# 11 Powerful Talks from rstudio::conf 2020 You Won‘t Want to Miss: An AI/ML Expert‘s Perspective - Canonical: https://33rdsquare.com/11-powerful-talks-rstudioconf-2020/ - Published: 2024-09-03 - Author: Jordan Brown - Categories: [Artificial Intelligence & Machine Learning & ChatGPT](https://33rdsquare.com/category/tech/ai/) --- ## Introduction For R users and data science professionals, rstudio::conf is the highlight of the year. The 2020 edition was no exception, bringing together over 1,700 attendees from more than 50 countries to explore the latest advancements in R. As an artificial intelligence and machine learning expert, I was particularly impressed by the growth and maturation of R‘s capabilities in these cutting-edge fields. From deep learning to graph neural networks to "tidy" machine learning workflows, the talks showcased how R is rapidly evolving to meet the demands of modern AI/ML projects. At the same time, the conference highlighted R‘s traditional strengths in areas like data visualization, statistical analysis, and data communication. Attendees were treated to inspiring case studies and demos that pushed the boundaries of what‘s possible with packages like ggplot2, shiny, and rmarkdown. After reviewing all the recordings, I‘ve selected 11 standout talks that I believe every R user and data science practitioner should watch. These presentations offer valuable insights and practical tips for anyone looking to level up their skills and stay abreast of the latest trends in AI, ML, and beyond. ## Data Visualization R has long been renowned for its powerful data visualization capabilities, and this year‘s conference showcased a range of innovative new approaches and packages. ### Effective Visualizations for Research by Miriah Meyer Miriah Meyer, a computer science professor at the University of Utah, presented a framework for designing visualizations that specifically address key research questions. Drawing on her experience collaborating with domain experts in fields like genomics and physics, Meyer stressed the importance of a question-driven approach that prioritizes iterative feedback and testing. To illustrate the framework, Meyer walked through several real-world examples of visualizations her team created for complex scientific datasets. She highlighted how techniques like sketching, wireframing, and rapid prototyping helped them quickly explore and refine design ideas. One particularly impactful visual showed how Meyer‘s team used interactive linking and brushing to help physicists identify key patterns in particle accelerator data. By allowing researchers to dynamically filter and highlight data points across multiple coordinated views, the tool enabled them to make new discoveries and generate hypotheses more efficiently. Throughout the talk, Meyer grounded her advice in established principles from cognitive psychology, graphic design, and visual perception theory. She also provided an extensive list of recommended resources, including classic books like Edward Tufte‘s "The Visual Display of Quantitative Information" and cutting-edge research papers on visualization techniques. As an AI/ML expert, I found Meyer‘s emphasis on domain-specific questions and iterative design highly relevant to my own work. Too often, I‘ve seen data scientists (myself included) get caught up in the technical details of modeling and overlook the importance of communication and visualization. By starting with clear research questions and collaborating closely with subject matter experts, we can ensure that our AI/ML projects generate meaningful insights and drive real-world impact. ### 3D Mapping and Visualization with Rayshader by Tyler Morgan-Wall While R has no shortage of mapping and GIS packages, rayshader stands out for its ability to create stunning 3D visualizations with relative ease. Tyler Morgan-Wall, a researcher at the Institute for Defense Analyses, demonstrated how the package can be used to generate interactive maps, animations, and even virtual reality scenes. The core functionality of rayshader lies in its ability to drape 2D images and maps over 3D elevation data. Morgan-Wall showed how this technique can be used to create striking visuals like topographic maps, city skylines, and natural landscapes. But rayshader‘s capabilities go beyond just mapping. Morgan-Wall demonstrated how the package can be used to visualize arbitrary 3D data, such as brain MRI scans, particle simulations, and even mathematical functions. By leveraging R‘s built-in support for matrix operations and linear algebra, rayshader makes it easy to transform and manipulate 3D data programmatically. One of the most impressive demos showcased rayshader‘s ability to generate real-time animations and fly-throughs. Using the package‘s built-in camera and lighting controls, Morgan-Wall created a cinematic tour of a mountainous landscape that had the audience gasping in awe. Throughout the talk, Morgan-Wall emphasized the importance of aesthetic considerations like color scales, shading, and perspective. He argued that well-crafted visuals are not only more engaging for audiences but can also communicate complex data more effectively by highlighting key patterns and relationships. From an AI/ML perspective, I see tremendous potential for packages like rayshader to enhance the interpretability and explanatory power of machine learning models. By visualizing high-dimensional datasets, decision boundaries, and feature spaces in intuitive 3D formats, we can better understand how models are making predictions and identify areas for improvement. Additionally, the ability to generate interactive and animated visualizations could enable richer communication of AI/ML results to non-technical stakeholders. ## Deep Learning While Python may be the go-to language for deep learning these days, R has made significant strides in recent years thanks to the keras and tensorflow packages. ### Deep Learning with R by Paige Bailey In her talk, Paige Bailey, TensorFlow‘s product manager, made a compelling case for using R for deep learning projects. She argued that R‘s simple syntax, unifying philosophy, and rich package ecosystem make it an ideal choice for data scientists who are already comfortable with the language. To demonstrate R‘s deep learning capabilities, Bailey walked through a complete example of building an image classification model using the keras package. She showed how to load and preprocess data, define a model architecture, train the model, and evaluate its performance using standard metrics like accuracy and loss. One of the key advantages of using keras in R, Bailey noted, is its tight integration with the tidyverse. By leveraging dplyr and tidyr functions for data manipulation, users can create clean, readable code that is easy to maintain and debug. She also demonstrated how to use ggplot2 to visualize training progress and model performance. Throughout the talk, Bailey emphasized the importance of data preprocessing and feature engineering. She showed how simple techniques like normalization, one-hot encoding, and data augmentation can significantly improve model accuracy and generalization. While acknowledging the strengths of Python‘s deep learning ecosystem, Bailey argued that R‘s simplicity and ease of use make it a viable alternative for many projects. She also highlighted the growing community of R users and developers contributing to packages like keras, tensorflow, and torch. As someone who has used both Python and R for deep learning, I appreciate the effort to make these powerful tools more accessible to R users. While Python may still have an edge in terms of raw performance and scalability, I believe that R‘s strengths in data exploration, visualization, and statistical modeling make it a valuable complement to Python in the deep learning workflow. ### Deep Learning on Graphs and Knowledge Representation by Daniel Falbel Daniel Falbel‘s talk on graph neural networks (GNNs) showcased one of the most exciting frontiers in deep learning research. GNNs are a class of models designed to learn from non-Euclidean data structures like graphs and networks, which are ubiquitous in fields like social network analysis, recommender systems, and computational biology. Falbel, the co-founder of Tidy Models, began by motivating the need for GNNs and explaining how they differ from traditional neural network architectures. He then introduced the key concepts behind GNNs, such as message passing and graph convolutions, using intuitive visualizations and analogies. To demonstrate how to implement GNNs in R, Falbel walked through a complete example using the torch and tidygraphs packages. He showed how to load and preprocess graph data, define a GNN architecture using torch‘s dynamic computational graph, and train the model using standard backpropagation techniques. One of the key advantages of using R for GNNs, Falbel argued, is the ability to leverage the language‘s rich ecosystem of graph and network analysis packages. By combining GNNs with tools like igraph, tidygraph, and ggraph, users can create powerful pipelines for tasks like node classification, link prediction, and community detection. Throughout the talk, Falbel emphasized the importance of model interpretability and visualization. He demonstrated how to use packages like visNetwork and gganimate to create interactive visualizations of graph data and GNN activations, which can help users gain insights into how the model is making predictions. As an AI/ML practitioner, I see GNNs as one of the most promising areas of research in deep learning. Many real-world problems involve complex, interconnected systems that are difficult to model using traditional tabular or sequential data formats. By enabling deep learning on graphs and networks, GNNs open up new possibilities for AI applications in fields like drug discovery, material science, and social network analysis. While the R ecosystem for GNNs is still relatively young compared to Python, I‘m excited to see the progress being made by packages like torch and tidygraphs. As more R users become aware of the potential of GNNs, I expect to see even more innovation and growth in this area in the coming years. ## Tidy Modeling and Machine Learning One of the key themes of the conference was the importance of "tidy" workflows for modeling and machine learning. Several talks showcased new packages and approaches for streamlining the process of building, evaluating, and deploying models in R. ### Confident Data Manipulation with dplyr by Mara Averick Mara Averick, a software engineer at RStudio, gave a highly practical talk on using the dplyr package for data manipulation. While dplyr is already a staple of the tidyverse, Averick argued that many users are not taking full advantage of its capabilities. To illustrate the power of dplyr, Averick walked through several real-world examples of data cleaning, filtering, and aggregation tasks. She demonstrated how to use functions like `across()`, `case_when()`, and `summarize()` to create concise, readable code that is easy to understand and modify. One of the key advantages of dplyr, Averick noted, is its ability to work with a wide range of data sources, including data frames, databases, and Spark clusters. She showed how to use the dbplyr package to translate dplyr code into SQL queries, enabling users to work with large datasets that don‘t fit in memory. Averick also highlighted the new tidylog package, which provides more informative error messages and output for dplyr operations. By displaying the intermediate steps of a pipeline and highlighting potential issues, tidylog can help users debug and optimize their code more efficiently. From an AI/ML perspective, I believe that tools like dplyr are essential for creating reproducible, scalable workflows. By providing a consistent, expressive syntax for data manipulation, dplyr enables users to focus on the high-level logic of their analysis rather than getting bogged down in low-level details. This is particularly important in machine learning projects, where data preprocessing and feature engineering can often take up the majority of development time. ### Predictive Modeling with text using tidy data principles by Julia Silge Julia Silge, a data scientist at Stack Overflow, gave a talk on using tidy data principles for text analysis and predictive modeling. She argued that by representing text data in a tidy format, with one token per row, users can leverage the full power of the tidyverse for tasks like sentiment analysis, topic modeling, and document classification. To demonstrate this approach, Silge walked through a case study of building a spam classifier for Stack Overflow comments. She showed how to use the tidytext package to preprocess the text data, creating a tidy data frame with columns for words, document IDs, and other metadata. Silge then demonstrated how to use dplyr and tidyr functions to filter, transform, and aggregate the text data into a format suitable for modeling. She showed how to calculate word frequencies, TF-IDF scores, and other features that capture the semantic content of the documents. Finally, Silge trained a logistic regression model using the glmnet package and evaluated its performance using standard metrics like accuracy, precision, and recall. Throughout the process, she emphasized the benefits of the tidy approach, which allowed her to use familiar tools like ggplot2 for data visualization and model diagnostics. As an AI/ML expert, I found Silge‘s talk highly relevant to my own work in natural language processing (NLP) and text mining. While Python has long been the go-to language for NLP, I believe that R‘s tidy data philosophy and rich ecosystem of text analysis packages make it a compelling alternative. By representing text data in a tidy format, users can take advantage of R‘s powerful data manipulation and visualization tools to explore patterns, test hypotheses, and build predictive models. This can lead to more efficient and effective text analysis workflows, particularly for users who are already familiar with the tidyverse. Moreover, I appreciate Silge‘s emphasis on interpretability and transparency in the modeling process. By using simple, explainable models like logistic regression and providing clear code examples, she demonstrates how to create models that are not only accurate but also understandable and trustworthy. ## Conclusion Attending rstudio::conf 2020 was a fantastic experience that left me inspired and energized about the future of R and data science. The talks showcased the incredible creativity, passion, and expertise of the R community, and demonstrated how the language is evolving to meet the challenges of modern AI/ML workflows. From cutting-edge deep learning techniques to elegant data visualization frameworks to tidy modeling approaches, the conference had something for everyone. As an AI/ML expert, I was particularly impressed by the growth and maturation of R‘s capabilities in areas like keras, torch, and tidymodels. At the same time, I appreciated the emphasis on core data science skills like data wrangling, visualization, and communication. The talks by Mara Averick, Julia Silge, and others demonstrated how mastering these fundamentals can lead to more effective and impactful AI/ML projects. Overall, I would highly recommend watching the recordings of these talks to anyone interested in staying up-to-date with the latest trends and best practices in R and data science. Whether you‘re a seasoned expert or a newcomer to the field, you‘re sure to learn something new and valuable. But perhaps more importantly, I encourage you to get involved in the R community and contribute your own knowledge and perspectives. As David Robinson noted in his keynote, the true power of R lies not just in the language itself but in the people who use and develop it. By sharing our work, collaborating with others, and building upon each other‘s ideas, we can push the boundaries of what‘s possible with data science and AI/ML. So here‘s to another incredible year of growth, discovery, and innovation in the R community. I can‘t wait to see what breakthroughs and achievements the future holds. --- Source: [11 Powerful Talks from rstudio::conf 2020 You Won‘t Want to Miss: An AI/ML Expert‘s Perspective](https://33rdsquare.com/11-powerful-talks-rstudioconf-2020/)