• *Physics* 16, 108

A new theory allows researchers to determine the ability of arbitrarily complex neural networks to perform recognition tasks on data with complex structure.

Every day, our brain recognizes and discriminates against the many thousands of sensory signals it encounters. Today’s best AI models, many of which are inspired by the neural circuits in the brain, have similar capabilities. For example, the so-called deep convolutional neural networks used for object recognition and classification are inspired by the layered structure of the visual cortex. However, scientists have yet to develop a comprehensive mathematical understanding of how biological or artificial intelligence systems achieve this recognition capability. Now SueYeon Chung of the Flatiron Institute in New York and her colleagues have developed a more detailed description of the connection of the geometric representation of objects in biological and artificial neural networks to the performance of networks in classification tasks [1] (Fig.1). The researchers show that their theory can accurately estimate the performance capability of an arbitrarily complex neural network, a problem that other methods have struggled to solve.

Neural networks provide coarse-grained descriptions of the complex circuitry of biological neurons in the brain. They consist of highly simplified neurons that signal to each other via synapses, connections between pairs of neurons. The strengths of synaptic connections change as a network is trained to perform a particular task.

During each phase of an activity, groups of neurons receive input from many other neurons in the network and fire when their activity exceeds a certain threshold. This activation produces a pattern of activity, which can be represented as a point in a high-dimensional state space where each neuron corresponds to a different dimension. A collection of activity patterns matching a specific input forms a “manifold” representation in that state space. The geometric properties of the multiple representations in neural networks depend on the distribution of information in the network and the evolution of the multiple representations during an activity is modeled through the algorithms that train the network to perform the specific activity.

The manifold geometries of a network also limit the network’s ability to perform tasks, such as invariant object recognition, the ability of a network to accurately recognize objects regardless of variations in their appearance, including size, location, or background (Fig. 2). In an earlier attempt to understand these constraints, Chung and a different group of colleagues studied simple binary classification tasks, those in which the network must sort stimuli into two groups based on some classification rules [2]. In such activities, the capacity of a network is defined as the number of objects it can correctly classify if the objects are randomly assigned categorization labels.

For networks where each object corresponds to a single point in the state space, a single-level network with *No* neurons can classify 2*No* objects before the classification error becomes equivalent to that of the random guess. The formalism developed by Chung and his colleagues allowed them to study the performance of complex deep (multilayered) neural networks trained for object classification. Constructing the multiple representations from the images used to train such a network, they found that the mean radius and number of manifold dimensions estimated from the data decreased sharply in deeper layers of the network than in shallower layers. This decrease has been accompanied by an increased classification capacity of the system [2–4].

This previous study and others, however, did not consider the correlations between different object representations when calculating network capacity. Object representations in biological and artificial neural networks are known to have complex correlations, which arise from structural features in the underlying data. These correlations can have important consequences for many tasks, including classification, because they are reflected in different levels of similarity between so-called class pairs in the neural space. For example, in a network tasked with classifying whether an animal was a mammal, the multiple representations of the dog and wolf would be more similar than those of an eagle and a hawk.

Now Chung’s group has generalized the computation of deep neural network performance capability to include correlations between classes of objects [1]. The team derived a set of self-consistent equations that can be solved to give the network the capability of a system with homogeneous correlations between the so-called axes (the dimensions along which the manifold varies) and the centroids (the centers of the manifolds) of different manifolds . The researchers show that axis correlations between manifolds increase performance ability, while centroid correlations push manifolds closer to the origin of the neural state space, decreasing performance ability.

In recent years, the study of neural networks has seen many interesting developments and more and more data analysis tools are being developed to better characterize the geometry of the representations obtained from neural data. The new results make a substantial contribution in this area, as they can be used to study the properties of learned representations in networks trained to perform a wide variety of tasks in which correlations present in the input data can play a crucial role in the learning and performance. These tasks include those related to motor coordination, natural language, and probing the relational structure of abstract knowledge.

## References

- AJ Wakhloo
*et al.*Linear classification of neural varieties with correlated variability, Phys. Rev. Lett.**131**027301 (2023). - SY Chung
*et al.*Classification and geometry of general perceptual manifolds, Phys. Reverend X**8**031003 (2018). - E. Gardner, The interaction space in neural network models, J. Fis. A: Mathematics. gen.
**21**257 (1988). - U. Cohen
*et al.*Separability and geometry of object manifolds in deep neural networks, Nat. Common. (2019).

## About the author

## Thematic areas

#Performance #capacity #complex #neural #network

Image Source : physics.aps.org