Blog

Aug 29, 2018 · newsletter

Hyperparameter Tuning and Meta-Interpretability: Track All Your Experiments!

From random forest to neural networks, many modern machine learning algorithms involve a number of parameters that have to be fixed before training the algorithm. These parameters, in contrast to the ones learned by the algorithm during training, are called hyperparameters. The performance of a model on a task given data depends on the specific values of these hyperparameters.

Hyperparamter tuning is the process of determining the hyperparameter values that maximize model performance on a task given data. The tuning of hyperparameters is done by machine learning experts, or increasingly, software packages (e.g., HyperOpt, auto-sklearn, SMAC). The aim of these libraries is to turn hyperparameter tuning from a “black art”, requiring expert expertise and sometimes brute-force search, into a reproducible science - reducing the need for expert knowledge, whilst keeping computational complexity at bay (e.g.,Snoek, Larochelle, & Adams; 2012).

Traditionally, hyperparameter tuning is done using grid search. Grid search requires that we choose a set of values for each hyperparameter and evaluate every possible combination of hyperparameter values. Grid search suffers from the curse of dimensionality; the number of joint values grows exponentially with the number of hyperparameters.

In 2012, Bergstra and Bengio showed that random hyperparameter search is more efficient than grid search, a perhaps counter-intuitive result. This is because only a few hyperparameters tend to really matter for the performance of a model on a task given data. Grid search tends to spend more time in regions of the hyperparameter space that are low-performing compared to random search. What’s more, random search allows one to easily add more experiments that explore even more sets of hyperparameter values without (expensive) adjustment of the grid (most recently, sequential approaches have shown great promise).

Grid (left) and random (right) search for nine experiments. With random search, all nine trials explore distinct values of the hyperparameters. Random search is more efficient. (Picture taken from Bergstra and Bengio, 2012)

If only a few hyperparameter values really matter, for a given model on a task given data, what are those parameters and what are their values? Current software libraries for hyperparameter tuning do not tend to discriminate important from unimportant hyperparameters and/or do not expose important parameters and their values. This limits insights into the workings of a model - which is important for a variety of reasons, as we explain in depth in our report on model interpretability; interpretability allows us to verify, for example, that a model gives high quality predictions for the right, and not the wrong, reasons.

A series of recent papers tackles this “meta-interpretability” problem: what hyperparameters matter for model performance on a task given data? In Hyperparameter Importance Across Datasets, Jan van Rijn and Frank Hutter first evaluate a model on a task given data and a set of randomly chosen hyperparameters to assess model performance. They then use these hyperparameters as inputs to a model, a so-called surrogate model, that they train to predict the oberved performance. Given a trained surrogate model, they predict model performance for hyperparameters not previously included in their experiments. Finally, they conduct an analysis of variance (ANOVA) to determine how much of the predicted model performance by the surrogate model can be explained by each hyperparameter or combination of hyperparameters. To draw conclusions across data sets, a more generalizable result, the authors repeat this procedure across several data sets. For random forests, for example, they find that only the minimum samples per leaf and the maximum numbers of features really matter. This finding is consistent with expert knowledge, which is great: it validates the method; we can use it to study more complex models for which we have no such intuition yet while it helps beginners to get started.

Using a related approach also based on surrogate models, Philipp Probst, Bernd Bischl, and Anne-Laure Boulesteix demonstrate that some default values of hyperparameters as set by software packages (e.g., scikit-learn) lie outside the range of hyperparameter values that tend to yield optimal model performance across tasks and data; we can use solutions to the meta-interpretability problem to define better default values, or to define prior distributions for even more efficient random hyperparameter search (Bergtra and Bengio sample from a uniform distribution which we can replace by a “more informed” distribution).

Within organizations, these results suggest that one should track and store the results of hyperparameter tuning - not only the set of parameters that result in the best performing model, but all results. These results can be used to train surrogate models that allow us insight into the importance of hyperparameter values and increase the efficiency of hyperparameter tuning by defining sensible default values (or distributions) for the classes of problems tackled by data teams at these organizations.

Read more

Newer
Sep 17, 2018 · post
Older
Aug 29, 2018 · newsletter

Latest posts

Nov 15, 2022 · newsletter

CFFL November Newsletter

November 2022 Perhaps November conjures thoughts of holiday feasts and festivities, but for us, it’s the perfect time to chew the fat about machine learning! Make room on your plate for a peek behind the scenes into our current research on harnessing synthetic image generation to improve classification tasks. And, as usual, we reflect on our favorite reads of the month. New Research! In the first half of this year, we focused on natural language processing with our Text Style Transfer blog series.
...read more
Nov 14, 2022 · post

Implementing CycleGAN

by Michael Gallaspy · Introduction This post documents the first part of a research effort to quantify the impact of synthetic data augmentation in training a deep learning model for detecting manufacturing defects on steel surfaces. We chose to generate synthetic data using CycleGAN,1 an architecture involving several networks that jointly learn a mapping between two image domains from unpaired examples (I’ll elaborate below). Research from recent years has demonstrated improvement on tasks like defect detection2 and image segmentation3 by augmenting real image data sets with synthetic data, since deep learning algorithms require massive amounts of data, and data collection can easily become a bottleneck.
...read more
Oct 20, 2022 · newsletter

CFFL October Newsletter

October 2022 We’ve got another action-packed newsletter for October! Highlights this month include the re-release of a classic CFFL research report, an example-heavy tutorial on Dask for distributed ML, and our picks for the best reads of the month. Open Data Science Conference Cloudera Fast Forward Labs will be at ODSC West near San Fransisco on November 1st-3rd, 2022! If you’ll be in the Bay Area, don’t miss Andrew and Melanie who will be presenting our recent research on Neutralizing Subjectivity Bias with HuggingFace Transformers.
...read more
Sep 21, 2022 · newsletter

CFFL September Newsletter

September 2022 Welcome to the September edition of the Cloudera Fast Forward Labs newsletter. This month we’re talking about ethics and we have all kinds of goodies to share including the final installment of our Text Style Transfer series and a couple of offerings from our newest research engineer. Throw in some choice must-reads and an ASR demo, and you’ve got yourself an action-packed newsletter! New Research! Ethical Considerations When Designing an NLG System In the final post of our blog series on Text Style Transfer, we discuss some ethical considerations when working with natural language generation systems, and describe the design of our prototype application: Exploring Intelligent Writing Assistance.
...read more
Sep 8, 2022 · post

Thought experiment: Human-centric machine learning for comic book creation

by Michael Gallaspy · This post has a companion piece: Ethics Sheet for AI-assisted Comic Book Art Generation I want to make a comic book. Actually, I want to make tools for making comic books. See, the problem is, I can’t draw too good. I mean, I’m working on it. Check out these self portraits drawn 6 months apart: Left: “Sad Face”. February 2022. Right: “Eyyyy”. August 2022. But I have a long way to go until my illustrations would be considered professional quality, notwithstanding the time it would take me to develop the many other skills needed for making comic books.
...read more
Aug 18, 2022 · newsletter

CFFL August Newsletter

August 2022 Welcome to the August edition of the Cloudera Fast Forward Labs newsletter. This month we’re thrilled to introduce a new member of the FFL team, share TWO new applied machine learning prototypes we’ve built, and, as always, offer up some intriguing reads. New Research Engineer! If you’re a regular reader of our newsletter, you likely noticed that we’ve been searching for new research engineers to join the Cloudera Fast Forward Labs team.
...read more

Popular posts

Oct 30, 2019 · newsletter
Exciting Applications of Graph Neural Networks
Nov 14, 2018 · post
Federated learning: distributed machine learning with data locality and privacy
Apr 10, 2018 · post
PyTorch for Recommenders 101
Oct 4, 2017 · post
First Look: Using Three.js for 2D Data Visualization
Aug 22, 2016 · whitepaper
Under the Hood of the Variational Autoencoder (in Prose and Code)
Feb 24, 2016 · post
"Hello world" in Keras (or, Scikit-learn versus Keras)

Reports

In-depth guides to specific machine learning capabilities

Prototypes

Machine learning prototypes and interactive notebooks
Notebook

ASR with Whisper

Explore the capabilities of OpenAI's Whisper for automatic speech recognition by creating your own voice recordings!
https://colab.research.google.com/github/fastforwardlabs/whisper-openai/blob/master/WhisperDemo.ipynb
Library

NeuralQA

A usable library for question answering on large datasets.
https://neuralqa.fastforwardlabs.com
Notebook

Explain BERT for Question Answering Models

Tensorflow 2.0 notebook to explain and visualize a HuggingFace BERT for Question Answering model.
https://colab.research.google.com/drive/1tTiOgJ7xvy3sjfiFC9OozbjAX1ho8WN9?usp=sharing
Notebooks

NLP for Question Answering

Ongoing posts and code documenting the process of building a question answering model.
https://qa.fastforwardlabs.com

Cloudera Fast Forward Labs

Making the recently possible useful.

Cloudera Fast Forward Labs is an applied machine learning research group. Our mission is to empower enterprise data science practitioners to apply emergent academic research to production machine learning use cases in practical and socially responsible ways, while also driving innovation through the Cloudera ecosystem. Our team brings thoughtful, creative, and diverse perspectives to deeply researched work. In this way, we strive to help organizations make the most of their ML investment as well as educate and inspire the broader machine learning and data science community.

Cloudera   Blog   Twitter

©2022 Cloudera, Inc. All rights reserved.