Researchers from IBM, MIT and Harvard Announced The Release Of DARPA “Common Sense AI” Dataset Along With Two Machine Learning Models At ICML 2021

Source: https://arxiv.org/pdf/2102.12321.pdf

Building machines that can make decisions based on common sense is no easy feat. A machine must be able to do more than merely find patterns in data; it also needs a way of interpreting the intentions and beliefs behind people’s choices.

At the 2021 International Conference on Machine Learning (ICML), Researchers from IBM, MIT, and Harvard University have come together to release a DARPA “Common Sense AI” dataset for benchmarking AI intuition. They are also releasing two machine learning models that represent different approaches to the problem that relies on testing techniques psychologists use to study infants’ behavior to accelerate the development of AI exhibiting common sense. 

This research work is the first of its kind to use psychology to create more fluid and better AI systems. It aims to develop machine commonsense that makes sense because it has innate human qualities – such as intuition, common knowledge, or understanding of social cues.

https://arxiv.org/pdf/2102.12321.pdf

The benchmark unveiled at the ICML event is called AGENT (Action, Goal, Efficiency, coNstraint, uTility) and consists of 8,400 3D animations. These videos are organized into four categories — goal preferences, action efficiency, unobserved constraints, and cost or reward tradeoffs.

Like other infant studies, this research project has two phases in each trial. ‘In the familiarization phase, an AI model is fed videos demonstrating a particular agent’s behavior in certain physical environments. In the test phase, the model is shown a video of the behavior of the same agent in a new environment‘.

The two machine learning approaches advance more real-world training of AI and machine learning models using traditional human psychology methods. They compare Bayesian inverse planning with a Theory of Mind neural network, where they offer strong baselines for building an effective plan that uses these methodologies to learn from humans while maintaining their autonomy.

The study suggests that models need representations of how agents plan in order to pass the tests of core intuitive psychology. This is because they must compute utility and have knowledge about objects and physics embedded into them.

Paper: https://arxiv.org/pdf/2102.12321.pdf

Source: https://research.ibm.com/blog/icml-darpa-agent

Related Products