Development of an Approach to Quantization of Neural Networks that Reduces Memory Requirements by Optimally Replacing the Activation Function with their Piecewise Constant Approximations

Student: Lapin Sergej

Educational Programme: Machine Learning and Data-Intensive Systems (Master)

Final Grade: 10

Year of Graduation: 2024

The work is devoted to the development of an approach to the optimization of neural networks by the method of quantization of their activations. A method not previously used in quantization is proposed, based on the selection of the optimal composition of activation functions to be binarized (replaced by piecewise constant approximations), based on the calculation of the value of "diversity" that they retain after replacement. The formalization of the proposed approach is given and experiments are carried out on the example of the ResNet50 network and the ImageNet1000 dataset.

Full text (added June 2, 2024)

Student Theses at HSE must be completed in accordance with the University Rules and regulations specified by each educational programme.

Summaries of all theses must be published and made freely available on the HSE website.

The full text of a thesis can be published in open access on the HSE website only if the authoring student (copyright holder) agrees, or, if the thesis was written by a team of students, if all the co-authors (copyright holders) agree. After a thesis is published on the HSE website, it obtains the status of an online publication.

Student theses are objects of copyright and their use is subject to limitations in accordance with the Russian Federation’s law on intellectual property.

In the event that a thesis is quoted or otherwise used, reference to the author’s name and the source of quotation is required.

Search all student theses