AI Performance Enhanced by Human Developmental Psychology

Researchers at Pennsylvania State University (Penn State) have published a new study that shows how a new artificial intelligence (AI) machine learning method that leverages concepts from human developmental psychology and early childhood learning achieves roughly 15% greater accuracy.

“The field of developmental psychology offers insight into what is missing in contemporary machine-vision learning,” wrote corresponding authors Brad Wible, professor of psychology, James Z. Wang, distinguished professor of information science and technology, and Penn State researchers Wongsek Lee and Li-Jen Zhu.

AI deep learning performance accuracy often requires huge amounts of training data. Even when exposed to vast amounts of training data, today's AI machine learning lacks the ability to generalize necessary for artificial general intelligence. For example, according to OpenAI, their Large Language Model (LLM) chatbot ChatGPT was trained from vast amounts of data from the internet, licensed third-party data, and information from users and human trainers.

“Despite being trained on massive datasets, current computer vision systems lag behind human children in learning about the visual world,” the researchers wrote.

Humans are naturally able to generalize concepts when presented with entirely new material. Technologists often argue that young children are smarter than the most sophisticated AI systems. For example, a child can learn the concept of cats and correctly identify other cats with limited exposure, without having been explicitly exposed to or taught all felines beforehand.

Researchers found that by the time a child is two years old, they have made roughly 90 million fixations on their environment, far fewer than the vast AI training datasets containing billions of images, and that fixations are primarily comprised of a limited set of faces, objects, and viewpoints in the home.

“As they view the world, children take advantage of a wealth of environmental information about how their bodies purposefully sample information through the controlled orienting of their senses and their interactions with the world,” the scientists write.

They can mimic the way human children learn Computer-based Will it lead to smarter, better-performing AI algorithms overall? Scientists have argued that human location awareness is key to infant visual learning, and have reasoned that incorporating environmental context into AI machine learning could enable learning that is flexible enough to better perform tasks like classification with less data.

Taking inspiration from human developmental psychology, the researchers decided to use the environment as a contextual data source, an approach they named Environmental Spatial Similarity (ESS). Using an interactive physics simulation platform called ThreeDWorld (TDW) and state-of-the-art ray tracing, the team created a simulation of an agent moving through a furnished house or apartment, with realistic shadows, reflections, and material properties.

The simulation had realistic features, such as glossiness, that “mimic the appearance of real gloss in human psychophysics.” Psychophysics is the branch of psychology that studies the relationship between the physical aspects of a stimulus and our reported sensations and perceptions, i.e., the relationship between matter and mind. The term was coined by physicist, professor, philosopher, and experimental psychologist Gustav Theodor Fechner (1801-1887).

For this study, the Penn State researchers modified a self-supervised AI learning algorithm called Momentum Contrast (MoCo) created in another study published in 2020 by researchers at Meta FAIR Labs (formerly Facebook AI Research), and implemented it using Pytorch.

Penn State's algorithm takes into account contextual spatial data from simulated images to identify image pairs that are close in spatial and rotational coordinates, allowing multiple pairs. The performance of their AI model was compared to the base model MoCo (version 2) algorithm under a range of conditions.

The new ESS approach outperforms baseline models across a range of tasks, improving by up to 15% (14.99%) on the task of identifying rooms in a virtual apartment with an average accuracy of 99.35%.

The researchers say there are many real-world applications that could benefit from their new AI approach, including human vision, neuroscience, computer vision, disaster relief, autonomous air vehicles, robotics and even future planetary and space exploration.

Copyright © 2024 Cami Rosso All rights reserved.