This work presents a very simple and biologically plausible model for cognitive map formation. We use primitive neurons and a discreet time simulation that we connect with biological evidence through rate firing model. Since cell assemblies are natural candidate for working memory we use them as substrate for information processing. Simple hand tuning reveals already the potential of this system. While being able to produce working memory behavior, the system remains scalable and efficient. Furthermore, we propose an unsupervised online learning algorithm. When the system encounters inputs, it dynamically creates the corresponding cell assemblies. Those cell assemblies can be stabilized in the system through a retroaxonal feedback. The retroaxonal feedback hypothesis states that when a neuron repeatedly fires for a given input, it would send back a feedback suggesting strengthening of its connection with that input. Results show that the obtained cell assemblies exhibit proper behavior reminiscent to the one observed with hand tuned system.
Recurrent Neural Network, Working Memory, Cognitive Map, Cell Assembly, Unsupervised Learning, Retroaxonal Feedback