LowPy: Simulation Platform For Machine Learning Algorithm Realization In Neuromorphic RRAM Based Processors (Preprint)
Abstract
A novel compilation of non-ideal characteristics which accompany hardware realizations of machine learning algorithms in the form RRAM-based neuromorphic ASIC processors is presented within a convenient, simple, and powerful simulation library named LowPy for use with Python. Simulations results are shown for four different networks; SLP,MLP, CNN, and LSTM. Each is subjected to six different GPU-accelerated nonideality functions backed by experimentally gathered data, as well as data provided by external research. Of the six nonideality functions, a spread of selected parameters is chosen such that any performance impacts of the algorithm are easily observed across several orders of magnitude. Main aspects differentiating LowPy from other neuromorphic simulation platforms include functions not yet implemented by other platforms, the most non-ideal functions provided by any platform known by the author to date, an event-driven architecture that provides the user full control over which and when nonideality functions are executed during training and testing, as well as its ability to wrap around an existing well-documented, popular, GPU-accelerated machine learning library.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2021
- Accession Number
- AD1136893
Entities
People
- Andrew J. Ford
Organizations
- University of Cincinnati