A UK-based research team has achieved an a-MAZE-ing development—it has created an artificial intelligence (AI) program with the ability to learn to find shortcuts through a maze to attain its goal. As part of the process, structures similar to those in the human brain were developed by the program.
The development of these computational “grid cells” has been explained in the Nature journal and could assist researchers in developing advanced navigational software for next-generation robots and could even offer an innovative platform through which mysteries of the mammalian brain can be probed.
In the recent past, AI scientists have designed and refined deep-learning networks - layered programs with the ability to provide innovative solutions to accomplish their assigned aim. For instance, one can notify a deep-learning network the face to be identified from a series of distinctive photos. Moreover, through multiple training rounds, it can refine its algorithms until the right face is spotted almost every time.
According to Francesco Savelli, a Johns Hopkins University neuroscientist who had no role in the paper, although these networks are inspired by the brain, they do not function anywhere close to it. To date, AI systems are not even close to simulating the architecture of brain, the complexity of individual neurons, the diversity of real neurons, or even the rules through which they learn.
“Most of the learning is thought to occur with the strengthening and weakening of these synapses,” Savelli stated in an interview, in relation to the connections between neurons. “And that’s true of these AI systems too - but exactly how you do it, and the rules that govern that kind of learning, might be very different in the brain and in these systems.”
In spite of this, AI has been quite helpful in various functions, from facial recognition to translating languages and deciphering handwriting, stated Savelli. However, higher level activities (e.g. navigating a complex environment) have been proven to be highly difficult.
Our brain has been observed to perform one navigational aspect without any conscious effort—path integration. This process is used by mammals to recalculate their position at the end of every step by taking into account the distance traveled and the direction faced by them. It has been considered to be central to the ability of the brain to produce a map of its surroundings.
Some of the neurons related to these “cognitive maps” are place cells that get initiated when the owner is in a specific spot in the environment, head-direction cells that inform the owner of the direction they are facing, and grid cells that seem to respond to an imaginary hexagonal grid mapped over the surrounding terrain. The neuron gets initiated every time the owner steps on a “node” in this grid.
“Grid cells are thought to endow the cognitive map with geometric properties that help in planning and following trajectories,” wrote Savelli and James Knierim, fellow Johns Hopkins neuroscientist, in a commentary on the paper. Three scientists were awarded the 2014 Nobel Prize in physiology or medicine for the discovery of grid cells.
Humans and other animals appear to have little, if any, difficulty moving through space since all these highly specialized neurons function together to inform a person where they are and where they are heading.
Researchers at DeepMind, owned by Google and University College London, wondered whether a program with the ability to also carry out path integration could be developed. Hence, they used simulations of paths used by rodents searching for food to train the network. In addition, they fed it with data related to the movement of a rodent and its speed, and also feedback from simulated head-direction cells and place cells.
At the time of this training, the scientists observed a peculiar phenomenon: The simulated rodent seemed to develop patterns of activity that were astonishingly similar to grid cells—despite the fact that grid cells were never a part of their training system.
“The emergence of grid-like units is an impressive example of deep learning doing what it does best: inventing an original, often unpredicted internal representation to help solve a task,” wrote Savelli and Knierim.
Grid cells seem to be so helpful for path integration that this simulated rodent produced a solution that had an uncanny resemblance to the brain of a real rodent. The scientists then wondered whether grid cells can also be used in another important aspect of mammal navigation.
Termed vector-based navigation, that aspect is essentially the potential to calculate the straight-shot, “as the crow flies” distance to a goal even though a less-direct, longer route was originally taken. Savelli indicated that it is a useful attribute for finding shortcuts to one’s destination.
The researchers tested this by challenging the grid-cell-enabled simulated rodent to solve a maze; however, in this case, they blocked off majority of the doorways such that the program would have to opt for the long route to reach its goal.
The program was also modified such that rewards were given for its actions that brought it closer to the goal. The network was trained on a specific maze and shortcuts were then opened to observe what happened. Most certainly, the faux-rodent with grid cells was able to quickly find and use the shortcuts, although those pathways were new and not known earlier.
Moreover, it performed much better than a simulated rodent, the start point and goal point of which were tracked only through the place cells and the head-direction cells. According to the authors of the paper, it even outperformed a “human expert.”
Savelli stated that ultimately, the outcomes of this study can prove useful for robots venturing into unknown territory. Furthermore, from a neuroscientific point of view, they could assist researchers in gaining better insights into the way these neurons perform their functions in the mammalian brain.