About Origins of Intelligence Testbed
This testbed is a toolkit for building computational models of newborn brains.
Why use this testbed?
Both biological and artificial neural networks (BNNs and ANNs) are heavily shaped by their training data. To compare BNNs and ANNs, we must therefore give them the same set of training data. This testbed is designed for this purpose.
The testbed contains a collection of environments for training and testing autonomous artificial agents. Each environment mimics the environment from which we collected precise behavioral data from newborn animals. By training and testing your agents in our environments, you can examine whether your algorithm learns like newborn brains.
Comparing Artificial and Biological Neural Networks
Each environment contains training data and test data, which are projected on display walls surrounding the animals and agents. As shown to the right, the animals (top) and agents (bottom) can be trained and tested with the same visual stimuli (images, videos). After training and testing your agent, the testbed will provide results and graphs showing how your agent compares with newborn animals across the experimental conditions.
Our testbed contains experiments testing a wide range of visual abilities, including visual parsing, object recognition, action recognition, object permanence, numerical cognition, and so on. We will also continue adding new experiments over time.
Our goal is to enable researchers to quickly get a sense of how their model scores against standardized behavioral benchmarks on multiple dimensions and facilitate comparisons to other state-of-the-art models.
This quantified approach lets us keep track of how close our models are to real animals on a range of experiments. The precise behavioral data collected from newborn animals provide essential benchmarks for comparing models of cognitive development and measuring progress.