate complex ETL tasks into efficient and elegant queries I started reading the Early Release version and I am enjoying it so far. I am looking forward to the final version of the book. #Python #DataFrameLibrary #Polars

40

2 commentaren

In a neural network training iteration, there are three main steps: forward propagation, backward propagation, and optimizer updates. Experiments show that the time taken for backward propagation is almost twice that of forward propagation, while the time spent on optimizer updates is negligible. This phenomenon can be explained by analyzing the floating-point operations (FLOPs) involved. Differences between Backward and Forward Propagation: 1. Forward propagation mainly calculates activations and the loss function by performing matrix multiplications to compute connections between neurons. 2. Backward propagation not only computes the gradients of the loss function with respect to weights and biases but also propagates the error from the output layer back to the hidden layers. This results in additional matrix multiplications and additions, making the FLOPs for backward propagation approximately twice that of forward propagation. Computational Complexity: - The primary overhead in forward propagation comes from matrix multiplication, with the FLOPs roughly calculated as 2*batch size * h_(l-1) * h_l. - Backward propagation involves error calculation and weight gradient computation, requiring extra matrix multiplications. Thus, the FLOPs for backward propagation are double that of forward propagation. The Impact of Batch Size and Network Depth on Computational Complexity: 1. Batch size increases cause the FLOPs for both forward and backward propagation to increase linearly, while the FLOPs for optimizer updates remain constant. Therefore, with larger batch sizes, the time taken by optimizer updates becomes negligible. 2. Network depth increases cause the backward-forward FLOPs ratio to approach 2:1, as intermediate and later layers dominate the computational load, while the first layer’s impact becomes minimal. Conclusion: In the training of large, deep neural networks, the FLOPs for backward propagation are approximately twice those of forward propagation, while the optimizer updates take a very small portion of time. As a result, in deep networks and large batch training, backward propagation becomes the most time-consuming part of the training process.

2

My colleague Jeroen Overschie made me finally understand how you can feed relevant documents into a large language model to enhance its context and provide better information and how important a good, relevant search engine still is for this method. In short: if your search engine and the results it turns up are better, the quality of answers from the LLM will be better.

9

🚀 uv (Python) enabling new workflows! Last night I had an idea: what if we could open any Jupyter notebook from GitHub with all its dependencies installed instantly? The result 2 hours later: opennb (https://lnkd.in/gGZ6CDPN) Example: uvx --with "pipefunc[docs]" opennb pipefunc/pipefunc/example.ipynb This single command: ✨ Creates a temporary environment 📦 Installs dependencies (instant with uv's cache!) 📓 Downloads and opens the notebook What made this possible? The amazing uv tool (https://docs.astral.sh/uv/), which handles Python package management at insane speeds. The best ideas are often the simplest - combining uv's speed with Jupyter's accessibility to make trying out Python tutorials friction-free. #Python #JupyterNotebook #OpenSource #DeveloperTools #Programming

92

5 commentaren

in Pandas 2.2 one can set pyarrow as the default backend. in this presentation the advantages of arrow vs numpy are presented: 1. Faster string operations 2. Consistent missing value handling 3. Efficient data serialization Minute 24 https://lnkd.in/gbbGpR56

22

The lowest hanging fruit for mobilising LLM's in production is handling outputs via data structures and controlling outputs with validators. Jason Liu explains this well as he demonstrates that classic software engineering tools are being backfit to LLM development. AI at scale is on its way - keep it simple. https://lnkd.in/gAitudHR #mlops #llmops #genai

49

4 commentaren

🚀 Groundbreaking advancements in the field of AI and language models! 🚀 𝐁𝐢𝐭𝐮𝐧𝐞: 𝐁𝐢𝐝𝐢𝐫𝐞𝐜𝐭𝐢𝐨𝐧𝐚𝐥 𝐈𝐧𝐬𝐭𝐫𝐮𝐜𝐭𝐢𝐨𝐧-𝐓𝐮𝐧𝐢𝐧𝐠 - a novel approach that dramatically enhances instruction-tuning of pretrained decoder-only large language models (LLMs). Our team, comprising Dawid J. Kopiczko (Vrije University of Amsterdam Amsterdam), Tijmen Blankevoort (Meta), and Yuki M. Asano (University of Amsterdam), developed Bitune to integrate both causal and bidirectional attention, achieving remarkable improvements in various downstream tasks. 🔍 𝐊𝐞𝐲 𝐈𝐧𝐬𝐢𝐠𝐡𝐭𝐬: Bitune leverages two sets of parameters: one for causal and another for bidirectional attention. These features are combined using a weighted average with trainable coefficients to generate new tokens. Our method introduces only a minimal set of new parameters through parameter-efficient fine-tuning techniques (PEFT), making it highly efficient. 📊 𝐌𝐚𝐭𝐡𝐞𝐦𝐚𝐭𝐢𝐜𝐚𝐥𝐥𝐲 𝐈𝐧𝐭𝐞𝐫𝐞𝐬𝐭𝐢𝐧𝐠 𝐅𝐢𝐧𝐝𝐢𝐧𝐠𝐬: During the prefilling phase, Bitune processes the instruction using both causal and bidirectional attention. This leads to an enhanced representation of the query. This method preserves the autoregressive nature of response generation while significantly improving performance on tasks such as commonsense reasoning, arithmetic, and language understanding. 📈 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 𝐁𝐨𝐨𝐬𝐭𝐬: Consistent zero-shot improvements across models like Gemma-2B, Llama2-7B, and Phi-2. Significant gains in commonsense reasoning and arithmetic tasks, demonstrating the versatility and robustness of Bitune. 🏆 𝐖𝐡𝐲 𝐁𝐢𝐭𝐮𝐧𝐞 (𝐁𝐋𝐒𝐓𝐌) 𝐢𝐬 𝐒𝐮𝐩𝐞𝐫𝐢𝐨𝐫 𝐭𝐨 𝐄𝐧𝐜𝐨𝐝𝐞𝐫-𝐎𝐧𝐥𝐲 𝐌𝐨𝐝𝐞𝐥𝐬: While encoder-only models like 𝐁𝐄𝐑𝐓 use bidirectional attention, they lack the autoregressive capabilities crucial for generative tasks. Bitune combines the strengths of bidirectional attention for richer context understanding with the generative prowess of decoder-only models, resulting in superior performance. This is a leap forward in making LLMs more effective and efficient in real-world applications, from general-purpose assistants to medical diagnostics and beyond. Excited to see the impact of Bitune in the AI community!

16

1 commentaar

Explore José Antonio Siqueira de Cerqueira's efforts in enhancing language models for greater safety and reliability. Dive into this article to discover his significant contributions to the field - https://lnkd.in/dA-EWMu2! 📚👨‍💻 #research #languagemodels

4

Closing thoughts on PyData Eindhoven, data scientists have a very good understanding regarding good uses cases for Generative AI, and, as well, use cases that don't really fit Generative AI and are better served by other approaches. For those use cases, the discussion is about monitoring, quality control and management of AI applications. My talk was the first talk in the back-up list, however, none of the titular roster speakers needed to excuse him/her self, so it would be next year. For those wondering what my talk was to be about, check the following repo :) https://lnkd.in/ehwC7h4H

12

LlamaIndex + MLflow RAG contains a lot of parameters to tune, from chunking to indexing - everything affects downstream answer accuracy. It’s important to have a systematic approach to track and tune these parameters, and define the right eval metrics/dataset to evaluate changes. Check out Jino Rohit's video on experimenting with a LlamaIndex pipeline with MLflow for experiment tracking. https://lnkd.in/gyYXfyyb

554

14 commentaren

Our PyData NYC tutorial with Christopher Fonnesbeck on Mastering #GaussianProcesses with PyMC is now online 🥳 If you missed it live and want to learn more about GPs in practice, come check it out! We cover: 1. Introduction to Probabilistic Modeling (10 min)   - Overview of modeling complex systems using Gaussian distributions. 2. What is a Gaussian Process? (15 min)   - Explanation of Gaussian Processes, their features, and properties.   - Discussion of the flexibility of GPs in capturing non-linear relationships. 3. Building a Gaussian Process Model in PyMC (10 min)   - How to choose and customize covariance functions to suit different problems. 4. Marginal, Latent, and Sparse GPs (10 min)   - Different classes of GPs and their properties. 5. Fast Gaussian Processes: HSGP Approximation (10 min)   - Introduction to the HSGP approximation and how it scales GPs to large datasets. 6. Sports Analytics Case Study (40 min)   - Application of GPs to a real-world sports analytics problem.   - Step-by-step walkthrough of modeling and prediction using GPs in PyMC. Feel free to comment below with questions, editions and suggestions! Enjoy & PyMCheers 🖖 Video: https://lnkd.in/dPee5PYp GitHub repo: https://lnkd.in/ddK28JDA

57

1 commentaar

Tired of keyword searches that miss the mark? 🔎 Imagine being able to ask questions of your data and get truly insightful answers. That's the power of GraphRAG, a new approach from Microsoft that combines the best of knowledge graphs and large language models. Think of it like this: GraphRAG doesn't just find documents with matching words, it actually understands the relationships between concepts and gives you answers with real depth and context. 🤯

2

Two of my favorite things: machine learning and turbulence. Interesting read by Rik, Wouter, and Daan Crommelin. "Fully resolving turbulent flows is computationally expensive. Therefore, we typically work with under resolved simulations with added turbulence model. Recently, Neural Networks are being used as turbulence model with as target to reproduce the fully resolved flow field. This is a high dimensional problem. However, scientists are often only interested in a few quantities which are derived from these flow fields, such as the average energy. We propose a simple method which only uses these quantities as "target", resulting in a much lower dimensional problem. We compare this method with convolutional neural networks and find that both can successfully correct long-term simulations." Read here: https://lnkd.in/etBe9BnX Aside, if you are interested in this overlap of CFD/AI and want to get started in your learning journey -> checkout my book I have written on this topic: https://geni.us/JaXp #turbulence #ai #machinelearning #datascience #aerospace #aerodynamics #cfd #cae

95

1 commentaar

That's awesome! Mistral's new features for building custom LLMs sound incredibly developer-friendly. The ability to fine-tune and create custom agents makes it a powerful tool for personalized AI solutions. Thanks for sharing the video link—can't wait to check it out! 🚀🔍 #CustomLLMs #AIInnovation #DeveloperTools #MachineLearning #MistralLLM

1

I had a really great time at PyData Amsterdam. Learned some good stuff about polars, very interesting library. A ton of interesting talks about causal estimation, drift detection, and CBPE. Breeze had a super interesting talk about algorithmic bias, highly relevant! The talk about regret-avoidance was really interesting, got me thinking about model-chains and the difficulties and feedback-loops those can cause. Game-theory and machine learning is a really interesting combination. And naturally there were a lot of talks about LLMs, always interesting to learn new LLM techniques and applications. The European AI act was talked about a lot, a super relevant topic for a lot of people at work right now. The legislation strikes a decent balance but is very vague in a lot of its definitions. It will be interesting to see how it's going to play out. One of the best talks was one of the technically simpler ones. Françoise Provencher talked about about measuring cities. Her lessons about good metrics were very universal, and so clearly talked about that they just burned themselves in my mind. There's an art to presenting simple important things well, and she has clearly mastered that. There were some inspirational talks about using data science for social good by the CorrelAid people. Inspired me to actually try their organization out, hoping I can find the hours to pick up some projects. One of the nicer moments was spent talking to the Neo4J guys. Told them that, while it was interesting technology, I personally used a relational database because it was a bit more practical for our use case. Was called a list-ist 😂 . Nerd holy wars are always fun. Good to see people being passionate about their product and vision. The venue was amazing, it was one of the best-organized conferences I've been to, the weather was great, the food was great, and it was good meeting some former colleagues and old acquaintances again. Certainly something to repeat next year.

24

2 commentaren

"We introduce phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such as Mixtral 8x7B and GPT-3.5 (e.g., phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench), despite being small enough to be deployed on a phone. The innovation lies entirely in our dataset for training, a scaled-up version of the one used for phi-2, composed of heavily filtered web data and synthetic data. The model is also further aligned for robustness, safety, and chat format. We also provide some initial parameter-scaling results with a 7B and 14B models trained for 4.8T tokens, called phi-3-small and phi-3-medium, both significantly more capable than phi-3-mini (e.g., respectively 75% and 78% on MMLU, and 8.7 and 8.9 on MT-bench)." https://lnkd.in/g9mSgaRC).

7

One of our recent AI projects involves wrangling arbitrary, unstructured data into clean, defined categories. The core challenge? LLMs hallucinating categories (among other things!). On the latest from the Dev Blog, Dan Girellini shares an approach using dynamic Pydantic models to enforce flexible, structured output from LLMs. This technique offers a robust way to control AI outputs while accommodating varying scenarios.

7