Montevideo once again played host to the premier Artificial Intelligence event in Latin America: Khipu 2023. As a driving force in AI innovation and collaboration, this conference aimed to empower the region through community building, advanced training, and the promotion of AI's potential to enrich lives across Latin America.

From March 6-10, Khipu brought together 200 students from 13 countries, with attendees hailing from a total of 18 nations, making it a truly global affair. Throughout the week, a whirlwind of enlightening talks, hands-on training sessions, and dynamic interactive experiences captivated the audience, immersing them in the ever-evolving world of AI.

In this blog post, we will walk you through the most memorable moments of Khipu 2023 and share the main takeaways that are shaping the future of AI.

Computer Vision horizon

Computer Vision consistently captures the spotlight at AI conferences, and Khipu 2023 was no exception. The recent surge in generative models, including groundbreaking advancements such as DALL-E-2 and Stable Diffusion, piqued everyone's anticipation for the computer vision talks at the conference.

The excitement began on Monday, when Jorge Sánchez delivered an insightful presentation on today's building blocks for visual perception. He provided a comprehensive and detailed overview of CNNs and transformer architectures applied to images, setting the stage for the transformative role of transformers throughout the conference.

On Tuesday, Ruben Villegas’s lecture on generative models for images and video introduced attendees to the widely-adopted diffusion models and their transformer-based counterparts. Be prepared for a deep dive into mathematics and architecture when reviewing his talk!

The conference also featured numerous applied research talks, showcasing how researchers are leveraging computer vision techniques to tackle real-world problems. Topics covered a wide range of applications, from object detection and its evaluation metrics to sketch-based image retrieval and breast cancer detection using images. The challenges of assembling high-quality datasets in the medical field were also addressed, underscoring the practical considerations in advancing computer vision research.

The potential of Graph Neural Networks

Over the past few years, there has been a significant rise in the use of Graph Neural Networks (GNNs) for addressing a wide array of challenges across various industries. This growing prominence of GNNs was reflected in Khipu's schedule, which featured two lectures and a practical session dedicated to this innovative approach.

The talks complemented each other perfectly. Alejandro Ribeiro began by introducing the concept of graphs and explaining how data can sometimes be directly represented using graph structures, while in other cases, additional efforts are required to adapt the data. He also covered the fundamental building blocks of GNNs, delving into graph filters and their properties.

Gonzalo Mateos, on the other hand, concentrated on the diverse properties and applications of GNNs, including their use in recommender systems and the detection of polypharmacy side effects. By watching these presentations, you will gain a comprehensive understanding of the field and learn how to employ GNNs for solving real-world problems effectively.

Unleashing creativity: Generative Models in focus

Generative models, especially those capable of producing text and images, have become a focal point in the AI community. It's no surprise that Khipu 2023 showcased a series of captivating talks on this cutting-edge subject.

Kyunghyun Cho, a distinguished researcher in NLP and genomics from New York University (NYU), delivered an insightful talk on classical methods for NLP and an intuitive introduction to the transformer architecture. Sporting the Uruguayan football team t-shirt he received at Khipu 2019, he highlighted some of the challenges facing modern Language Models and suggested that novel algorithms, which go beyond current data-centric approaches, are necessary to address these issues.

Keeping it up with cherished acquaintances from Khipu 2019, we were thrilled to welcome back Nando de Freitas as a speaker at this year’s conference. In his talk, he captivated the audience with an engaging and accessible exploration of modern transformer models, delving into their applications for generating text, images, and videos, while keeping attendees informed on the latest trends.

Joan Bruna, another renowned researcher and professor at NYU, illuminated the most pressing problems of Deep Learning models and emphasized the importance of establishing a solid theoretical foundation to better understand and address these model behaviors and limitations.

Finally, Maria Lomeli, researcher at Meta, introduced Atlas, a Retrieval-Augmented Language Model. This innovative model not only generates high-quality responses to text prompts but also references the training samples that informed its response. For those interested in learning more about Atlas, Maria recommended this Atlas blog post as a valuable resource.

(Deep) Reinforcement Learning advances

Reinforcement Learning (RL) remains a hot topic in AI, with even the most cutting-edge models, such as Large Language Models (LLMs), employing RL to tackle complex challenges like alignment. As a result, RL and Deep RL in particular held a prominent position among Khipu topics.

For those interested in theory, Pablo Samuel Castro talk was the place to be. Castro presented Reinforcement Learning I, a comprehensive introduction to Reinforcement Learning. He explained fundamental concepts like reward and policy functions, their interaction through Bellman Equations, building an agent using either Value Iteration or Policy Iteration algorithms, and the importance of balancing exploration and exploitation.

To dive deeper into DRL, Doina Precup, Head of the Montreal office of DeepMind, presented Reinforcement Learning II. This talk provided an overview of more advanced RL techniques, including the choice between Value Iteration and Policy Iteration and the circumstances in which one is more useful than the other. She also presented various methods for agents to learn from the environment and shared some tricks to improve learning speed.

For a hands-on experience with these techniques, explore the practicals section of this blog post. The RL practical session offers a chance to apply several of these methods in real-world scenarios.

Demystifying AI ethics and fairness

Sara Hooker's enlightening talk, Beyond the Myth of the Perfect Model, delved into the crucial aspects of ethics, fairness, and safety in AI, examining how algorithmic bias can lead to unjust outcomes. She emphasized the importance of considering factors beyond top-line performance metrics, noting that high accuracy does not guarantee "true" learning.

Watch out for Clever Hans moments when doing ML

In her discussion of fairness and safety, Hooker stressed the potential for design decisions to disproportionately impact different groups. She highlighted geographic biases in dataset collection, which can cause models to underperform in underrepresented regions. Furthermore, she addressed the challenges of mitigating harm in AI, considering the complexities arising from dynamic data, labeling difficulties, and the lack of a universal definition of harm.

Hooker concluded by asserting that bias is more than just a data issue; it can also be amplified by model choices and treatment of data. Her talk serves as a valuable resource for those interested in the ethical dimensions of AI. Don't miss Sara Hooker's talk for a deeper exploration of these topics.

Mathematics unveiled

Mathematics, the bedrock of all sciences, serves as the basis for the models we have come to appreciate and rely on in AI. This series of talks focused on providing fundamental, principled, and provable approaches to model reality and enhance our understanding.

Luciana Ferrer, opened Khipu by introducing Expected Cost (of class misclassification), a metric widely used for training but not for evaluating probabilistic classification models. She showcased the metric's various useful properties in terms of calibration and applicability under different levels of class imbalance, proving it to be superior to the more commonly used F-beta. This led to a highly practical and widely applicable result with proven theoretical guarantees.

Rachel Ward shared insights into The Power of Adaptive Stochastic Gradient Descent. She demonstrated the favorable convergence properties of AdaGrad under a less restrictive version of Lipschitz smooth functions, offering insights into why the method performs well on a wide variety of objective functions.

Jeremías Sulam presented Neural Networks as Sparse Local Lipschitz Functions. In his talk, he examined the robustness and generalization guarantees of Neural Networks by investigating their sparsity structures.

Research at the forefront

Khipu brought together leading researchers from academia and industry to share their experiences and provide insights on how to become better researchers and professionals. Here are some highlights from this section of the event:

Samy Bengio shared with us some promising research lines that Apple is currently working on. One of these lines focuses on proving that modern language and vision models aren’t able to reason as humans do. Another line of work, focused on the subject of bias in modern LLMs, is their research on Self-Conditioning Pre-Trained Language Models, which tries to condition a model to be biased toward certain concepts in order to eradicate harmful biases that these models have.

A panel consisting of Sara Hooker, Kyunghyun Cho, Joan Bruna, and Martin Arjovsky had a live conversation on How to Write a Great Research Paper. They provided some useful tips, such as referencing similar works that have inspired your research and recognizing the limitations of your solution positively to pave the way for future improvements. Overall, they emphasized the importance of trusting your ideas and doing work that you are proud of. Stay tuned for more updates on our social media channels, as we will be sharing further insights and tips based on this panel discussion.

Women in AI

Khipu hosted the ‘Women in AI’ reception aimed at fostering, encouraging, and supporting greater diversity in AI in the Latin American community.

The event brought together attendees, speakers, and other guests in a relaxed atmosphere, featuring a panel discussion by four remarkable women (Magdalena Fuentes, Sara Hooker, Paula Martinez, and Aiala Rosá) who shared their valuable experiences. One key insight from the discussion was the power of weaving networks and mutual support among professionals.

The event also recognized the outstanding contributions of Alicia Fernández and Dina Wonsever, two emblematic female professors from UdelaR.

To cap off the memorable evening, the Uruguayan rock band No Te Va Gustar performed live, and some even joked that the world's nerdiest mosh pit was formed there.

Hands-on learning: practical sessions

Dive into the practical side of AI at Khipu 2023 with these comprehensive practical sessions. Ranging from high-performance computation to social impacts of AI, these Google Colab notebooks are designed to help you consolidate your knowledge and enhance your skills.

Here's a summary of the available practicals:

• JAX, Optax, Haiku and CNN: Explore high-performance, hardware-agnostic computation library, and Optimization and Neural Networks libraries built on top of them and optimization libraries for neural networks.
• Transformers and Attention: Learn the basic architecture behind current LLMs and build an attention model from scratch.
• Graph Neural Networks: Delve into graph theory and implement popular GNNs.
• Deep Generative Models: Create new samples of training data distribution with Denoise Diffusion Models.
• Intro to Reinforcement Learning: Implement RL techniques and train an agent to balance a cartpole.
• Social Impacts of Artificial Intelligence: Explore unintentional biases in trained models, particularly LLMs, and learn how to measure them.
• Building ML in an Open and Collaborative Way with the Hugging Face ecosystem: Discover the relevance of open-source in ML and build an end-to-end news summarizer service.

Time series forecasting: from theory to practice

Last but not least: A Hands-On Forecasting Guide. In a remarkable collaboration between Google and Tryolabs, this code camp provided invaluable insights and best practices for addressing time series forecasting problems. We were privileged to welcome Mathieu Guillame-Bert and Richard Stotz from Google's Zurich team to Khipu, where they co-hosted the code camp alongside our team.

Explore the key structural components of time series data and delve into various modeling approaches, including the Gradient Boosted Decision Tree (GBDT) implementation of the TensorFlow Decision Forests library by Google. Acquire versatile, time-series specific feature engineering techniques that can be effortlessly integrated into any forecasting project. These techniques are being incorporated into Temporian, an innovative, general-purpose time series forecasting library under active development by Tryolabs and Google. Stay tuned for more updates on this pioneering project!

Wrapping up Khipu 2023

The closing event was held at Teatro Solís, a historic building in Montevideo that accommodated a larger audience than the previous days. The speakers shared inspiring AI stories from Latin America, and keynote speaker Peter Norvig discussed the future of programming with a focus on Generative AI. One of the event's highlights was Martin Rocamora's presentation on musical diversity and representation, which included an incredible live performance of Candombe, a Uruguayan rhythm played by a group of drummers.

Khipu 2023 was a resounding success, bringing together experts, researchers, and enthusiasts from across the globe to discuss and share the latest advancements in AI. The conference not only showcased state-of-the-art techniques and applications in fields like computer vision, graph neural networks, generative models, and reinforcement learning but also emphasized the importance of community building and collaboration in Latin America.

We hope the insights and experiences shared continue to inspire and empower the AI community, fostering innovation and driving positive change in the region and beyond. As we look forward to future Khipu events, we invite you to stay connected and engaged with the thriving AI ecosystem in Latin America.

For those who made it to the end of this recap, we hope you enjoyed revisiting Khipu 2023's main takeaways. Stay tuned for future events and opportunities to join the thriving AI community in Latin America, and be sure to check our socials for the full coverage of this year's edition. Explore our social media coverage and keep an eye out for upcoming events!