IRENA (Invariant Representations Extraction in Neural Architectures)

Team NeuroTHIx codebase.

1st Place at Merck Future of AI Research Challenge.

https://app.ekipa.de/challenge/future-of-ai/about

Team:

NeuroTHIx

Team members:

Du Xiaorui, Erdem Yavuzhan, Cristian Axenie

Abstract

In our proposed solution we aim at a combination of Neuroscience principles, Abstract thinking and Prototyping, towards a solution aiming at bringing the efficiency and robustness of biological intelligence to technical systems that solve real-world problems. We start from the proposed Neuroscientific Research Challenges and propose a novel model and system capable of learning invariant representation.

Core idea

Using cortical maps as neural substrate for distributed representations of sensory streams, our system is able to learn its connectivity (i.e., structure) from the long-term evolution of sensory observations. This process mimics a typical development process where self-construction (connectivity learning), self-organization, and correlation extraction ensure a refined and stable representation and processing substrate. Following these principles, we propose a model based on Self-Organizing Maps (SOM) and Hebbian Learning (HL) as main ingredients for extracting underlying correlations in sensory data, the basis for subsequently extracting invariant representations.

Quick Start

1-Download the whole files from the link:

https://drive.google.com/open?id=1I6a21i7N86tNrdttCgA19UEJR9vV_SUQ

Test step:

1-Set train = 0 in config_FGV.txt(config_IG.txt)

2-Run

python3 trainFGV_dense.py

python3 trainIG.py

Note:

After runing, you will get results in test_results folder

Train step:

1-Set train = 1 in config_FGV.txt(config_IG.txt)

2-Run

python3 trainFGV_dense.py

python3 trainIG.py

Note:

After runing, you will get .pkl files in the current folder.