ML space exploration (ARM)

Tanszéki konzulens: 
A munkatárs fényképe
mesteroktató
Szoba: IE336
Tel.:
+36 1 463-2066
Email: szanto (*) mit * bme * hu
Külső konzulens: 
Gyimesi Tamás

A kiírás adatai

A téma státusza: 
Aktív (aktuális, lehet rá jelentkezni)
Kiírás éve: 
2021
A kiírás jellege: 
önálló labor, szakdolgozat/diplomaterv

Background and goals

Machine Learning applications have skyrocketed in recent years in all areas of
industry due to their relatively easily accessible resources needs and generic
usefulness. Though beeing widespread and popular, little effort has bee put into
customizing their resource needs to the field of actual application. The very same
processes are being observed in digital logic verification, where ML-based
approaches are gaining ground, not clearly setting requirements and articulating
expected benefits leading to enormous efforts put into innovation with
unsatisfactory results, wasting resources. Here at ARM we are starting to use ML
assisted verification techniques hoping to decrease required simulation cycles and
thus computing capacity and development times of projects while achieving same or
better metrics.
 
The goal of this research is to find relationships between module (and verification)
complexity, number of test cases and expected metrics and the size of the machine
learning network and training set with used weighting approach. The outcome
should be an estimating utility or library that advises on neural network size and
weighting approach.

Main tasks of the student:

  • Get familiar ML/AI technologies
  • Get familiar with functional verification principles, methods and metrics.
  • Review application of ML engines in verification, their requirements and benefits
  • Implement a tool that demonstrates evolution of parameter sets and error function results and estimates required iteration counts
    • by changing NN sizes (layers, layer sizes)
    • by changing target metrics
    • by changing number of input parameters (model size)
    • by changing the learning model (method and feedback strength)
  • Find and implement an algorithm that can discover unused or redundant parameters of model and suggest reductions
  • Invent a method that would automatically partition the parameter space and reformulate the problem to the sub-spaces creating less complex targets if that consumes less resources (time, computation effort, iterations)

Professional requirements:

  • Verilog / SystemVerilog knowledge
  • Knowledge of ML/AI principles

A téma kiírója az ARM Magyarország, a jelentkezés elfogadását interjú előzi meg.

A témára MSc hallgatók jelentkezését várjuk.

© 2010-2022 BME MIT | Hibajelentés | Használati útmutató