Budapest University of Technology and Economics, Faculty of Electrical Engineering and Informatics

Péter Sárközy research assistant, Department of Measurement and Information Systems

A kötelező előtanulmányi rend az adott szak honlapján és képzési programjában található.

This course introduces artificial intelligence algorithms for information processing in embedded systems. The speciality is that information is basically data derived from physical processes, and the implementation of the algorithms will be specifically addressed in the context of the realization on embedded platforms.

Introduction

1. Introduction, description of the subject requirements. Overview of the areas of artificial intelligence, its application in embedded systems and the focus of the subject.

Information processing in embedded AI systems

2. Description of the data analysis workflow. Outlier detection and data cleaning, handling missing data, exploring the possibilities of knowledge modeling.

3. Analysis of the problems and solutions of regression and classification in a hardware environment, introduction to the related linear and logistic models.

4. Examination of the clustering problem, study of dimensionality reduction options.

5. Introduction to artificial intelligence sensor fusion methods for embedded applications.

6. Introduction to neural networks. Demonstration the effect of noise on the learning process. Examining the problems of overlearning, early stopping and backtracking on different platforms. Decomposition of the sample set into training, test and validation sets.

7. Description of the functionality of convolutional neural networks. Presentation of a pattern recognition system that can be run in an embedded environment.

8. Study of feedback neural networks. Introduction to the possibilities of prediction.

9. Interpreting the output of neurons. Demonstration of the importance of representation learning, description of autoencoder.

Embedded platforms for artificial intelligence applications

10. Overview of application limitations of general purpose devices (microcontroller, FPGA, general purpose processor).

11. Presentation of target hardware for implementing artificial intelligence on embedded platforms.

12. Presentation of smart devices, smart watches capabilities for embedded AI.

Detailed topics of the exercises

1. Application of linear and logistic regression and classification in an embedded environment, using examples with known physical models, testing the representational capabilities of linear models, adding new variables to the model.

2. Challenges of high dimensionality data, removing linear dependencies, applying principal component analysis and singular value decomposition to dimensionality reduction on an embedded platform, quantifying information loss, testing reversibility.

3. Sensor data integration, noise management, measurements from different sources, fusion of different measurement methods in hardware implementation.

4. Implementing applied neural networks in embedded systems, investigating the impact of noise on learning, calculating confidence of convergence, and discussing coupled over-learning, early shutdown and backtracking challenges. Decomposition of samples into training, test and validation sets.

5. Embedded application examples of convolutional neural networks, impact of kernel sizes on representability, explanatory analysis of learned feature vectors.

6. Time-series data analysis on embedded platforms, comparative analysis of autoregressive (ARIMA) methods and feedback neural network-based prediction architectures.

7. Unsupervised feature vector learning, the impact of latent dimensionality on the representativeness of models, sampling of generative models.

The theoretical part of the course will be given in the form of a frontal lecture, the practical part will be in the form of a computational exercise.

There will be a laboratory for the course in the next semester: Embedded artificial intelligence laboratory

During term time: Successful completion of 1 midterm exam (min. 40%)

During the examination period: Oral examination

One remedial course according to the TVSZ*, during the semester.

  (* CODE OF STUDIES AND EXAMS OF BME)

If required, a consultation can be arranged by prior appointment.

Stuart Russel, Peter Norvig: Artificial Intelligence - A Modern Approach, 4. Edition, 2021

Péter Sárközy   research assistant, Department of Measurement and Information Systems

Dr. Balázs Renczes   associate professor, Department of Measurement and Information Systems