#ICML #cognition #GrowAI We spent 2 years carefully curated every single experiment (i.e. object permanence, A-not-B task, visual cliff task) in this dataset (total: 1503 classic experiments spanning 12 core cognitive concepts).
We spent another year to get 230 MLLMs evaluated
Hokin Deng
2,854 posts
- Replying to @DengHokin🚀 Dive in 👇 🌐 Project: williamium3000.github.io/core-knowledge/ 📄 Paper: arxiv.org/abs/2410.10855 📝 OpenReview: openreview.net/forum?id=EIK6x… 📊 Dataset: huggingface.co/datasets/willi… 💻 Code: github.com/williamium3000… 😎 Team: growing-ai-like-a-child.github.io 💥 Huge thanks to my amazing collaborators
- Replying to @DengHokinFirst, MLLMs exhibit a reversed developmental trajectory 📉 compared to humans 📈: they perform better on "higher-level" abilities, sometimes comparable to or even surpassing humans, yet struggle with "lower-level" abilities 🙀, which humans typically develop these "lower-level"
- #embodied All forms of biological intelligence are grounded movements🏃♂️ muscles & motor neurons 🧠 emerge before visual cortex & rods & cones in eyes 👁️ Building monocular better-than-mocap-studio #video2motion is our critical step towards human embodied intelligence.💥We are excited to share that our #Video2Animation feature is now live at @MyolabAI's discord servers. We're giving away massive 𝐟𝐫𝐞𝐞 credits for early users. Head now to 👉 discord.gg/Gc9hapPgQg
00:00 - Replying to @DengHokin7) The result is astonishing 😱 All models - either they become totally illusory or they become completely short-cutted. None of the models is even close to core knowledge. No models are able to get both the manipulation tasks and the control tasks right at the same time,
- Replying to @DengHokinStarted with this old man, Jean Piaget, developmental psychology has become a formal discipline with generations of dedicated scientists and over a century of sustained, yet fruitful, curious endeavors. We first built our CoreCognition framework leveraging Piaget's four-stage
- Replying to @DengHokin2) The topology of internal cognitive representation of MLLMs is akin to human, suggesting "cognition" as a natural kind. Having 1503 individual cognitive experiments spanning 12 core cognitive concepts on 230 models allows us to interrogate the topology of internal cognitive
- Finally finish making our website @GrowAiLikeChild growing-ai-like-a-child.github.io 🙌 with @Williamiumli @carrot0817_ @ziqiao_ma @zory_zhang @Pinyuan3 @Pooyanrg and many others 🚶🚶♂️🚶♀️
- Replying to @DengHokinLiving in the AI world, all classic cognitive experiments from developmental psychology literature are instantiated as cognitive prototypes. Our cognitive scientists purchase real-world equipments and manually set up physical lab-setting environments to acquire media materials
- Replying to @MudrikNogaThis is totally unacceptable! I am very sorry to what happened to you.
- Replying to @DengHokin4) 🧐 🧐 Would Core Knowledge Emerge from Pure Scaling? ‼️‼️‼️Nope‼️‼️‼️ By applying linear regression to the performance data of 230 models with different model sizes & training data sizes, we quantify the scaling effect as the slope of the resulting regression line. The
- big neuroscience paper dropped in Nature today: memory re-consolidation in hippocampus is supported by neurogenesis
- Replying to @DengHokin‼️Results‼️We ran 1503 cognitive experiments spanning 12 core concepts on 230 MLLMs with 11 different prompts for 5 times to get over 3800K inference data points. Here is what we found ... (4/n)
- Replying to @DengHokin6) Last, we introduce "Concept Hacking" to reveal core knowledge deficiencies in the control experiment set-up. Concept Hacking systematically manipulates the task-relevant features while preserving all task-irrelevant conditions ... (10/n)














