INTERPRETABLE VIDEO ANALYSIS WITH THE USE OF DEEP LEARNING TECHNIQUES.
The project aims at investigating whether classifiers trained on short videos can be more resistant to disturbance data (robust) from classifiers trained on single images.
The main driving force of the conducted research is the existence of the so-called adversarial attacks. Adversarial attacks rely on the fact that, imperceptible to the human eye, disturbances in input data result in misclassification, which is a serious security gap in systems based on artificial intelligence.
The more advanced variants of these attacks do not even require detailed knowledge of architecture of the classifier or access to training data, which can lead to the fact that emails with a competitor’s logo will be classified as spam by a significant proportion of e-mail boxes. This is a serious obstacle to the spread of intelligent systems. MIM Solutions is responsible for designing, implementing and testing appropriate algorithms in terms of their resistance to data disturbances, in particular to adversarial attacks.
The latest machine learning algorithms are making advances in solving problems that we previously thought were the exclusive domain of humans – tasks such as speech and object recognition.