For some things, humans have a leg (or two) up on computers. While machines excel in speed and accuracy, we trounce our technological competitors when it comes to ditching the rules, recognizing visual patterns, and learning from our past. In the last decade, scientists have harnessed these talents by recruiting people to play citizen science computer puzzles – data-inspired games in which everyday people can work to solve science’s biggest questions from the comfort of their homes. Through gaming, thousands of individuals each year help advance our understanding of distant galaxies, quantum computers, and the proteins in our bodies.
These programs capitalize on human intuition for scientific gain, but some researchers believe that, with the rise of machine learning, citizen science is on the verge of a change. Right now, scientists rely on droves of citizens to help further their research. In the future, artificially intelligent computers may do the same work.
Many attributes of the human brain have historically made us better than machines at interpreting certain data. While computers are limited by their programmed parameters, humans are more adaptable. Consider Foldit, a game that determines protein structures by presenting the data as a puzzle – the accurate a model is, the more points a player achieves. Current protein model-building algorithms are designed to create the most stable protein designs possible, since stability reflects protein model accuracy. However, sometimes making counterintuitive moves will produce the most stable end point. “That’s where [human] Foldit players really excel,” said Amanda Winburn, a doctoral candidate at Indiana University. “If they make this move which gives them a short-term disadvantage, then they can have a long-term advantage.”
Humans outcompete computer programs when it comes to identifying patterns as well. The game EyeWire relies on pattern recognition to map the structures of neurons. Players rack up points by coloring in the cross-sections of a neuron in a two-dimensional slice of the brain. Layer by layer these slices add up to create a three-dimensional neuron model.
However, recently the researchers involved in projects like Foldit and EyeWire have turned their attention towards neural networks, a popular type of machine-learning programming inspired by the human brain. This kind of technology is still relatively new, but some researchers predict that, with neural networks, computers may soon replace human volunteers, acting as “citizen science cyborgs.”
Many citizen science directors want to implement networks to make their projects better equipped to analyze mountains of data. Volunteers have been a valuable resource to scientists, but learning algorithms may be a more efficient alternative. Volunteers who play Foldit normally take 2 to 4 weeks to complete a protein puzzle, but neural networks could speed up the rate to just a few days.
Neural network systems comprise a series of branching nodes that operate like neurons. Computer programmers feed information to the input node, which then transmits the information across the entire network. The network then processes this information until it generates the desired product.
The crucial component of these neural networks is their ability to learn from past experience. Winburn compares this memory-driven computation to an educated guess. “For example, if I have the credit card from somebody and I’m trying to guess their PIN number, one approach would be a brute force guess, where I just have a list of all the possible PIN numbers,” she said. “But if I know this person’s PIN for another non-related thing, then I would try that one first.”
Before a neural network can act on its memory, it first needs a memory. After developing the software, programmers “train” their computers to perform a function, feeding the system massive amounts of data, like a package of digital notecards to learn from.
There are three ways to train a network. One strategy, called supervised learning, provides the system with a labeled dataset. On the other hand, unsupervised learning gives a dataset with no labels, leaving the system to pick up on patterns by itself. For example, if you give the computer system a television show reel in a foreign language, eventually the system will discern patterns in the videos from educated guesswork. The last technique is reinforced learning, in which the programmer grades the algorithm on how correctly it guessed.
Some volunteers worry that neural networks will replace their work entirely. Many players dedicate their time to these projects because they like to feel integral to the scientific process. Sometimes researchers even include online players as co-authors when publishing crowd-sourced research results.
Program directors don’t want to leave behind this work force and instead, plan to redefine the human role in their projects. Scientists predict that citizens will stay involved as overseers, checking in on the programs to ensure they are on track. In the future, volunteers may act more as citizen science guides, aiding learning machines as they process unexplored territory.
