Hyperparameter self-tuning for data streams

Question

1. What contributions have the authors mentioned in the paper "Hyperparameter self-tuning for data streams" ?

2. What are the future works in "Hyperparameter self-tuning for data streams" ?

3. What is the method for optimisation of hyperparameters?

4. What is the main limitation of the approaches available in the literature?

Accepted Answer

In this article, the authors present SSPT, an extension of the Self Parameter Tuning ( SPT ) optimisation algorithm for data streams.

Accepted Answer

Future work will address: ( i ) the ability to select not only hyperparameters but also models ; ( ii ) the comparison of SPT with other emergent AutoML optimisation algorithms ; and ( iii ) the analysis of the robustness of SPT with different types of drifts.. Further research will be required to assess the behaviour of SPT in face of these challenges.

Accepted Answer

It relies on incremental k-fold cross-validation to perform the incremental tuning of the support vector machine hyperparameters.

Accepted Answer

The main limitation of approaches available in the literature is that they require offline processing, with train and validation stages, making them applicable to batch learning problems, but not in streaming scenarios.

Accepted Answer

The experiments were250run on an Intel Xeon CPU E5-2680 2.40 GHz Central Processing Unit (CPU), 32 GB DDR3 Random Access Memory (RAM) and 1 TB hard drive, running the Linux Ubuntu 16.04 OS.

Accepted Answer

Kar et al. [20] apply an exponentially decaying centrifugal force to all vertices of the Nelder-Mead algorithm to80obtain better objective values.

Accepted Answer

The authors argue that the optimisation of hyperparameters in algorithmsthat learn from data streams requires a specialised AutoML approach.

Accepted Answer

In terms of online hyperparameter tuning algorithms, there are few works [17, 23–26] that use support vector machines together with batch processing, gradient solutions combined with brute force or genetic algorithms to opti-95mise hyperparameters.

Accepted Answer

The extended algorithm – Hyperband – regularly adjusts the trade-off between the available70computational budget and the number of candidates.

Accepted Answer

Lin et al. [17] propose a forward hyper-gradient110optimisation method that uses the gradient of the response of a function to adjust the hyperparameters [27].

Accepted Answer

This new implementation only requires a single-pass over the data to optimise the set of parameters, making it more versatile and less computationally expensive.

Accepted Answer

Hyperparameters are used to tune certain aspects of an algorithm that have155direct impact on its performance, depending on the data it learns from [38].

Accepted Answer

Regardless of the version, SPT uses obviously350more memory than the default algorithm, since it requires enough memory to hold ten (three learning and seven experimental) models during exploration2.

Accepted Answer

Li et al. [15] extend the Successive Halving multi-armed bandit algorithm [18] to continuously reduce the number of configuration candidates, by focusing on the most promising ones.

Accepted Answer

The size of the sample increases when the variance of the error grows, and decreases otherwise, down to a lower bound of30 samples, to avoid using small-sized samples.

Accepted Answer

The problem with the automation of the hyperparameter tuning process can be stated as follows: given a machine learning algorithm A with a default hyperparameter configuration c, represented by Ac, and a data stream S, thegoal is to find an optimal algorithm configuration A∗c where A ∗ c yields higher160performance score than Ac.

Accepted Answer

SPT updates the learning and experimental models incrementally with each example in the data stream until the learning models converge to an optimal configuration.

Accepted Answer

Regarding existing AutoML tools, the authors focus on those that support multiplemachine learning tasks, automatically select algorithms and optimise hyperpa-115rameters.

Accepted Answer

In accuracy ranking, SSPT is only outperformed by the grid and random search offline techniques, in the case of the Postures data set.

Hyperparameter self-tuning for data streams

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What contributions have the authors mentioned in the paper "Hyperparameter self-tuning for data streams" ?

2. What are the future works in "Hyperparameter self-tuning for data streams" ?

3. What is the method for optimisation of hyperparameters?

4. What is the main limitation of the approaches available in the literature?

5. How many xeon CPUs were used in the experiments?

6. How do they achieve better objective values?

7. What is the main argument against the AutoML approach?

8. What is the way to tune the hyperparameters?

9. How does the extended algorithm adjust the trade-off between the number of candidates?

10. What is the optimisation method used by Lin et al.?

11. What is the new implementation of the auto-learning module?

12. What is the purpose of the hyperparameter tuning module?

13. How many models can be accessed by the default algorithm?

14. How do they extend the Successive Halving multi-armed bandit algorithm?

15. How does the size of the sample change when the error increases?

16. What is the problem with the automation of the hyperparameter tuning process?

17. How does the algorithm update the learning and experimental models?

18. What are the main features of AutoML?

19. What is the ranking of SSPT?

Figures

Citations

Electrical Load Forecasting Using LSTM, GRU, and RNN Algorithms

5G Technology in Healthcare and Wearable Devices: A Review

A survey on machine learning for recurring concept drifting data streams

Unsupervised concept drift detection for multi-label data streams

Unsupervised online anomaly detection in Software Defined Network environments

References

Scikit-learn: Machine Learning in Python

A simplex method for function minimization

Scikit-learn: Machine Learning in Python

Model-agnostic meta-learning for fast adaptation of deep networks

Random search for hyper-parameter optimization

Related Papers (5)

Fast Efficient Hyperparameter Tuning for Policy Gradients

Efficient hyperparameter optimization through model-based reinforcement learning

Bayesian Optimization for Iterative Learning

Efficient Online Hyperparameter Optimization for Kernel Ridge Regression with Applications to Traffic Time Series Prediction

Combination of Hyperband and Bayesian Optimization for Hyperparameter Optimization in Deep Learning