Cortex Specification Summary

A Brief Summary of the CognitionLab Cortex-Hopfield Specification

- Hopfield Network Overview
Hopfield networks are members of a broader class known as auto-associative neural networks. These are networks that can be trained to return to desired states from initial states which are close to yet different from those desired. They associate network states with themselves, hence their name.

Auto-associative networks consist of one layer of neural elements that are all interconnected, except that self-connections are disallowed. The neural elements are nonlinear, with states bounded from zero to one; so called two-state-neurons. Any state, whether initial or desired, will be represented by some pattern of zeros and ones over the units. Recurrence and nonlinearity lie at the heart of auto-associative network behavior. Recurrence allows desired states to emerge via interactions among the units, and the nonlinearity prevents the whole network from running away, and instead encourages the network to settle into a stable state. The auto-associative network will, in most cases, relax from any initial state into the desired state closest to it.

- Cortex-Hopfield Summary
The Cortex-Hopfield specification is a proprietary application layer used to normalise Hopfield network training data and link it with processed objects, allowing arbitrary input strings of any predefined fixed length and any syntax or encoding to return fully flexible results loaded with data, protocols and methods, and even unique functions built to requirement.

This makes a Cortex network instantly deployable to an infinite number of diverse tasks with no modification of core code, allowing modular expansion and cross-compatibility between Cortex network data and open source Hopfield implementations, allowing Cortex-based networks to be deployed on any number of platforms with little to no modifications, and no need to use any Cortex-specific interpreter. The Cortex Application Layer provides an extremely flexible training structure, but the resulting data can be exported and deployed without any CognitionLab technology used in deployment if needs be.

The implementation of the Cortex specification we use at CognitionLab is written in JavaScript and the application layer is loaded "on top" of the open-source node.js library Synaptic by pseudonym "cazala", and to him we are extremely grateful for the foundations Cortex was developed with, however the specification is easily ported to other languages and libraries, and is by no means fixed to one environment.

Cortex networks can be used alone, or in tandem with other machine learning technologies such as Tensor networks to provide actions or data models in response to the output from more specialised deep learning, giving it the potential to be used as a data-normalising API for less tangible data sources.

Cortex opens up the possibilities of machine cognition and deep learning by making it intuitive and simple to apply to any task, and removing many of the barriers to developing new technologies that make use of the vast capabilities of computational neurobiology.

 - Erin MacDonald, CognitionLab LTD. CEO