The goal of supervised learning is to learn a function that maps an input to an output based on input-output pairs.
At training time deep learning algorithms generally require a large number of labeled
training instances which are fairly rare in many domains. In practice, sets of labeled data are often
curated manually which is not only an unattractive job but also time-consuming and expensive.
In this project, which is an industry project commissioned by Harman International Inc., the students developed a labeling tool for object detection in image data which is additionally supported by active learning to reduce the amount of manual labeling effort. While training a deep object detection network in the background, the tool automatically selects unlabeled images that are, with respect to some evaluation metric, expected to improve the object detection network most. These images are shown to human labelers and subsequently used for training the network.
Have you ever seen a movie trailer and asked yourself whether you should spend the money watching this movie at the cinema? What are the decisive factors: the cast, the budget, the genre, the plot?
In this project, a group of students tackled this question by developing an AI approach to predict the average movies’ IMDb ratings. Via the web interface, the user can provide any information about a movie, e.g., the plot as text, the movie’s poster as an image file, and/or simply meta information such as actors, duration, genre, and so on. This information is collected, preprocessed, and given as input to the multi-component neural network in the backend which performs the actual regression for predicting the IMDb movie rating.
The KDD Cup is an annual competition in Data Mining and Knowledge Discovery organized by SIGKDD alongside the KDD conference.
In 2018, the participants were requested to predict the concentration of several air pollutants for London and Beijing. Given the historical measurement data from several air quality stations and weather data from meteorology stations, the task was to combine the weather forecast for the next 48 hours with this data to obtain a forecast of the air pollutants’ concentrations. Inspired by this competition, the group built a system that automatically retrieves the data, stores them in a database, and trains a variety of machine learning models on the collected data. Also, the process of evaluating the different models and their configurations is automated.
An appealing web interface allows exploring the data as well as the predictions of the models as well as their quality.
Despite the overwhelming success of deep convolutional networks in a broad variety of applications, most prominently image classification, their inner workings remain not fully understood. These networks comprise many layers, and each layer consists of a number of so-called channels. When fixing the layer and the channel, the values computed across different image locations are called activations and measure the presence of the learned feature. It is known that deeper layers learn more abstract and invariant features than early layers. In order to facilitate a better understanding of these features, the group built a system for explorative analysis of the activations of the famous Inception V3 model. Accessed via a web interface, the distribution of activation values faceted by class labels can be investigated.
Moreover, the tool shows a ranking of image patches that yielded the highest activation values, i.e. the patches that a feature responds to the most.