Neural Circuit Policies Enabling Auditable Autonomy by Lechner, M., Hasani, R., Amini, A., Henzinger, T. A., Rus, D., & Grosu, R. (2020)

source
(Lechner et al. 2020)
tags
Neural networks

Summary

This article introduces a type of RNN called Neural Circuit Policies (NCP). This architecture is said to be inspired from the wiring diagram of the C. elegans nematode.

The main building block is a Recurrent neural network called liquid time constant (LTC) introduced in (Hasani et al. 2020).

LTC Neurons

These neurons are bio-inspired. For a given neuron in state x_i(t), the continuous temporal evolution is described by an ODE: $\dot{x}_i = - \left(\frac{1}{\tau_i} + \frac{w_{ij}}{C_{m_i}} \sigma_i(x_j) \right) x_i + \left( \frac{x_{\text{leak}_i}}{\tau_i}+ \frac{w_{ij}}{C_{m_i}} \sigma_i(x_j) E_{ij} \right)$

where $$w_{ij}$$ is a synaptic weight from neuron $$i$$ to $$j$$, $$C_{m_i}$$ is the membrane capacitance of the neuron $$i$$, and $$\sigma_i(x_j(t)) = 1/\left( 1 + e^{-\gamma_{ij}(x_j-\mu_{ij})}\right)$$. $$x_{\text{leak}_i}$$ is called the resting potential and $$E_ij$$ is a reversal synaptic potential.

More details about the biological analogy for those quantities are given in (Lechner et al. 2020; Hasani et al. 2020).

Results

The authors reports good results on a car steering task compared to many RNN architectures, although not much better than LSTMs. The main advantages of this method according to the authors are:

• Better noise robustness
• Better interpret-ability
• Less parameters needed