Łukasz Neumann’s research while affiliated with Warsaw University of Technology and other places

In this work we present an automated approach to allergy recognition based on neural networks. Allergic reaction classification is an important task in modern medicine. Currently it is done by humans, which has obvious drawbacks, such as subjectivity in the process. We propose an automated method to classify prick allergic reactions using correlated visible-spectrum and thermal images of a patient’s forearm. We test our model on a real-life dataset of 100 patients (1584 separate allergen injections). Our solution yields good results—0.98 ROC AUC; 0.97 AP; 93.6% accuracy. Additionally, we present a method to segment separate allergen injection areas from the image of the patient’s forearm (multiple injections per forearm). The proposed approach can possibly reduce the time of an examination, while taking into consideration more information than possible by human staff.

In this paper, we discuss an evolutionary method for training deep neural networks. The proposed solution is based on the Differential Evolution Strategy (DES) – an algorithm that is a crossover between Differential Evolution (DE) and the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). We combine this approach with Xavier’s coefficient-based population initialization, batch processing, and gradient-based mutations—the resulting weight optimizer is called neural Differential Evolution Strategy (nDES). Our algorithm yields results comparable to Adaptive Moment Estimation ADAM for a convolutional network training task (50K parameters) on the FashionMNIST dataset. We show that combining both methods results in better models than those obtained after training by either of these algorithms alone. Furthermore, nDES significantly outperforms ADAM on three classic toy recurrent neural network problems. The proposed solution is scalable in an embarrassingly parallel way. For reproducibility purposes, we provide a reference implementation written in Python.

Recurrent neural networks are key tools for sequential data processing. Existing architectures support only a limited class of operations that these networks can apply to their memory state. In this paper, we address this issue and introduce a recurrent neural module called Deep Memory Update (DMU). This module is an alternative to well-established LSTM and GRU. However, it uses a universal function approximator to process its lagged memory state. In addition, the module normalizes the lagged memory to avoid gradient exploding or vanishing in backpropagation through time. The subnetwork that transforms the memory state of DMU can be arbitrary. Experimental results presented here confirm that the previously mentioned properties of the network allow it to compete with and often outperform state-of-the-art architectures such as LSTM and GRU.

Predicting customer decisions allows companies to obtain higher profits due to better resource management. The accuracy of those predictions can be currently boosted by the application of machine learning algorithms. We propose a new method to predict a car driver’s decision about taking a replacement car after a vehicle accident happens. We use feature engineering to create attributes of high significance. The generated attributes are related to time (e.g., school holidays), place of collision (e.g., distance from home), time and conditions (e.g., weather), vehicles (e.g., vehicle value), addresses of both the victim and the perpetrator. Feature engineering involves external sources of data. Five machine learning methods of classification are considered: decision trees, multi-layer perceptrons, AdaBoost, logistic regression and gradient boosting. Algorithms are tested on real data from a Polish insurance company. Over 80% accuracy of prediction is achieved. Significance of the attributes is calculated using the linear vector quantization method. Presented work shows the applicability of machine learning in the car insurance market.

The next generation sequencing techniques produce a large amount of sequencing data. Some part of the genome are composed of repetitive DNA sequences, which are very problematic for the existing genome assemblers. We propose a modification of the algorithm for a DNA assembly, which uses the relative frequency of reads to properly reconstruct repetitive sequences. The new approach was implemented and tested, as a demonstration of the capability of our software we present some results for model organisms. The new implementation, using a three-layer software architecture was selected, where the presentation layer, data processing layer, and data storage layer were kept separate. Source code as well as demo application with web interface and the additional data are available at project web-page: http://dnaasm.sourceforge.net.