Sensory Coding with Dynamically Competitive Networks

M. I. Rabinovich,R. Huerta,A. Volkovskii,Henry D. I. Abarbanel,G. Laurent

DOI: https://doi.org/10.48550/arXiv.q-bio/0309033

1999-05-24

Abstract:Studies of insect olfactory processing indicate that odors are represented by rich spatio-temporal patterns of neural activity. These patterns are very difficult to predict a priori, yet they are stimulus specific and reliable upon repeated stimulation with the same input. We formulate here a theoretical framework in which we can interpret these experimental results. We propose a paradigm of ``dynamic competition'' in which inputs (odors) are represented by internally competing neural assemblies. Each pattern is the result of dynamical motion within the network and does not involve a ``winner'' among competing possibilities. The model produces spatio-temporal patterns with strong resemblance to those observed experimentally and possesses many of the general features one desires for pattern classifiers: large information capacity, reliability, specific responses to specific inputs, and reduced sensitivity to initial conditions or influence of noise. This form of neural processing may thus describe the organizational principles of neural information processing in sensory systems and go well beyond the observations on insect olfactory processing which motivated its development.

Neurons and Cognition

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is to understand how the complex spatio - temporal neural activity patterns in the insect olfactory system encode and process sensory information, especially odor information. Specifically, the paper aims to explore the following questions:

1. **Theoretical framework of the dynamic competition network**: How to explain the experimentally observed phenomenon that odors are represented by rich spatio - temporal neural activity patterns, and these patterns are specific and reliable under repeated stimulation.

2. **Advantages of dynamic encoding**: What are the advantages of such complex distributed spatio - temporal behaviors for stimulus representation or classification? Can it provide greater information capacity, reliability, reduced sensitivity to initial conditions or noise, etc.?

3. **Dynamic competition mechanism**: How is the representation of inputs (such as odors) achieved through the dynamic competition between neuron populations? How is each pattern generated through the dynamic movement within the network, rather than through the "winner - takes - all" mechanism?

### Specific objectives of the paper

- **Propose a dynamic competition model**: The author proposes a "dynamic competition" paradigm, in which inputs (such as odors) are represented by internally competing neural assemblies. Each pattern is the result of dynamic movement within the network, without involving a "winner".

- **Simulate biological experimental phenomena**: The model can produce results very similar to the experimentally observed spatio - temporal patterns and has many desirable pattern - classifier characteristics, such as large information capacity, reliability, specific responses to specific inputs, and low sensitivity to the influence of initial conditions or noise.

- **Explain how the brain processes complex sensory information**: Through the dynamic competition model, the paper attempts to describe the organizational principles of neural information processing, not only limited to the insect olfactory system, but also potentially applicable to other sensory systems.

### Key features of the dynamic competition model

- **Transient dynamics**: The transient dynamics of the system are determined by the input itself, and the trajectory in the state space defines a unique global attractor related to a specific input.

- **Heteroclinic orbits**: In the state space, continuous competition is manifested in the trajectories close to heteroclinic orbits, which are presented as a series of "ribs" connecting unstable fixed points or limit cycles.

- **Noise resistance and robustness to initial conditions**: Even in the presence of noise and wide variations in initial conditions, the dynamic competition network still exhibits significant noise resistance and stability.

### Examples of mathematical formulas

The mathematical models mentioned in the paper include the following equations:

1. **Average neuron model**:
\[
\frac{dY_i(t)}{dt} = Y_i(t)\left\{\sigma(E_i(t)+S_i(t))-\sum_{k = 1}^3\rho_{ik}Y_k(t)\right\}
\]
where:
\[
\sigma(x)=1-\frac{2}{1 + e^{10(x - 0.4)}}
\]
\[
E_i(t)=g_e\sum_{k\neq i}Y_k(t)
\]

2. **FitzHugh - Nagumo model**:
\[
\epsilon\frac{dU_i(t)}{dt}=f(U_i(t))-I_i(t)-\gamma_1(U_i(t)-\tilde{U}_I)v_{i - 1}(t)+S_i(t)
\]
\[
\frac{dI_i(t)}{dt}=U_i(t)-bI_i(t)+a
\]
\[
\alpha_1\frac{dv_i(t)}{dt}=\sigma(U_i(t))-v_i(t)
\]

Through these models, the paper shows how to achieve complex spatio - temporal neural activity patterns through the dynamic competition mechanism and explains its potential advantages in sensory information processing.