AEP

We introduce the Self-Learning Autonomous Edge Learning and Inferencing Pipeline (AEP). AEP is a software system we have developed for resource-constrained devices. Beside the inference capability, the AEP allows autonomous field training, by means of a two stage pipeline, involving a label generation with a confidence measure step and on-device training.

For label generation, AEP implements the following clustering algorithm:

• k-means:

For on-device training, AEP currently implements the following ML algorithms (training and classification) for binary classification problems:

• Decision Tree
• K-NN

Usage

You can compile the code as an executable or as a static library, using gcc/g++ for a Microcontroller or a desktop (e.g., thorugh Eclipse CDT or Visual Studio Code).

The program must be configured first in main.h, where the user has to specify some #define, such as:

• Algorithm: to enable one of the two supported classifirers

  • AutoDT: the DT training algorithm is executed,
  • AutoKNN: the samples together with their corresponding pseudo-labels (resulting from k-means clustering) are ready to be used by the K-NN classifier,

• Confidence: to enable/disable the confidence algorithm

  • CONFIDENCE: the k-means labels are evaluated by a confidence algorithm which filters the samples to be used for training one of the two supported classifiers (can be enabled/disabled),

• Filter: to enable one of the three supported memory filtering strategies

  • FIFO: removes the older samples from memory,
  • RND: removes random samples from memory,
  • CONF: removes samples having low confidence values,

Second, the memory should be configured in pipeline.h, where the user has to choose values for:

• INITIAL_THR: initial number of samples to start the pipeline,
• UPDATE_THR: after this number of samples, the k-means clustering is performed again for all samples in the memory, providing the new set of pseudo-labels and updating the classifiers accordingly,
• MEMORY_SIZE: maximum number of samples that fit in memory,

Third, the k-means algorithm should be configured in kmeans.h, where the user has to choose values for:

• K: number of clusters,
• ITERATION: maximum number of iterations of the k-means algorithm for a single run.
• CONFIDENCE_THR: the threshold for the confidence algorithm, so to remove all samples having weights lower than the threshold

AutoDT

If AutoDT was set, the program must be configured in decision_tree_training.h, where the user has to choose values for:

• MAX_DEPTH: maximum depth of the tree,
• MIN_SIZE: minimum number of samples required to split an internal node,

AutoKNN

If AutoKNN was set, the program must be configured in knn_classification.h, where the user has to choose value for:

• K_NEIGHBOR: number of neighbors k to be considered in each decision,

main.h exposes the following functions:

• kmeans, k-means clustering algorithm. More details about this function can be found in the kmeans.c file
• decision_tree_training, the decision tree training algorithm. More details about this function can be found in the decision_tree_training.c file
• decision_tree_classifier, the decision tree classification algorithm. More details about this function can be found in the decision_tree_classifier.c file
• knn_classification, the K-NN classification algorithm. More details about this function can be found in the knn_classification.c file kmeans_classifier
• pipeline, the pipeline function followed after clustering and training. More details about this function can be found in the pipeline.c file
• quicksort_idx and update_mem, these two functions are responsible of sorting the samples in memory to keep the most confident ones when CONF filtering is enables.
• random_func, this function is responsible of choosing the random samples to be removed when RND filtering is enables.

Data

• float 32 data are used
• The dataset is saved as header and source files (dataset and test) for the training and testing phases respectively to evaluate the AEP and mimic a "reading sensory data" scenario (the data in this repository represent 80% trainng data and 20% testing data from the Pima Indian Dataset)

Results

The obtained results after running the AEP with the user defined parameters are saved in a text file named log.txt

Reference article for more infomation

F., Sakr, R. Berta, J. Doyle, A. De Gloria, and F., Bellotti, "Self-Learning Pipeline for Low-Energy Resource-Constrained Devices," Energies 2021, 14, 6636. https://www.mdpi.com/1996-1073/14/20/6636