Fedot

View the Project on GitHub ITMO-NSS-team/FEDOT.Docs

State-of-the-art AutoML solutions: a brief overview

We cannot deny that there are a lot of well-implemented AutoML frameworks that are quite popular.

We tried to summarize the main advantages and disadvantages of the state-of-the-art solutions to the breif table:

Framework Pipeline Optimisation algorithm Input data Ensembling Scaling Additional features Source
TPOT Variable GP Tabular - Multiprocessing, Rapids Code generation https://github.com/EpistasisLab/tpot
H2O Fixed Grid Search Tabular, Texts + Hybrid - https://docs.h2o.ai/h2o/latest-stable/h2o-docs/automl.html
AutoSklearn Fixed SMAC Tabular + - - https://github.com/automl/auto-sklearn
ATM Fixed BTB Tabular - Hybrid - https://github.com/HDI-Project/ATM
FEDOT Variable GP + GridSearch Tabular, Timeseries + - Composite models https://github.com/nccr-itmo/FEDOT
AutoGluon Fixed Fixed Defaults Tabular, Images, Texts + - NAS, AWS integration https://github.com/awslabs/autogluon
LAMA Fixed Optuna Tabular + - Profiling https://github.com/sberbank-ai-lab/LightAutoML
NNI Fixed Bayes Tabular, Images ? Hybrid, Kubernetes NAS, WebUI https://github.com/microsoft/nni

Why we want better AutoML?​

  1. «Machine learning from scratch» is too hard – a lot of time and resources is required (https://ai.googleblog.com/2020/07/automl-zero-evolving-code-that-learns.html)​ Expert knowledge should be involved (i.e. as initial assumption);​
  2. Pipelines with fixed structure – suitable for simple, routine tasks but not very effective for real problems;​
  3. Raw data should be processed – preprocessing is not well-supported by modern AutoML;​
  4. Modelling for multimodal data (i.e. table + images + time series + text) is not well-automated now for common tasks;​
  5. Specialized for different modelling problems – current solutions are mostly over-specialised;​
  6. Modular pipelines – composite structures instead of complex neural networks / End-To-End Learning for better controllability;