SA/MA: Hierarchical sequence learning on mixed-signal neuromorphic hardware
Hierarchical chunking models (HCMs) learn hierarchically structured sequential data online. Study-ing them and simulating them can lead to insights in how mechanisms of working memory and natural intelligence work, and to potential solutions for practical applications in the real world. However, serial implementations of HCM neural network models on von Neumann computers consume large amounts of memory and computation time.
Spiking neural networks (SNN) implemented on mixed-signal neuromorphic hardware are intrinsically parallel and could potentially accelerate the computation through recurrent connections and spike-timing-dependent plasticity. In this project the student will investigate implementations of HCM compatible with the SNN hardware available (exclusively) at our institute. The results could be used for neuromorphic temporal sequence recognition and robotic control in the future.
You will implement the HCM model on our dynamic neuromorphic asynchronous processor (DYANP-SE2), including the following stages:
1. Pilot implementation of learning size-2 chunks on spiking neuromorphic hardware.
2. Using the design architecture of Stage 1 as a building block for higher order sequential chunk representations, and eventually scale up the neuromorphic chunk representation to learn arbitrary sequential hierarchies.
You will be mostly using the python API to control the chip, so basic python skill is expected.Previous knowledge about spiking neural network especially neuromorphic engineering is recommended.
This is a project conducted at the Institute of Neuroinformatics at UZH-ETH in collaboration with the Max Planck Institute for Biological Cybernetics in Tübingen, Germany.
Chenxi Wu: chenxi (at) ini.uzh.ch & Prof. Dr. Giacomo Indiveri: giacomo (at) ini.uzh.ch
Shuchen Wu: shuchen.wu (at) tue.mpg.de & Dr. Eric Schulz: eric.schulz (at) tue.mpg.de