Review - Agenda 2020

The perception of what defines AI has become very blurry. The more attention is drawn to the topic and the more activities we see evolving, the bigger the need to reflect on the actual success factors of AI solution development. It’s overdue to take a look under the hood and establish a common understanding of what building blocks constitute an AI solution, what’s easily doable today, where things get experimental and where we drift into science fiction. The keynote will demystify the current AI hype and clarify its misperception. A strong focus will be on the question why AI initiatives get stuck in the experimental or PoC phase and what countermeasures can be taken to really scale AI solutions.

The talk will consist of a detailed discussion of the economic impact, regulatory constraints and technology challenges for inspecting commercial aircraft with the help of drones and analyzing the data with deep learning strategies. or solving the anomaly detection problem, we applied unsupervised deep reconstruction algorithms, in which Artificial Neural Network uses a corrupted original image as an input and performs reconstruction of the corrupted part.

Image classification has been solved successfully for many tasks e.g. using deep learning techniques. However, in many application scenarios, the set of classes from which input images are drawn is not completely known at time of modeling. This is typically the case when leaving controlled environments like in production systems and entering e.g. domestic settings. It is then of outmost importance to a) reduce false positives during model inference and b) enable the description of “unknown” image content during annotation for efficient iteration of the modeling pipeline. This case study shows how we addressed these issues at Miele for recognition of food items. Based on a comparison of state-of-the-art techniques, the presentation shows how to best reduce false positive rates. Additionally, a concept and tooling for image annotation is presented that allows to deal with “unknowns” early in the modeling process.

Deep Learning image classifiers represent a breakthrough in image recognition and classification tasks, however their practical use in industry remains limited: A “clean” product image can be easily classified, but what about a context-dependent stock image of “a man jogging on the beach”? Is it a pair of shorts? His shoes? His watch which hardly shows up in the picture? While results can be improved by retraining and extending existing architectures, this exercise is very resource intensive (labeling, training, tuning, etc.). Shopify was tasked with mapping a catalogue of products that had highly problematic images and inconsistent metadata. In this talk Yizhar will describe their practical approach to the solution: Why they ended up using only some outputs of pre-trained image classifiers, how they combine these outputs with domain knowledge and graph theory to achieve our goals, and how this process allows them to avoid some of the typical pitfalls and expensive re-training.

To remain cost-competitive the wind industry must find new efficiencies such as transitioning from reactive maintenance to a more proactive strategy that exploits advanced analytics to identify damage. In this talk, Jonathan and Ulrich will discuss their work with Sulzer&Schmid Laboratories, a Swiss company that provides wind turbine inspection services, to automate the image pre-annotation stage of their inspection pipeline using deep learning-based predictive analytics. They will discuss the challenges related to training large supervised models with incomplete and noisy labels, and demonstrate the solution in action, showing how it not only reduces cost and increases throughput, but also actually improves upon the performance of the human annotators in terms of both accuracy and utility.

1,628 kidnapped children were retrieved from a single railway station in northern India in 2018. These children were in the age group of 4 to 15 years and included 134 girls amongst which the youngest was only four years old. This has become an epidemic in India where gangs kidnap children from weaker sections of society, later to be sold either to brothels for prostitution or factories for child labor. The local law enforcement agencies only raid these places if they have confirmed identity for any of the children. Skylark labs ChildrEN SafEty Retrieval (CENSER) System is being used by non-profits in India to identify and then convince law enforcement to raid these places to retrieve the kidnapped children. The volunteers of the non-profits go to sell makeup to the brothels while wearing hidden cameras. The children buying makeup are recorded on the hidden cameras which are used by the CENSER system to establish a match with the database of missing children. Since the facial features of children have changed significantly since being kidnapped, the proposed system can perform age-invariant face recognition. A majority of the time, the image of the child is not available at the age when he/she was kidnapped. In that case, the CENSER system uses kinship analysis as well as matches with the sketch of the child. The talk will detail this very severe problem and how the CENSER system is making a change to save these kidnapped children.

This presentation addresses the challenges of applying Deep Learning use cases for business innovation. Over the past years, UnternehmerTUM has worked with more than 80 cross-functional teams at Digital Product School and deployed the best practices of lean startup, data-driven design thinking (data thinking) and agile framework for our corporate partners. This presentation will summarize the key challenges when deep learning and data intelligence contribute to the customer experience and value proposition. The key learnings on successful data strategies to overcome those challenges will be overviewed and adoption of various techniques will be contrasted.

Do you know that it is possible to create glasses to bypass facial recognition system? The use of facial recognition technology is on the rise, and you can find it in different areas of human activity including social media, smart homes, ATMs, and stores. Recently, researchers have discovered that deep learning algorithms are vulnerable to various attacks called adversarial examples. Our client, a smart home solution provider, asked us to test software and hardware to reveal a real threat or academic research and select the most secure solution available on the market. Here is the way we conducted the security assessment. Facial recognition systems have their specific deep learning models. The systems usually work in the physical environment, and their attack surface differs from the digital one presented in research papers. Furthermore, all examples of attacks and defensive measures were given for various models, datasets, and conditions. It does not help to understand the real situation even if you examine approximately 100 research papers on this subject. To test properly, we’ve composed our own attack taxonomy to check the effectiveness of the recent approaches to attacking facial recognition systems. I will present our research conducted in the real environment with various cameras and algorithms and show how to protect production systems from this kind of attacks.

It is common to think that our AI technology works in ways similar to how our brain works. But this is not really true. This presentation will explain some stunning differences in how our brain works and how deep learning works. It will explain the theoretical fundamentals as well as show you the implications of those limitations for business applications of neural networks.

The goal of this project was to detect illegal mines from satellite data in Suriname. Illegal mines are known to heavily use chemicals, which eventually contaminate nearby water resources. This talk sheds light on which spectral data we chose, which computer vision / deep learning algorithms we used and how accurate our predictions are. One of our most prominent challenge was the imbalanced data set. Philipp will comment on their experiences with pretraining, network architectures, feature engineering and selecting a loss function.

The use of satellite imagery becomes more popular and important in analysis for various fields of the economy and business especially in agriculture. This talk will present intelligent system that automates process of counting grazing cattle and sheep for the Department of Agriculture, Environment and Rural Affairs (DAERA) of the United Kingdom. DAERA’s data of animal locations and movements are currently based on herd keepers’ addresses, complemented by on-site surveys. In order to reduce the need for physical inspections and manual livestock counting to validate subsidy claims and improve the assessment and management of disease outbreaks, enabling authorities to target their actions we’ve proposed solution based on aerial/satellite imagery and Convolutional Neural Networks. This talk will highlight orthophotomap and satellite imagery data preparation process, geospatial data processing, object detection and semantic segmentation modelling for counting cattle and sheep.

Voice assistants are becoming more and more popular. Today, most available voice assistants still talk with non-emotional tone. However, recent press announcements and patents of big IT companies indicate that this is about to change. From a marketing perspective, this raises the question: How will voice assistants talking with emotional tone impact consumer’s behavior? For example, could a voice assistant with an exciting tone make consumers impulsively buy products they hadn’t planned to buy before? In this talk, Carolin and Rene present a new deep learning model which synthesizes speech with emotional tone requiring only a small amount of voice recordings and they show how the emotional tone of a voice assistant affects consumer’s behavior in a shopping experiment.

Today’s enterprises are gaining access to ever-increasing volumes of data. However, a large fraction of potentially useful information remains effectively unused as it is locked in unstructured formats that are difficult to analyze or search. Jörn and Till have tackled the challenge of extracting information from unstructured documents, particularly tabular data in financial reports, using deep visual perception algorithms. They present their approach to establishing neural network architectures, outline the full-lifecycle of operating the models in production, and discuss challenges regarding how non-technical users can leverage these tools. Jörn and Till demonstrate how learning-based solutions can help to continuously generate insights from previously inaccessible sources of information and thereby reduce risks in substantiating decisions.

Larger companies face the problem of large numbers of incoming documents, binding significant resources in reading and redirecting to the corresponding sections. In this presentation an approach is presented, which discusses the solution regarding incoming emails at ERGO Nuremberg. Approx. 1k emails receives this location per day, which were manually read and redirected. Expecting an estimated increase of the email rate, a solution based on current machine learning approaches has to be implemented to support the company. This talk will cover the data preparation, the architecture and the final operationalization of this model in detail. The additional challenges of data science in the field of insurances regarding data protection law will be discussed as well as the technical requirements and boundaries as well as the usual pitfalls.