It is an exquisite challenge to understand how the brain organizes and access spatial information. The process involves recalling an entire network of memories and storing spatial data from thousands of other people. Scientists have identified elements such as grid cells. It is difficult to study slices of human gray matter to see how location-based memories of images, sounds and smells flow through and connect to each other.
Another way is offered by artificial intelligence. Neural networks are the engines that power most deep learning applications. The hippocampus, a structure of the brain critical to memory, is actually a transformer, according to recent research. Their model tracks spatial information in a way that is similar to the brain. Remarkable success has been seen by them.
James Whittington is a cognitive neuroscientist who splits his time between the lab of Tim Behrens at the University of Oxford and the one at Stanford University.
The ability of neural network models to mimic computations carried out by grid cells and other parts of the brain is said to be improved by transformer studies. Whittington said that such models could push our understanding of how artificial neural networks work.
David Ha said that they are not trying to create a new brain. Can we come up with a way to do what the brain does?
Five years ago, transformer appeared as a new way to process language They are the secret sauce in the programs that can create convincing song lyrics, write Shakespearean sonnets and impersonate customer service representatives.
Every input is connected to every other input using a mechanism called self- attention. Neural networks don't connect inputs to other inputs. While transformers were designed for language tasks, they have excelled at other tasks as well.
A group led by Sepp Hochreiter used a transformer to retool a model of memory retrieval called a Hopfield network. The networks were first introduced 40 years ago by the physicist John Hopfield.
The connection between how Hopfield networks retrieve memories and how transformers perform attention was found by Hochreiter and his colleagues. The Hopfield network was upgraded in order to turn it into a transformer. The model was able to store more memories because of that change. A transformer-based Hopfield network was proved to be biologically plausible by Hopfield and his co-workers at IBM.
Whittington and Behrens helped modify the transformer so that Hochreiter's approach of treating memories as a linear sequence instead of a string of words in a sentence could be used. The model's performance on neuroscience tasks was further improved by that "twist" The model was similar to models of the grid cell firing patterns shown in fMRI scans.
According to a neuroscientist at University College London, grid cells have an exciting, beautiful, regular structure and are not likely to be random. The new work shows how the patterns in the hippocampus are reproduced by transformers. It is possible for a transformer to figure out where it is based on previous states and how it has moved, in a way that is similar to models of grid cells.
All Rights Reserved © 2024
