New Neural Machine Code for Programming Reservoir Computers

Reservoir computing is a promising computational framework based on recurrent neural networks (RNNs). An RNN essentially maps input data into a high-dimensional computational space, keeping some parameters of the artificial neural network (ANN) fixed and updating others. This framework helps improve the performance of machine learning algorithms while reducing the amount of data required to properly train the algorithms.

RNNs basically leverage iterative connections between different processing units to process sequential data and make accurate predictions. RNNs have been found to perform well on many tasks, but identifying the most relevant parameters for the task at hand and optimizing their performance can be difficult and time-consuming.

Two researchers at the University of Pennsylvania, Jason Kim and Dani S. Bassett, recently introduced an alternative approach to designing and programming RNN-based reservoir computers. It is inspired by how programming languages ​​work on computer hardware. This approach nature machine intelligencecan identify appropriate parameters for a particular network and program computations to optimize performance for the target problem.

“Whether it’s calculating chips or simulating multiple moves in a game of chess, we’ve always been interested in how the brain represents and processes information,” says Kim. told Tech Explore. “We were inspired by the success of recurrent neural networks (RNNs) in both modeling brain dynamics and learning complex computations. What if we could program RNNs like we would computers? Studies in control theory, dynamical systems, and physics have shown us that it’s not an impossible dream.”

The neural machine code introduced by Kim and Bassett was realized by decompiling the internal representation and dynamics of the RNN to guide the analysis of the input data. Their approach resembles the process of compiling an algorithm on computer hardware, requiring a detailed description of where and when individual transistors should be turned on and off.

“In RNNs, these operations are specified simultaneously with weights distributed across the network, and both neurons execute operations in parallel, saving memory,” explained Kim. “We use mathematics to define a set of operations (connection weights) that perform a desired algorithm (like solving an equation, simulating a video game, etc.) and then defining the algorithm running on the existing set of weights. A unique advantage of our approach is that no data or sampling is required, and that we can define not just one connection, but a space of connection patterns that exercise the desired algorithm.”

Researchers have demonstrated the framework’s benefits by using it to develop RNNs for a variety of applications, including virtual machines, logic gates, and AI-powered ping-pong video games. All of these algorithms were found to perform very well without requiring trial-and-error tuning of parameters.

“One of the notable contributions of our research is the paradigm shift in RNN understanding and research methods from data processing tools to full-scale computers,” said Kim. “This change shows that we can look at a trained RNN to see what problem it solves, and design an RNN to perform a task without training data or backpropagation. In effect, it allows the network to be initialized using hypothesis-driven algorithms instead of random weights.” Or using a pre-trained RNN, extracting the learned model directly from the RNN can. ”

The programming framework and neural machine code introduced by this team of researchers will soon be available for other teams to design better performing RNNs and easily tune their parameters. Kim and Bassett hope to eventually use their framework to create full-fledged software that runs on neuromorphic hardware. In their next work, they also plan to devise an approach to extract algorithms learned by trained reservoir computers.

“Neural networks are great at processing complex, high-dimensional data, but these networks tend to consume a lot of energy to run, and it is very difficult to understand what they have learned,” says Kim. said Mr. “Our work provides a stepping stone to directly decompile the trained weights into an explicit algorithm, which can be performed much more efficiently without the need for an RNN and is scientifically proven. understanding and performance can be further scrutinized.”

Bassett’s research group at the University of Pennsylvania is also working to replicate human mental processes and abilities using machine learning approaches, particularly RNNs. Their recent neural machine code may support their efforts in this area of ​​research.

“A second exciting research direction is to design RNNs to perform tasks characteristic of human cognition,” added Professor Dani S. Bassett, who oversees the research. “Using theories, models, or data-derived definitions of cognitive processes, we envision designing RNNs to exercise attention, proprioception, and curiosity. would like to understand the connectivity profiles that support such different cognitive processes.”