-
Flaws of ImageNet, Computer Vision's Favorite Dataset
Since its release, ImageNet-1k has been a gold standard for evaluating model performance. It has served as the foundation of numerous other datasets and it has been widely used for pretraining.
As models have improved, issues related to label correctness have become increasingly apparent. In this blog post, we analyze the issues, including incorrect labels, overlapping or ambiguous class definitions, training-evaluation domain shifts, and image duplicates. The solutions for some problems are straightforward. For others, we hope to start a broader conversation about how to improve this influential dataset to better serve future research. -
Flow With What You Know
This tutorial provides an accessible introduction to flow-matching and rectified flow models, which are increasingly at the forefront of generative AI applications. Typical descriptions of them are often laden with extensive probability-math equations, which can form barriers to the dissemination and understanding of these models. Fortunately, before they were couched in probabilities, the mechanisms underlying these models were grounded in basic physics, which provides an alternative and highly accessible (yet functionally equivalent) representation of the processes involved.
-
How do we interpret the outputs of a neural network trained on classification?
This post shows how neural networks trained for classification approximate the Bayesian posterior, explaining the theoretical basis and providing empirical examples.
-
How to visualize training dynamics in neural networks
Deep learning practitioners typically rely on training and validation loss curves to understand neural network training dynamics. This blog post demonstrates how classical data analysis tools like PCA and hidden Markov models can reveal how neural networks learn different data subsets and identify distinct training phases. We show that traditional statistical methods remain valuable for understanding the training dynamics of modern deep learning systems.
-
In Search of the Engram in LLMs: A Neuroscience Perspective on the Memory Functions in AI Models
Large Language Models (LLMs) are enhancing our daily lives but also pose risks like spreading misinformation and violating privacy, highlighting the importance of understanding how they process and store information. This blogpost offers a fresh look into a neuroscience-inspired perspective of LLM's memory functions, based on the concept of engrams-the physical substrate of memory in living organism. We discuss a synergy between AI research and neuroscience, as both fields cover complexities of intelligent systems.
-
Linear Recurrences Accessible to Everyone
Investigating linear RNNs such as Mamba, can be challenging because they are currently not efficiently expressible in PyTorch. We propose the abstraction of linear recurrences to gain intuition for the computational structure of these emerging deep learning architectures. After deriving their parallel algorithm, we gradually build towards a simple template CUDA extension for PyTorch. We hope that making linear recurrences accessible to a wider audience inspires further research on linear-time sequence mixing.
-
LLMs' Potential Influences on Our Democracy: Challenges and Opportunities
With growing research and attention on LLMs' potential influence on political discourse and democratic processes, this blog post discusses the path forward and proposes future research questions in four broad areas: (1) evaluation of LLM political leanings, (2) understanding LLMs' influence on our democracy, (3) better policy frameworks for AI development, and (4) technical solutions to adjust or mitigate political leanings. As LLMs become increasingly integrated into society, continued investigation of how they will reshape democracy is essential to maximize their benefits while minimizing risks to democratic processes.
-
Lost in Prediction: Why Social Media Narratives Don't Help Macroeconomic Forecasting?
Can we predict the macroeconomy by analyzing the narratives people share on social media? We dove deep into the world of Narrative Economics, using NLP models to analyze millions of viral tweets and see if they could help us predict the fluctuations of macroeconomic indicators. 🚨 Spoiler alert: it's not that easy! Join us as we explore the interesting relationship between narratives, social media, and macroeconomy, and uncover the challenges of turning narratives into treasure.
-
Mechanistic Interpretability Meets Vision Language Models: Insights and Limitations
Vision language models (VLMs), such as GPT-4o, have rapidly evolved, demonstrating impressive capabilities across diverse tasks. However, much of the progress in this field has been driven by engineering efforts, with a limited understanding of how these models work. The lack of scientific insight poses challenges to further enhancing their robustness, generalization, and interpretability, especially in high-stakes settings. In this work, we systematically review the use of mechanistic interpretability methods to foster a more scientific and transparent understanding of VLMs. Specifically, we examine five prominent techniques: probing, activation patching, logit lens, sparse autoencoders, and automated explanation. We summarize the key insights these methods provide into how VLMs process information and make decisions. We also discuss critical challenges and limitations that must be addressed to further advance the field.
-
Models trained with unnormalized density functions: A need for a course correction
Training a generative model with energy or unnormalized density functions is considered an important problem for physical systems such as molecules. This provides a path to train generative models to sample from the much desired Boltzmann distribution in situations of data scarcity. As of late, several generative frameworks have been proposed to target this problem. However, as we show in the following blog post, these methods have not been benchmarked sufficiently well against traditional Markov Chain Monte Carlo (MCMC) methods that are used to sample from energy functions. We take the example of two recent methods (iDEM and iEFM) and show that MCMC outperforms both methods in terms of number of energy evaluations and wall clock time on established baselines. With this, we suggest a “course correction” on the benchmarking of these models and comment on the utility and potential of generative models on these tasks.
-
Multi-modal Learning: A Look Back and the Road Ahead
Advancements in language models has spurred an increasing interest in multi-modal AI — models that process and understand information across multiple forms of data, such as text, images and audio. While the goal is to emulate human-like ability to handle diverse information, a key question is: do human-defined modalities align with machine perception? If not, how does this misalignment affect AI performance? In this blog, we examine these questions by reflecting on the progress made by the community in developing multi-modal benchmarks and architectures, highlighting their limitations. By reevaluating our definitions and assumptions, we propose ways to better handle multi-modal data by building models that analyze and combine modality contributions both independently and jointly with other modalities.
-
Open-Source vs Close-Source: The Context Utilization Challenge
This blog post aims to evaluate how well the most capable open-source long context large language models (LLMs) utilize context, using the Needle In A Haystack test. We adopt the task of chapter summarization for recently published books to minimize data contamination while ensuring a challenging test. Our results show that open-source models still have room to improve in context utilization compared to close-source models.