Khipu 2023 Recap
Takeaways and highlights from the main AI event in Latin America
During the week of March 6–10, Montevideo was decked out to impress as it played host to Khipu 2023, the premier Artificial Intelligence conference in the region. Khipu has a vital mission of promoting progress and development in Latin American AI through student training and building a robust AI community. For an entire week, students, researchers, and professionals from across the region gathered to exchange valuable experiences and knowledge.
More than 350 attendees from 18 different nations, including 200 students from Latin America and around the world, participated in the event, making this edition of Khipu truly regional and diverse.
This week was jam-packed with inspiring lectures, tutorials, research and spotlight talks, covering the latest research topics in AI, including Generative AI, Graph Neural Networks and AI Fairness. Attendees had the privilege of listening to first-level speakers like Nando de Freitas, Samy Bengio, and Peter Norvig, who spoke at length about these and many more topics. With their impressive knowledge and expertise, this remarkable speakers lineup provided invaluable insights into the future of AI.
As if this wasn’t enough, the principal companies in the region that work in AI were there, giving to students immersing experiences and a first-hand sight of what it looks like to work on the industry developing AI systems.
At IDATHA we love collaborating with the AI community and we define ourselves by nature, as strong supporters of such honorable initiatives like this. For this reason and as we did in 2019, we supported once again this amazing cause.
After this long introduction, in this post, we will cover the most memorable moments of Khipu 2023 and share the main takeaways from this extraordinary conference, through the eyes of our lucky Khipu attendees.
Computer Vision: Past, Present and Future
Khipu caught the attention of Computer Vision practitioners on day one with a remarkable recap on this field by Jorge Sánchez: Today’s buildings blocks of visual perception. From the early Hubel and Wiesel experiment (1959) to the present Convolutional Neural Networks models and Multi-Head Attention, explained in simple words.
On day two, Ruben Villegas from Google Brain gave us an extended lecture on generative models applied for image and video generation. He started the talk mentioning early approaches like GANs, VAEs, and Normalizing Flows and concluded going deep into two state-of-the-art approaches: Diffusion Probabilistic Models and Vision Transformers. Don’t miss Ruben’s talk Generative Models to go deeper on this fascinating field.Also check out the Practicals Sessions for a hands-on experience on this area.
Also, check out Parallel Sessions on day four and Practicals Sessions for a hands-on experience in this area.
Generative Models beyond the hype
Khipu was no stranger to the generative AI boom of the past few months. Rodrigo Nogueira opened the field on day one, with a spotlight talk centered around the advantages and disadvantages of using Large Language Models for information retrieval. Minutes later Joan Bruna Associate Professor at New York University, perfectly complemented with a lecture focused on the main challenges behind building Foundations Models.
On day two, professor Kyunghyun Cho dazzles on stage with his outstanding introduction lecture on Natural language processing. In just one hour, he covered the relevance of language in human evolution, the intuitions behind language representation, and simple strategies such as the co-occurrence matrix, to Transformers and Attention. He concluded the lecture by pointing out that — Statistical patterns alone cannot distinguish between intrinsic meaning and extrinsic use. For this reason, researchers must go beyond the statistical paradigm for True language understanding — So much truth…
By the end of the third day, Nando de Freitas took the stage wearing the Uruguay soccer team jersey and led us on an extensive but very clear journey through Large Language Vision Models. Without any difficulty he overflew Transformers and Self-Attention, passing through VQ-VAE to finally end with Diffusion Models. For each of these models architectures, mentioning fantastic implementations that have woow us in the last weeks.
On day four and as part of a parallel session, Maria Lomeli who is a researcher at Meta Research, explained ATLAS: a Retrieval Augmented Language Model (RALM). Unfortunately the video of presentation is not available but you can read more in Atlas: Few-shot learning with retrieval augmented language models.
Reinforcement Learning for Deep Learners
For those curious about RL, the talk Reinforcement Learning I by Pablo Samuel Castro from Google Brain is the place to go. Pablo provided a comprehensive introduction to Reinforcement Learning, covering fundamentals such as reward and policy functions, value and policy iteration, implementation challenges, and the importance of balancing exploration and exploitation. He also mentioned valuable educational resources to delve deeper into Deep RL: an extended Introduction to reinforcement learning by himself (Khipu’s presentation is a summary), Spinning Up in Deep RL by OpenAI, and the Dopamine prototyping tool by Google. Don’t miss Pablo’s lecture for more tips!
Doina Precup, a professor at McGill University and researcher at DeepMind, continued with Reinforcement Learning II. In this lecture, Doina tackled the problem of how to do RL at Scale through Function Approximation and covered some classic implementations. Watch Doina’s presentation to dive deeper.
To see RL in practice, check the Intro to Reinforcement Learning at the practical sessions. Also, in the context of parallel sessions, several researchers complemented the RL experience, sharing their valuable knowledge and expertise on the field.
Graph Neural Networks
Another central topic at the conference was Graph Neural Networks (GNN). On day two, Alejandro Ribeiro presented several large-scale problems (such as authorship attribution and recommendation systems) typically modeled with graphs and solved using Machine Learning techniques applied to Graphs. He also pointed out that image CNNs, and RNNs for time series, share the same idea if we think of them as graph models! In other words, we can generalize the convolution concept to a graph and fit a Neural Network (the main idea of GNNs). Don’t miss Alejandro Ribeiro’s lecture.
Later, Gonzalo Mateos continued with Graph Neural Networks II: historical evolution of the field and scenarios where GNNs are used for supervised, semi-supervised and unsupervised learning. Gonzalo emphasized that to obtain a satisfactory modeling of the problem, it is essential to understand what the links in our graph mean, in terms of our original problem. Gonzalo, citing René Vidal, concluded that we started optimizing features, and now we optimize architectures. He ended the lecture by introducing state-of-the-art GNN solutions.
GNN also had a relevant spot during practical sessions with a comprehensive tutorial on the field (including theoretical and practical remarks). During the tutorial, students had the chance to model a collaborative filtering problem using Graphs and then use a GNN to predict customer reviews over products.
Ethics and AI Fairness
Reflecting on how we work, the objectives and behavior of some ML models, the role of women in AI, and the role of Latin America in AI had substantial relevance in this edition of Khipu. In particular, Sara Hooker Director at Cohere, opened the discussion on day three with an enlightening talk about Ethics and Fairness in AI: protected features bias, detecting minority groups to measure fairness, hallucinations, and misinformation of LLMs, and challenges in building fair models, such as model drifts or catastrophic forgetting.
Moving on, the practical session on the Social Impacts of Artificial Intelligence was an opportunity to see this topic in action. Particularly to explore how stereotypes encoded in Transformers models can result in discriminatory behavior. Later, the Ethics in AI Panel formed by Paola Ricaurte, Sasha Luccioni, Andrés Morales, and Laura Ación, added diverse and valuable views to the topic.
Undoubtedly the icing on the cake was a the statement on the impact of Artificial Intelligence in Latin America, signed by many at the organization Committee and then, conference attendants (read the statement here).
Practical Sessions
In addition to keynotes and research talks, Khipu provided a majestic hands-on experience (some already cited in this post). The practical sessions were led by experts in the field, covering last trends in AI and providing invaluable insights and best practices for participants and AI practitioners. Here is a summary of them:
- JAX, Optax, Haiku and CNN: Hands-on with JAX, Optax and Haiku: high-performance libraries for computation, optimization, and model learning. Explore these libraries training Convolutional Neural Networks (CNN), from scratch.
- Building ML in an Open and Collaborative Way with the Hugging Face ecosystem: Explore the Hugging Face hub and unlock the power of working collaboratively, reusing open data, publishing your datasets as open data, and building on top of the best open-source models.
- Attention and Transformers: Learn the basics behind the Transformers architecture and see in greater detail the attention mechanism. Finally, build the entire architecture block by block.
- Graph Neural Networks: Recap Graph Theory and then learn Graph Neural Networks (GNNs): how it works from a high level, some popular implementations, and how they work in practice.
- A Hands-On Forecasting Guide: From Theory to Practice: Outstanding code camp, providing invaluable insights and best practices for tackling time series forecasting problems. The tutorial was created by Tryolabs friends in collaboration with Google fellows, making it an exceptional learning opportunity.
- Intro to Reinforcement Learning: Explore various RL approaches for solving the classic CartPole, an inverted pendulum system, where an agent must learn to balance a vertical pole by displacing the cart.
- Deep Generative Models: Walk through the challenges in developing an effective generative model, specifically for Denoise Diffusion Models (a.k.a. a Score-Based Generative Model), the backbone of the recent and exciting Dalle-2 and Imagen models.
- Social Impacts of Artificial Intelligence: Learn to use different tools to assess how stereotypes can result in discriminatory behavior in language technologies and reflect on their social impacts.
Research & Spotlight Talks: Honorable Mentions
Many interesting and enlightening research works were presented during research and spotlight talks. Unfortunately, we can not mention every single presentation in this post, or this would not be a summary. For this reason, we want to make two honorable mentions and encourage you to also review these presentations and parallel sessions.
- Samy Bengio: Perhaps unnoticed, since it was one of the last Research Talks on day four, but worth every minute of your attention. Samy put together different research problems that Apple is working on both in company products and for the common good of society, with long-term expectations (3–5 years). The main topics are: Self-supervised learning, continuous pseudo-labeling for speech recognition, debiasing large language models, and learning to reason.
- Mónica Tentori: She is a Professor of Computer Science at CICESE (Baja California, México). In a Spotlight Talk, she presented two Inspiring stories about AI in the context of her research work. Spoiler alert: Children with ASD use a different force amount and perform different gesture patterns interacting with haptic interfaces. In this context, Digital Biomarkers and simple Machine learning models such as Random Forest could be a novel approach to early diagnosis of Autism with high Precision. Want to know more, watch the presentation here.
Secondary Room: Sponsors area
Another of the high points of the week was the sponsors room, where each company stood out with extremely original stands, and interactive activities to impress students and create a relaxed atmosphere to just talk and meet new people. From Machine Learning quizzes, free raffles to an interactive squat wars app that count all the squats you can do in 30 seconds.
We want to thank all the students, other sponsors and attendees in general, who came to our stand to chat, especially those who said that our stickers were amazing 🙌. ️Finally, congratulations once again to the winners of the free raffles. See you in Khipu 2024 with more raffles and new stickers designs 💪💪️
— We are nearing the end of this post, almost there! —
Women in AI
On Thursday night, Khipu hosted the Women in AI reception, an event intended to unite women in the field and foster, encourage and support greater diversity in the field of AI in the Latin American community.
With the best views of Montevideo as living background, drinks and chill music, the organizers brought together a panel of four remarkable women: Paula Martínez, Aiala Rosá, Magdalena Fuentes and Sara Hooker. This extraordinary group of women inspired attendees, speakers, and special guests with their personal career journey and experiences, to finalize with a Q&A session. In addition, the organization took the opportunity to recognize the professional trajectory of Alicia Fernández and Dina Wonsever, two local pioneers in the field and emblematic professors from UdelaR.
This evening was without a doubt, the pinnacle of the conference, filling the audience with a renewed sense of empowerment and creating a relaxed environment for engaging in stimulating conversations and cultivating new bonds.
The organizers reserved a very special surprise for the end. No Te Va Gustar, a popular Uruguayan band that sold out stadiums, surprised the audience with an intimate show. The band played several hits and completed the show by playing the iconic song No Era Cierto — My next memory may be distorted by drinks and the adrenaline (keep that in mind). I swear I saw my IDATHA colleague Jairo bravely encourage everyone in the audience to assemble a giant pogo (mosh pit). Even the conference speakers were there!! — Would It be crazy if I say then, I saw Jairo starting the nerdiest pogo in history?
(I think is good material for a t-shirt at least…)
Closing event
The beautiful Teatro Solís, an opera house from 1856 (main stage at the country), was chosen to host the closing ceremony. The event was also extended to general public (academia, industry and press) who had the privilege to hear from leading researchers and experts the latest advances and trends in AI.
Three researchers from Latam were in charge of raising the curtain. Among these presentations, Martín Rocamora stands out without a doubt, presenting Generative AI in music, candombe drum line on live included! Later Nando de Freitas, Sara Hooker and Kyunghyun Cho dazzled on stage with different presentations about recent advances in AI.
Later, Peter Norvig, a distinguished educator in the Artificial Intelligence field, researcher at Google, and co-writer of the iconic book Artificial Intelligence: A Modern Approach (one of these books that we should all read once), took the stage. In his keynote, Peter presented a software engineering journey, from the past’s logic and math-focused languages to the present LLM-based coding assistants like Github-Copilot, ruled by probabilistic and observational strategies. Coding example included!
Concluding with a week full of knowledge and experiences, the ceremony ended with a round table on How and Why We Should be Fostering AI in Latin America. The panel of experts in charge of guiding the discussion were: Jocelyn Dunstan (Assistant Professor at Universidad Católica de Chile), Fabrizio Scrollini (Executive Director at ILDA), Peter Norvig (Stanford Fellow and Researcher at Google) and Sebastian Barrios (SVP of Technology at Mercado Libre).
Did you miss it? Want to experience it again? Watch the closing ceremony here.
Final thoughts
This edition of Khipu was special and incredible in every aspect: from bringing together some of the brightest minds in the field of Artificial Intelligence, bringing students from all over Latin America, designing a week full of priceless lectures and tutorials, to the extraordinary effort to have the most amazing closing ceremony and Women in AI reception we could ever have dreamed of.
An standing ovation to the organizing committee, which really surpassed itself remarkably and gave us a first class conference.
As we conclude this journey, we say goodbye to this edition of Khipu, with joy from the bonds we have forged and the friendships we have cultivated. Filled with inspiration drawn from the fascinating career paths and personal journeys in AI we have come across. We are also very proud of the incredible work that researchers and young talents are doing in the region. Ultimately, we depart with a renewed spirit, committed more than ever to fostering Artificial Intelligence in Latin American community that is not just strong and united, but resoundingly diverse.
See you at the next Khipu!!
P.S. If you enjoyed this article, consider following me on my social media to support my work. Additionally, you’ll be notified whenever I release new content!!
This blog post wouldn’t be possible without the invaluable contributions and support of Gabriel Illanes, Felipe Chavat, Jairo Bonanata and Manu Reynaert. My delegation fellows have truly elevated the quality of this post, sharing their unique and personal Khipu experiences.
