Researchers suggest that AI can learn common sense from animals

AI researchers developing reinforcement learning tools could learn a lot from animals. This is based on recent analysis by Google DeepMind, Imperial College London, and University of Cambridge researchers evaluating AI and non-human animals.

In a decade-long quest to advance machine intelligence, the AI ​​research community has turned frequently to neuroscience and behavioral science for inspiration and a better understanding of how intelligence is formed. However, these efforts have mainly focused on human intelligence, especially that of babies and children.

“This is especially true in an intensified learning context in which, thanks to advances in deep learning, it is now possible to bring comparative cognition methods to bear directly,” says the researchers’ article. “Animal knowledge provides a compendium of well-understood, non-linguistic, intelligent behavior. experimental methods of assessment and benchmarking are proposed; and it can guide the environment and task design. ”

DeepMind introduced some of the first forms of AI that combine deep learning and enforcement learning, such as the Deep Q-Network (DQN) algorithm, a system that played numerous Atari games at superhuman levels. AlphaGo and AlphaZero also used deep learning and enforcement learning to train the AI ​​to beat a human Go champion and perform other feats. More recently, DeepMind has developed an AI that automatically generates reinforcement learning algorithms.

On the human cognition side, DeepMind’s Embassy Director for Neuroscience Matthew Botvinick urged machine learning practitioners to become more interdisciplinary with neuroscientists and psychologists at an HAI conference in Stanford earlier this month.

Unlike other methods of training AI, deep reinforcement learning gives an agent a goal and a reward, an approach similar to training animals with food rewards. Previous animal recognition studies have looked at a number of species, including dogs and bears. Cognitive behaviorists have discovered higher levels of intelligence in animals than previously thought, including the self-esteem of dolphins and the crows’ ability to retaliate.

Studies of animal cognitive abilities can also inspire AI researchers to look at problems in other ways, especially in-depth learning. While researchers draw parallels between animals in test scenarios and reinforcement learning tools, the idea of ​​testing the cognitive abilities of AI systems has developed. Other forms of AI, such as the Alexa or Siri assistants, cannot search a maze for a box that contains a reward or meal.

The team’s paper, “Artificial Intelligence and Common Sense of Animals,” published in CellPress Reviews – cites cognitive experiments with birds and primates.

“Ideally, we would like to develop an AI technology that can capture these interrelated principles and concepts as a systematic whole and that manifests itself in an ability on a human level for generalization and innovation,” the paper says. “How one builds such AI technology remains open. However, we advocate an approach where RL agents, possibly with undeveloped architectures, acquire what is needed through advanced interaction with large virtual environments. ”

When it comes to creating systems like the ones mentioned in the paper, the challenge is to convey to the agents that they exist in an independent world. Training agents to understand the concept of common sense is another hurdle, along with identifying the environments and tasks that are best suited to the task.

A prerequisite for the use of common sense by training agents is 3D simulated worlds with realistic physics. These can simulate objects such as B. Trays that can be torn apart, lids that can be unscrewed, and packages that can be torn open.

“This is one of the technological capabilities of today’s physics engines, but such rich and realistic environments have yet to be deployed on a large scale for training RL agents,” the paper says.

The researchers argue that common sense is not a unique human characteristic, but depends on some basic concepts, such as understanding what an object is, how the object occupies space, and the cause-effect relationship. These principles include the ability to perceive an object as a semi-permanent thing that can remain quite persistent over time.

One of the forms of knowledge shown by animals is an understanding of the permanence of objects and that a reward can be in a container, just as a shell can contain a seed. The challenge of providing agents with such common sense principles can be viewed as the problem of finding assignments and curricula that, when properly architected, result in trained agents that can pass appropriately designed transmission assignments.

“Although modern deep RL agents can learn to solve several tasks very effectively, and some architectures have rudimentary forms of transmission, it is by no means clear that every current RL architecture is able to acquire such an abstract concept. But suppose we had a candidate, how would we test whether he has acquired the concept of a container? ”

The researchers believe that common sense tasks for training agents should be able to develop an understanding without seeing many examples, approaches known as low-shot or zero-shot learning.

The common sense assessment included in the article focuses on part of common sense physics and does not take into account other forms of common sense such as psychological concepts, the ability to determine various forms of objects such as liquids or gases, or the understanding of objects that are these can be manipulated like paper or a sponge.

In other recent developments in enhanced learning, UC Berkeley professor Ion Stoica spoke at VentureBeat’s recent Transform conference about why supervised learning is far more common than enhanced learning. Stanford University researchers introduced LILAC to enhance enhanced learning in dynamic environments, and Georgia Tech researchers collaborated to learn NLP and reinforcement to create AI that excels in text adventure games.