Coding the Future: Pennsylvania Researchers Transform the World of Neural Networks
It’s time for a change in the world of AI.
Neural networks are a powerful machine learning tool that allows you to analyze and predict complex data. However, in order for the neural network to work effectively, it is necessary to select the optimal parameters that depend on the specific task. This can be a laborious and lengthy process.
Researchers at the University of Pennsylvania suggested a new approach to designing and programming neural networks, inspired by the principles of programming languages. This approach, published in the journal Nature Machine Intelligence, allows you to determine the appropriate parameters for a given network, program its calculations and optimize its performance to solve targeted problems.
“We’ve always been interested in how the brain represents and processes information, whether it’s calculating tips or simulating multiple moves in chess,” said Jason Kim, one of the authors of the paper. “We were inspired by the success of Recurrent Neural Networks (RNNs) for both modeling brain dynamics and teaching complex computations. Based on this inspiration, we asked ourselves a simple question: what if we could program RNNs in the same way that we do with computers? Previous work in control theory, dynamical systems and physics has told us that this is not an impossible dream.”
Kim and his colleague Dani Bassett developed a custom neural machine code that they obtained by decompiling the internal representations and dynamics of RNNs to guide their analysis of input data. Their approach resembles the process of compiling an algorithm on computer hardware, which consists in specifying the places and times when individual transistors should be turned on and off.
“In RNN, these operations are simultaneously given by weights distributed throughout the network, and neurons both perform operations in parallel and store memory,” Kim explained. “We use math to define a set of operations (link weights) that will perform a desired algorithm (e.g. solve an equation, simulate a video game) and extract an algorithm that runs on an existing set of weights. The unique advantages of our approach are that it does not require data or sampling, and that it defines not only a single connection, but also a space of connection patterns that execute the desired algorithm.”
The researchers demonstrated the benefits of their approach by using it to develop RNNs for a variety of applications, including virtual machines, logic gates, and an AI-powered ping-pong video game. These algorithms have shown excellent results without requiring trial and error selection of parameters.
“One of the notable contributions of our work is a paradigm shift in how we understand and study RNNs, from data processing tools to full-fledged computers,” said Kim. “This shift means we can examine a trained RNN and know what problem it solves, and we can design an RNN to perform tasks without training data or backpropagation. In practice, we can initialize our networks with a hypothesis driven algorithm rather than random weights or a pretrained RNN, and we can directly extract the learned model from the RNN.”
The software platform and neural machine code presented by this group of researchers may soon be used by other teams to design higher performance RNNs and easily tune their parameters. Kim and Bassett eventually hope to use their platform to build full-fledged software that runs on neuromorphic hardware. In their next research, they also plan to develop an approach to extract the algorithms learned by trained tank computers.
“Although neural networks are excellent at processing complex and high-dimensional data, these networks require a lot of energy to operate, and understanding what they have learned is an exceptionally difficult task,” said Kim. “Our work provides a stepping stone to directly decompile and translate the trained weights into an explicit algorithm that can perform much more efficiently without the need for an RNN, and be further analyzed in terms of scientific understanding and performance.”
Bassett’s research team at the University of Pennsylvania is also working on using machine learning approaches, specifically RNNs, to replicate human mental processes and abilities. The neural machine code they recently created could help them in this area of research.