DeepMind: Deep learning algorithms can outperform human intelligence in many ways: from classifying images to reading lips to accurately predicting the future. But despite their superhuman levels of proficiency, they lag behind in the rate at which they learn.
Some of the best machine learning algorithms take hundreds of hours to study and master classic video games, something a human can learn in an afternoon. That may have something to do with the neurotransmitter dopamine, according to a paper by Google’s DeepMind subsidiary in the journal Nature Neuroscience.
Meta-learning, or the process of quickly learning from examples and learning rules from those examples over time, is thought to be one way humans acquire new knowledge more efficiently than algorithms. But the underlying mechanisms of meta-learning are currently poorly understood.
In an effort to shed light on the process, researchers at DeepMind in London modeled human physiology using a recurrent neural network, a type of neural network that is able to internalize past actions and observations and learn from those experiences. The system, which mathematically optimizes the algorithm over time through trial and error, reportedly uses dopamine, a chemical in the brain that influences emotions, movements, and sensations of pain and pleasure, and plays a key role in the learning process.
The researchers then created a similar system in six neuroscientific meta-learning experiments, comparing its performance with that of animals that had undergone the same tests. One of the tests, known as the Harlow Experiment, gave the algorithm two randomly selected images, one of which was associated with a reward. In the original experiment, a group of monkeys very quickly learned a strategy for collecting rewards. They chose an object at random the first time, but immediately after the objects that had the reward.
The algorithm worked much like the animals did, choosing images that were directly associated with rewards from new images that it had not “seen again.” Furthermore, the researchers noted that the learning took place through the neural network, supporting the theory that dopamine plays a key role in meta-learning.
The dopamine study shows that medical science has a lot to gain from research into neural networks, just as computer science does.
“The use of AI data that can be applied to explain findings in neuroscience and psychology highlights the value of each field to the other,” the DeepMind team says. “Moving forward, we expect many benefits in the opposite direction, with guidance from the specific organization of brain circuits to design new models that learn from augmented AI.”
- Data centers: Using the heat of servers for heating!
- Microsoft, Apple, Google, Facebook, Amazon and captivity
- The strangest programming principles
