In search of a more general theory of computational science, Stanford University’s John Ioannidis and a team of researchers report their new theory of the structure and function of neural networks

Stanford University researchers have developed a theory of how neurons are organized to perform computations.

They have developed the theory, called computational biology, which can be applied to a range of scientific and technological problems.

They will present their findings at the annual meeting of the American Physical Society (APS) in Orlando, Florida, next week. 

They are the first to develop a general framework for understanding how networks work.

In a paper published in the journal Science, the Stanford researchers, led by PhD student John Ioaniadis, describe the theory and explain how it applies to their own work on neural networks. 

In a nutshell, the theory proposes that network activity is a function of information processing speed. 

To find out more, Ioannides and his team created a network that consists of several tens of thousands of neurons.

The researchers trained this network to perform a simple task in a series of steps.

For example, the network might be trained to detect a new image on a screen. 

The task involves a series and then two tests.

One is to determine whether the network has enough neurons to perform the task.

The second test is to compare the network’s performance to a previous set of tasks. 

A network that has trained successfully will perform the first task.

A network that does not perform well will have to learn from previous training mistakes. 

 When training is done, the neurons will be reorganized in a network’s structure.

This reorganization is what is called network-level learning. 

If the network is organized in a way that it learns better from previous errors, then this reorganization will allow it to perform better. 

This reorganization can be done either by adding more neurons to the network or by taking some of the network away from it. 

These network-scale changes can be useful for performing tasks in real time, such as analyzing data. 

However, the authors say that there is an additional benefit. 

For example, if the network does not reorganize properly and instead keeps working on the task at a certain speed, the learning process can be slowed down. 

It could be that this reorganizing process can also be used to solve other problems, such to build robots or to automate software tasks. 

 For their study, the researchers used a network of more than one hundred thousand neurons. 

Their model predicts that the reorganization process is not restricted to the neurons but that it can also affect other parts of the neural network. 

When a network learns to perform certain tasks, for example by performing a task faster than a previous time, it is called a neural network trained with network-based structure. 

One problem that the authors face is how to describe this reorganized structure.

For instance, they are currently unable to describe it in terms of the specific parts of a neuron that are changed or whether the reorganized parts are related to the function of the neuron. 

So the model has several limitations. 

Its most important limitation is that the model does not account for other factors that might affect the structure of the neurons. 

 To explore this problem, Ioaniads and his colleagues constructed an artificial neural network that was able to perform both tasks.

They then asked the network to be trained on a computer task, to compare how well it performed with a set of similar tasks that it had never been trained on, and to test whether the neural system can learn from past training mistakes using data that was never seen before. 

Each time the neural model learned a new task, it performed it in parallel, but the researchers did not control for the effects of other variables. 

After training the network, the team then used this new network to run a series-of-2 tests. 

Then, after a further set of trials, they repeated the training process to see if the model improved over the previous trial. 

As expected, the neural-network-trained network improved more than the neural one that had never trained on the computer task. 

Finally, the same team tested whether the new neural network could solve a number of computational tasks that previously required the same network structure.

The results indicated that the network could do these tasks as well as the network that had not trained. 

According to the Stanford team, the model is “a powerful tool to develop an understanding of how network dynamics are generated and to understand the underlying principles underlying the way networks are organized.” 

 To further explore this topic, the group plans to explore how networks of neurons interact and adapt to different types of problems. 

While this is a promising model for understanding the basic properties of networks, the research team hopes to use it to test new methods and develop new ideas in the field of artificial intelligence. 

Source: Science, Nature, Scientific American, APS, John Ioannis, Stanford University