Pigeons don’t just think about bread all day long. A study by Ohio State University found that pigeons actually solve problems using the same principles as some AI systems.
Psychology professor Brandon Turner and his colleague Edward Wasserman from the University of Iowa devised an AI model that learns via “principles of response association through reinforcement learning and stimulus generalisation”.
This is because reinforcement learning combined with visual stimuli, such as an image or an object, is widely thought to be how pigeons learn. For example, if seeing a child at the park stomping his feet means no food and possible danger, pigeons would typically fly away. If seeing an old lady sitting at the bench with bread means food and no danger, pigeons would fly towards said lady.
If the AI model was able to learn in the same way as pigeons, then the experiment would be considered a success.
As smart as a pigeon, or…?
In the experiment, the AI model was programmed to carry out two mechanisms related to reinforcement learning and error correction. The model contained a grid of “representational points” to plot out how the AI model learnt specific tasks.
Meanwhile, pigeons were shown a variety of stimuli that included lines, angles and rings, and could peck a button to categorise which group the stimuli belonged to. If they got it right, they received a food pellet. If they got it wrong, they got nothing and stayed hungry.
The pigeons had to solve four simple tasks, and their performance whilst solving them was recorded.
In the end, Turner and Wasserman theorised that the AI model was able to solve problems in the way that pigeons do. Through reasoning and trial and error, the AI model was able to increase the number of correct answers throughout the experiment.
Similarly, as the pigeons progressed through the four tasks, they gradually learned to associate wrong answers with no food. As a result, their correct answers increased from 55% to 95% for the easier tasks. For the more difficult task, their correct responses improved from 55% to 68%.
“We found really strong evidence that the mechanisms guiding pigeon learning are remarkably similar to the same principles that guide modern machine learning and AI techniques,” Turner explained.
“Our findings suggest that in the pigeon, nature may have found a way to make an incredibly efficient learner that has no ability to generalise or extrapolate like humans would,” he continued.
This means that pigeons aren’t as dumb as we make them out to be. Or maybe some AI models … aren’t as intelligent as we think they are!
AI’s human-like intelligence
Both researchers said that the experiment was significant as it provided new insight into how pigeons thought and operated.
Pigeons are historically thought to be a bit thick, but we may need to rethink our perception.
“We celebrate how smart we are that we designed artificial intelligence, at the same time we disparage pigeons as dim-witted animals,” Turner pointed out.