DiMO: Differentiable Model Optimization

Brando Miranda

Permalink

https://hdl.handle.net/2142/109086 Copy

Description

Title

DiMO: Differentiable Model Optimization

Author(s)

Brando Miranda

Issue Date

2020-03-01

Keyword(s)

• meta-learning
• learning to learn
• NAS
• automl
• machine learning

Abstract

I am interested in automated methods for model selection - architecture selection, hyperparameter selection and optimizer selection. To this end, I am developing DiMO (Differentiable Model Optimization): a novel technique for 1) model optimization (i.e. model selection for a fixed dataset by improving the selected architecture, hyperparameters and custom optimizers) and 2) meta-learning (i.e. automatic selection of model architectures for new tasks not observed during training). The main idea of DiMO is to develop a framework (flexible for both meta-learning and model optimization) that models the entire machine learning pipeline and minimize that objective directly with minimal hacks and tricks. The hope is that by optimizing the objective directly - with minimal modifications - performance of constructed models can be optimal and the search efficient, due to it's differentiable formulation.

Type of Resource

text

Language

en

Permalink

http://hdl.handle.net/2142/109086

Owning Collections