TASK Quarterly

Object detection and multimodal learning for product recommendations

2024-11-13T12:08:42+01:00

This study showcases how deep learning can be applied to automated information extraction in fashion data to create a recommendation system. The proposed approach is an algorithm for recommending multiple products based on visual and textual features, ensuring compatibility with query items. The object detection model can detect many products across different garment categories. The study utilized public e-commerce datasets and trained models using deep learning methods. The compatibility model has shown promising results in automating recommendations of compatible products based on user interests. The study experimented with multiple pre-trained feature extraction models and successfully trained the object detection model for fashion article detection and localization task. Overall, the goal is to deploy the method to enhance its effectiveness and usefulness in providing a satisfying shopping experience for e-commerce users.

Edge coloring of small signed graphs

2024-02-14T10:01:40+01:00

In 2020, Behr introduced the problem of edge coloring of signed graphs and proved that every signed graph (G, sigma) can be colored using Delta(G) or Delta(G) + 1 colors, where Delta(G) denotes the maximum degree of G. Three years later, Janczewski et al. introduced a notion of signed class 1, such that a graph G is of signed class 1 if and only if every signed graph (G, sigma) can be colored using Delta(G) colors.

It is a well-known fact that almost all graphs are of class 1. In this paper we conjecture that the similar fact is true for signed class 1, that almost all graphs are of signed class 1. To support the hypothesis we implemented an application that colored all the signed graphs with at most 8 vertices. We describe an algorithm behind the application and discuss the results of conducted experiments.

Breath Detection from a Microphone Using Machine Learning

2024-10-15T13:06:07+02:00

This project investigates and implements various artificial intelligence techniques for the real-time detection of breath sounds using audio data captured via a computer microphone. The primary objective is to develop and compare methodologies to identify distinct phases of breathing, namely inhalation, exhalation, and the silent intervals between breaths, in order to determine the most accurate, efficient, and practical approach.
The study explores three innovative approaches:
1. VGGish Model for Feature Extraction and Classification with Random Forest: This method utilizes the VGGish model to extract sound feature vectors, followed by classification using a random forest classifier.
2. Spectrogram Classification Using Convolutional Neural Networks: This approach involves classifying spectrograms of half-second or quarter-second audio segments using a convolutional neural network adapted for image classification tasks.
3. Mel-Frequency Cepstral Coefficients (MFCC) for Feature Extraction and Neural Network Classification: This method employs MFCCs as set of sound features for classification using a neural network.
The experimental results show that methods 1 and 3 achieved an accuracy of 87% in the test data, while method 2 achieved an accuracy of 83%. The dataset comprised approximately 1,000 recordings of inhalations, exhalations, and
silences between breaths, collected using four different microphones and recorded by three different individuals.
All implementations and training data are available on a public GitHub repository: github.com/tomaszsankowski/Breathing-
Classification.