AI Breakdown

In this episode, we discuss MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI by Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, Wenhu Chen. MMMU is a new benchmark for evaluating multimodal models using college-level questions from various disciplines to test advanced reasoning and subject knowledge. The benchmark contains 11.5K questions across six core disciplines and 30 subjects, featuring diverse visual content like graphs and music sheets. Initial testing on 14 models, including the sophisticated GPT-4V, showed a best accuracy of 56%, suggesting ample scope for improvement in artificial general intelligence.

In this episode we discuss MLP-Mixer: An all-MLP Architecture for Vision by Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, Alexey Dosovitskiy. The paper presents MLP-Mixer, an architecture that relies solely on multi-layer perceptrons (MLPs) for image classification tasks, demonstrating that neither convolutions nor attention mechanisms are necessary for high performance. The MLP-Mixer operates with two types of layers: one that processes features within individual image patches, and another that blends features across different patches. The model achieves competitive results on benchmarks when trained on large datasets or with modern regularization techniques, suggesting a new direction for image recognition research beyond conventional CNNs and Transformers.

In this episode we discuss Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine by Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, Renqian Luo, Scott Mayer McKinney, Robert Osazuwa Ness, Hoifung Poon, Tao Qin, Naoto Usuyama, Chris White, Eric Horvitz. The paper discusses enhancing the performance of GPT-4, a generalist language model, in medical question-answering tasks without domain-specific training. By innovatively engineering prompts, the researchers created Medprompt, which significantly outperformed specialized models, achieving state-of-the-art results on the MultiMedQA benchmark suite with fewer model calls. Moreover, Medprompt was also successful in generalizing its capabilities to other fields, demonstrating its broad applicability across various competency exams beyond medicine.

In this episode we discuss Nash Learning from Human Feedback by Remi Munos, Michal Valko, Daniele Calandriello, Mohammad Gheshlaghi Azar, Mark Rowland, Daniel Guo, Yunhao Tang, Matthieu Geist, Thomas Mesnard, Andrea Michi, Marco Selvi, Sertan Girgin, Nikola Momchev, Olivier Bachem, Daniel J. Mankowitz, Doina Precup, Bilal Piot from Google DeepMind. The paper introduces Nash Learning from Human Feedback (NLHF), a new approach for tuning large language models (LLMs) based on human preferences, different from the traditional reinforcement learning from human feedback (RLHF). The NLHF technique involves learning a preference model from paired comparisons and refining the LLM's policy towards a Nash equilibrium, where no alternative policy produces more preferred responses. They developed a Nash-MD algorithm and gradient descent approaches for implementing NLHF, and demonstrated its effectiveness on a text summarization task, suggesting NLHF as a promising direction for aligning LLMs with human preferences.

In this episode we discuss Animate Anyone: Consistent and Controllable Image-to-Video Synthesis for Character Animation by Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, Liefeng Bo. The paper presents a novel framework designed for character animation that synthesizes consistent and controllable videos from still images using diffusion models. It introduces a ReferenceNet that utilizes spatial attention to keep the character's appearance consistent and integrates a pose guider for movement controllability along with a technique to ensure smooth temporal transitions. The method exhibits superior performance on character animation, including fashion video and human dance synthesis benchmarks, outperforming other image-to-video methods.

This podcast explores the relationships between neural network models trained on diverse datasets, revealing clusters in weight space. By navigating between these clusters, new models can be created with stronger performance. Starting fine-tuning from specific regions within the weight space achieves improved results.

In this episode we discuss MobileCLIP: Fast Image-Text Models through Multi-Modal Reinforced Training by Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, Oncel Tuzel. The paper introduces MobileCLIP, a new efficient image-text model family optimized for mobile devices with a novel multi-modal reinforced training method that enhances accuracy without increasing on-device computational demands. MobileCLIP achieves better latency-accuracy trade-offs in zero-shot classification and retrieval tasks and outperforms existing models in speed and accuracy. The reinforced training method improves learning efficiency by factors of 10 to 1000 times, demonstrated by advancements in a CLIP model with a ViT-B/16 image backbone across 38 benchmarks.

In this episode we discuss Simplifying Transformer Blocks by Bobby He, Thomas Hofmann. The paper studies the possibility of simplifying standard transformer blocks without reducing training speed by experimenting with the removal of certain components such as skip connections and normalization layers. Using signal propagation theory along with empirical research, the authors justify modifications that allow for these simplifications. Their findings indicate that the streamlined transformer models match the performance and training speed of traditional transformers while offering increased training throughput and reduced parameter count.

Visual In-Context Prompting is a new framework in vision tasks that improves zero-shot learning capabilities. It allows an encoding-decoding architecture to utilize prompts like strokes, boxes, points, and context reference segments. The framework extends to a broader range of tasks including open-set segmentation and detection. The authors demonstrate performance enhancements and competitive results on various datasets.

In this episode we discuss GAIA: a benchmark for General AI Assistants by Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, Thomas Scialom. The paper introduces GAIA, a benchmark designed to assess the capabilities of General AI Assistants in performing tasks that are simple for humans yet difficult for AIs, such as reasoning, multi-modal tasks, web browsing, and general tool-use. It highlights a significant performance discrepancy, with humans scoring a 92% success rate contrasting with a mere 15% for an advanced AI model (GPT-4 with plugins). The authors propose this benchmark as a measure to guide AI research towards achieving robustness in tasks where humans excel, challenging the prevailing focus on skills that are difficult for humans, and establishing a leaderboard for tracking AI progress.