<p><span style="font-size:14.0px"><p>Background/Motivation:<br>Face-based age estimation is central to many applications, such as crime prevention, identity verification, youth protection, and also in the medical field. Age estimation systems often show different performance on subgroups (e.g., regarding age, gender, ethnic affiliation). Reasons include, on the one hand, the availability or balance of training data and, on the other hand, classical training methods that optimise global metrics and ignore problems in certain subgroups.<br>Techniques such as oversampling or probabilistic sampling attempt to create a balance in the training data through statistical analysis in advance, with the hope that this will result in uniform performance across all subgroups. However, the result of the measure usually does not feed back into the training process; it remains unaffected.<br><br>Objective: The aim of this master's thesis is the development and systematic evaluation of training strategies that dynamically adjust the training process at runtime based on the current performance of the subgroups.<br><br>To this end: </p>
<ul>
<li>during training, subgroup-specific metrics are calculated and used for control.</li>
<li>Strategies are developed and implemented: dynamic oversampling of weak subgroups, uncertainty sampling, as well as curriculum strategies (first general/easy, then specific/hard).</li>
<li>appropriate aggregation metrics over the subgroups are examined (e.g., worst-group performance, harmonic mean, quantiles instead of macro-average).</li>
<li>Comparisons with baselines such as classic oversampling, probabilistic sampling, GroupDRO [1], and JTT [2] should be conducted.</li>
<li>AutoML/hyperparameter search will be used to explore combinations.</li>
</ul>
<p> </p>
<p>Results: The developed methods are intended to allow the training of balanced but also specialised computer vision models, particularly in the field of face-based age estimation. Suitable and successful measures are presented, guidelines on when which strategy (or combination) works, as well as limitations, pitfalls, and unexpected results. The methods are evaluated using freely available benchmark datasets and compared with existing methods. The code used is well-documented, reusable, and the results are reproducible.</p></span></p>
<p> </p>
<p><span style="color:#16a085"><span style="font-size:14.0px"><strong>Be part of change</strong></span></span></p>
<p><span style="font-size:14.0px"><ul>
<li>Researching and compiling information on a current topic in the field of machine learning.</li>
<li>Researching and implementing novel machine learning and computer vision approaches.</li>
<li>Self-critical evaluation of the obtained results.</li>
<li>Presenting the results.</li>
<li>Preparing a scientific paper in the form of a master's thesis with the results.</li>
</ul></span></p>
<p> </p>
<p><span style="color:#16a085"><span style="font-size:14.0px"><strong>What you contribute</strong></span></span></p>
<p><span style="font-size:14.0px"><ul>
<li>Good knowledge in the field of machine learning and training neural networks.</li>
<li>Ideally, knowledge in computer vision and facial recognition.</li>
<li>Good Python skills, preferably some experience with PyTorch, OpenCV, etc.</li>
<li>Motivation to independently delve into new and current research topics.</li>
<li>Interest in robustness and evaluation metrics.</li>
<li>Interest in scientific research. </li>
</ul></span></p>
<p> </p>
<p><span style="font-size:14.0px"><span style="color:#16a085"><strong>What we offer</strong></span></span></p>
<p><span style="font-size:14.0px"><ul>
<li>Independent work schedule management</li>
<li>Insights into the intersection of academic research and industrial application </li>
</ul>
<p> </p>
<p> Related works: <br>[1] Sagawa et al., Distributionally Robust Neural Networks for Group Shifts (GroupDRO) <a href="https://arxiv.org/pdf/1911.08731">https://arxiv.org/pdf/1911.08731</a> <br>[2] Liu et al., Just Train Twice: Improving Group Robustness Without Training Group Information (JTT) <a href="https://arxiv.org/pdf/2107.09044">https://arxiv.org/pdf/2107.09044</a> <br>[3] Hacohen, Weinshall (2019). On the Power of Curriculum Learning in Training Deep Networks <a href="http://proceedings.mlr.press/v97/hacohen19a/hacohen19a.pdf">http://proceedings.mlr.press/v97/hacohen19a/hacohen19a.pdf</a><br>[4] Roh et al., FairBatch: Batch Selection for Model Fairness — <a href="https://arxiv.org/pdf/2012.01696">https://arxiv.org/pdf/2012.01696</a> <br>[5] Ren et al., Learning to Reweight Examples for Robust Deep Learning — <a href="https://arxiv.org/pdf/1803.09050">https://arxiv.org/pdf/1803.09050</a> <br>[6] Cui et al., Class-Balanced Loss Based on Effective Number of Samples — <a href="https://arxiv.org/pdf/1901.05555">https://arxiv.org/pdf/1901.05555</a> <br>[7] Hashimoto et al., Fairness Without Demographics in Repeated Loss Minimization — <a href="https://arxiv.org/pdf/1806.08010">https://arxiv.org/pdf/1806.08010</a></p></span></p>
<p> </p>
<p><span style="font-size:14.0px">We value and promote the diversity of our employees' skills and therefore welcome all applications – regardless of age, gender, nationality, ethnic and social origin, religion, ideology, disability, sexual orientation and identity. Severely disabled persons are given preference in the event of equal suitability. Our tasks are diverse and adaptable – for applicants with disabilities, we work together to find solutions that best promote their abilities. </span></p>
<p><span style="font-size:14.0px">With its focus on developing key technologies that are vital for the future and enabling the commercial utilization of this work by business and industry, Fraunhofer plays a central role in the innovation process. As a pioneer and catalyst for groundbreaking developments and scientific excellence, Fraunhofer helps shape society now and in the future. </span></p>
<p><strong><span style="font-size:14.0px">Ready for a change? Then apply now and make a difference! Once we have received your online application, you will receive an automatic confirmation of receipt. We will then get back to you as soon as possible and let you know what happens next.</span></strong></p>
<p style="text-align:justify"> </p>
<p><span style="font-size:14.0px"><span style="font-size:14.0px"><span id="cke_bm_258S" style="display:none"> </span><span style="font-size:14.0px"><span style="font-size:14.0px"></span></span></span></span></p>
<p><span style="font-size:14.0px"><span style="font-size:14.0px"><span style="font-size:14.0px"><span style="font-size:14.0px">Fraunhofer Institute for Secure Information Technology SIT</span></span></span></span> </p>
<p><a href="http://www.sit.fraunhofer.de" target="_blank" rel="noopener"><span style="font-size:14.0px">www.sit.fraunhofer.de</span></a> </p>
<p> </p>
<p><span style="font-size:14.0px"><span style="font-size:14.0px"><span style="font-size:14.0px"><span style="font-size:14.0px">Requisition Number: <span style="font-size:14.0px"><span style="font-size:14.0px"><span style="font-size:14.0px"><span style="font-size:14.0px">82685</span></span></span></span> Application Deadline: <span style="font-size:14.0px"><span style="font-size:14.0px"><span style="font-size:14.0px"><span style="font-size:14.0px"></span></span></span></span></span></span></span></span></p>
<p style="text-align:right"> </p>
