BI 173 Justin Wood: Origins of Visual Intelligence
Aug 30, 2023
auto_awesome
In this podcast, Justin Wood discusses his work comparing the visual cognition of newborn chicks and AI models. He uses controlled-rearing techniques to understand visual intelligence and build systems that emulate biological organisms. They explore topics like object recognition, reverse engineering, collective behavior, and the potential of transformers in cognitive science.
Testing whether machine learning systems can develop similar capacities as newborn animals by comparing the visual cognition of newborn chicks and AI models.
The significance of slowness and smoothness in learning, where newborn chicks require objects to move slowly and smoothly to develop high-level visual capabilities.
The comparison between convolutional neural networks (CNNs) and transformers in replicating chick visual capabilities, suggesting a possible universal principle of fitting to data distributions in learning.
The importance of considering embodiment, development, and morphology in understanding intelligence, with the goal of creating a unified model that explains the development of various cognitive capacities.
Deep dives
Comparison of machine learning systems and newborn animal capabilities
The podcast episode discusses the goal of testing whether machine learning systems, given the same input as newborn animals, can develop similar capacities. The focus is on understanding the gap between animals and machines and the potential role of critical periods in AI research. The guest, Justin Wood, runs the Wood Lab at Indiana University and compares the early development of natural organisms to artificial agents. The lab conducts controlled rearing experiments with chicks, controlling their visual experiences, and builds AI models trained on the same visual data. The experiments explore visual capabilities such as view-invariant recognition, and present findings on how the trained AI models compare to the newborn chicks in performing tasks.
Importance of slowness and smoothness in learning
The episode highlights the significance of slowness and smoothness in learning. It is observed that newborn chicks require objects to move slowly and smoothly in order to develop high-level visual capabilities. Rapid or erratic movement of objects leads to abnormal object representations. The experiments show that matching these constraints in artificial neural networks through accurate training data enables them to solve the same visual tasks as the chicks. The discussion emphasizes the exploration of time-based constraints and fitting to underlying data distributions as fundamental principles of visual learning.
Comparing CNNs and Transformers in replicating chick visual capabilities
The podcast highlights the comparison between convolutional neural networks (CNNs) and transformers in replicating chick visual capabilities. CNNs trained on data simulating visual experiences of the chicks perform better than fully trained CNNs, suggesting a closer match to the chicks' cognitive abilities. Transformers, although performing slightly worse than CNNs, still demonstrate the same capacities. The experiments challenge the notion that AI models require significantly larger amounts of data than newborn brains, suggesting a possible universal principle of fitting to data distributions in learning.
Further exploration and future trajectory
The podcast discusses the ongoing research and future trajectory in studying chick visual development and comparing it to machine learning systems. The aim is to create a closed-loop system where animals and machines are connected without human bias. Researchers plan to explore different architecture sizes and continue investigating the role of training data and fitting to data distributions. The discussion introduces new virtual reality chambers for studying more naturalistic environments and the potential for studying other cognitive abilities in addition to vision. The goal is to gradually close the gap between animal and machine capabilities.
Distinguishing between naturalism of stimuli and naturalism of data
The podcast episode emphasizes the importance of distinguishing between the naturalism of stimuli and the naturalism of data in computational neuroscience. The naturalism of stimuli refers to the images or stimuli presented to a model, while the naturalism of data refers to the data that the model can acquire from its environment. The speaker argues that while computational neuroscience often focuses on using disembodied models and providing training data to these models, it is crucial to consider the naturalism of embodiment as well. This distinction is seen as important in understanding the impact of both stimuli and embodiment on learning capacities.
Exploring collective behavior and imprinting with reinforcement learning agents
The podcast episode discusses experiments conducted with embodied AI agents to study collective behavior and imprinting. The researchers aim to build closed-loop systems that connect biological and artificial systems through embodiment. The experiments involve using reinforcement learning agents to explore criteria that lead to collective chick behavior. By giving agents reinforcement learning and a convolutional neural network (CNN) as an encoder, the experiments demonstrate that these agents can exhibit collective behavior and imprinting. The findings suggest that curiosity and reinforcement learning algorithms, combined with the experience of being raised together, contribute to the development of collective behavior in both biological and artificial systems.
Importance of embodiment, development, and morphology in understanding intelligence
The podcast episode highlights the importance of considering embodiment, development, and morphology in understanding intelligence. The speaker argues that building a unified model of intelligence requires taking development seriously and embracing morphology. The morphology of an organism, including its visual system, body, and physiological needs, can place constraints on how it interacts with the world and acquires training data. The speaker suggests that exploring the role of embodiment and agency, along with understanding how different constraints impact learning and behavior, will contribute to bridging the gap between animals and machines. The goal is to create a unified model that explains the development of various cognitive capacities, such as object recognition, navigation, and decision-making.
Justin Wood runs the Wood Lab at Indiana University, and his lab's tagline is "building newborn minds in virtual worlds." In this episode, we discuss his work comparing the visual cognition of newborn chicks and AI models. He uses a controlled-rearing technique with natural chicks, whereby the chicks are raised from birth in completely controlled visual environments. That way, Justin can present designed visual stimuli to test what kinds of visual abilities chicks have or can immediately learn. Then he can building models and AI agents that are trained on the same data as the newborn chicks. The goal is to use the models to better understand natural visual intelligence, and use what we know about natural visual intelligence to help build systems that better emulate biological organisms. We discuss some of the visual abilities of the chicks and what he's found using convolutional neural networks. Beyond vision, we discuss his work studying the development of collective behavior, which compares chicks to a model that uses CNNs, reinforcement learning, and an intrinsic curiosity reward function. All of this informs the age-old nature (nativist) vs. nurture (empiricist) debates, which Justin believes should give way to embrace both nature and nurture.
0:00 - Intro
5:39 - Origins of Justin's current research
11:17 - Controlled rearing approach
21:52 - Comparing newborns and AI models
24:11 - Nativism vs. empiricism
28:15 - CNNs and early visual cognition
29:35 - Smoothness and slowness
50:05 - Early biological development
53:27 - Naturalistic vs. highly controlled
56:30 - Collective behavior in animals and machines
1:02:34 - Curiosity and critical periods
1:09:05 - Controlled rearing vs. other developmental studies
1:13:25 - Breaking natural rules
1:16:33 - Deep RL collective behavior
1:23:16 - Bottom-up and top-down
Get the Snipd podcast app
Unlock the knowledge in podcasts with the podcast player of the future.
AI-powered podcast player
Listen to all your favourite podcasts with AI-powered features
Discover highlights
Listen to the best highlights from the podcasts you love and dive into the full episode
Save any moment
Hear something you like? Tap your headphones to save it with AI-generated key takeaways
Share & Export
Send highlights to Twitter, WhatsApp or export them to Notion, Readwise & more
AI-powered podcast player
Listen to all your favourite podcasts with AI-powered features
Discover highlights
Listen to the best highlights from the podcasts you love and dive into the full episode
