Refine
H-BRS Bibliography
- yes (12)
Departments, institutes and facilities
- Fachbereich Informatik (12) (remove)
Document Type
- Preprint (12) (remove)
Year of publication
- 2020 (12) (remove)
Language
- English (12)
Has Fulltext
- no (12)
Keywords
- Evolutionary Computation (2)
- Autoencoder (1)
- Automatic Short Answer Grading (1)
- BERT (1)
- Bayesian Deep Learning (1)
- Bayesian optimization (1)
- Black-Box Optimization (1)
- Computational Fluid Dynamics (1)
- Deep Learning (1)
- ELMo (1)
Object detectors have improved considerably in the last years by using advanced CNN architectures. However, many detector hyper-parameters are generally manually tuned, or they are used with values set by the detector authors. Automatic Hyper-parameter optimization has not been explored in improving CNN-based object detectors hyper-parameters. In this work, we propose the use of Black-box optimization methods to tune the prior/default box scales in Faster R-CNN and SSD, using Bayesian Optimization, SMAC, and CMA-ES. We show that by tuning the input image size and prior box anchor scale on Faster R-CNN mAP increases by 2% on PASCAL VOC 2007, and by 3% with SSD. On the COCO dataset with SSD there are mAP improvement in the medium and large objects, but mAP decreases by 1% in small objects. We also perform a regression analysis to find the significant hyper-parameters to tune.
In this paper we introduce the Perception for Autonomous Systems (PAZ) software library. PAZ is a hierarchical perception library that allow users to manipulate multiple levels of abstraction in accordance to their requirements or skill level. More specifically, PAZ is divided into three hierarchical levels which we refer to as pipelines, processors, and backends. These abstractions allows users to compose functions in a hierarchical modular scheme that can be applied for preprocessing, data-augmentation, prediction and postprocessing of inputs and outputs of machine learning (ML) models. PAZ uses these abstractions to build reusable training and prediction pipelines for multiple robot perception tasks such as: 2D keypoint estimation, 2D object detection, 3D keypoint discovery, 6D pose estimation, emotion classification, face recognition, instance segmentation, and attention mechanisms.
Reinforcement learning (RL) algorithms should learn as much as possible about the environment but not the properties of the physics engines that generate the environment. There are multiple algorithms that solve the task in a physics engine based environment but there is no work done so far to understand if the RL algorithms can generalize across physics engines. In this work, we compare the generalization performance of various deep reinforcement learning algorithms on a variety of control tasks. Our results show that MuJoCo is the best engine to transfer the learning to other engines. On the other hand, none of the algorithms generalize when trained on PyBullet. We also found out that various algorithms have a promising generalizability if the effect of random seeds can be minimized on their performance.
Comparative Evaluation of Pretrained Transfer Learning Models on Automatic Short Answer Grading
(2020)
Automatic Short Answer Grading (ASAG) is the process of grading the student answers by computational approaches given a question and the desired answer. Previous works implemented the methods of concept mapping, facet mapping, and some used the conventional word embeddings for extracting semantic features. They extracted multiple features manually to train on the corresponding datasets. We use pretrained embeddings of the transfer learning models, ELMo, BERT, GPT, and GPT-2 to assess their efficiency on this task. We train with a single feature, cosine similarity, extracted from the embeddings of these models. We compare the RMSE scores and correlation measurements of the four models with previous works on Mohler dataset. Our work demonstrates that ELMo outperformed the other three models. We also, briefly describe the four transfer learning models and conclude with the possible causes of poor results of transfer learning models.
Graph drawing with spring embedders employs a V x V computation phase over the graph's vertex set to compute repulsive forces. Here, the efficacy of forces diminishes with distance: a vertex can effectively only influence other vertices in a certain radius around its position. Therefore, the algorithm lends itself to an implementation using search data structures to reduce the runtime complexity. NVIDIA RT cores implement hierarchical tree traversal in hardware. We show how to map the problem of finding graph layouts with force-directed methods to a ray tracing problem that can subsequently be implemented with dedicated ray tracing hardware. With that, we observe speedups of 4x to 13x over a CUDA software implementation.
Facial emotion recognition is the task to classify human emotions in face images. It is a difficult task due to high aleatoric uncertainty and visual ambiguity. A large part of the literature aims to show progress by increasing accuracy on this task, but this ignores the inherent uncertainty and ambiguity in the task. In this paper we show that Bayesian Neural Networks, as approximated using MC-Dropout, MC-DropConnect, or an Ensemble, are able to model the aleatoric uncertainty in facial emotion recognition, and produce output probabilities that are closer to what a human expects. We also show that calibration metrics show strange behaviors for this task, due to the multiple classes that can be considered correct, which motivates future work. We believe our work will motivate other researchers to move away from Classical and into Bayesian Neural Networks.
Deep learning models are extensively used in various safety critical applications. Hence these models along with being accurate need to be highly reliable. One way of achieving this is by quantifying uncertainty. Bayesian methods for UQ have been extensively studied for Deep Learning models applied on images but have been less explored for 3D modalities such as point clouds often used for Robots and Autonomous Systems. In this work, we evaluate three uncertainty quantification methods namely Deep Ensembles, MC-Dropout and MC-DropConnect on the DarkNet21Seg 3D semantic segmentation model and comprehensively analyze the impact of various parameters such as number of models in ensembles or forward passes, and drop probability values, on task performance and uncertainty estimate quality. We find that Deep Ensembles outperforms other methods in both performance and uncertainty metrics. Deep ensembles outperform other methods by a margin of 2.4% in terms of mIOU, 1.3% in terms of accuracy, while providing reliable uncertainty for decision making.
In complex, expensive optimization domains we often narrowly focus on finding high performing solutions, instead of expanding our understanding of the domain itself. But what if we could quickly understand the complex behaviors that can emerge in said domains instead? We introduce surrogate-assisted phenotypic niching, a quality diversity algorithm which allows to discover a large, diverse set of behaviors by using computationally expensive phenotypic features. In this work we discover the types of air flow in a 2D fluid dynamics optimization problem. A fast GPU-based fluid dynamics solver is used in conjunction with surrogate models to accurately predict fluid characteristics from the shapes that produce the air flow. We show that these features can be modeled in a data-driven way while sampling to improve performance, rather than explicitly sampling to improve feature models. Our method can reduce the need to run an infeasibly large set of simulations while still being able to design a large diversity of air flows and the shapes that cause them. Discovering diversity of behaviors helps engineers to better understand expensive domains and their solutions.
Grasp verification is advantageous for autonomous manipulation robots as they provide the feedback required for higher level planning components about successful task completion. However, a major obstacle in doing grasp verification is sensor selection. In this paper, we propose a vision based grasp verification system using machine vision cameras, with the verification problem formulated as an image classification task. Machine vision cameras consist of a camera and a processing unit capable of on-board deep learning inference. The inference in these low-power hardware are done near the data source, reducing the robot's dependence on a centralized server, leading to reduced latency, and improved reliability. Machine vision cameras provide the deep learning inference capabilities using different neural accelerators. Although, it is not clear from the documentation of these cameras what is the effect of these neural accelerators on performance metrics such as latency and throughput. To systematically benchmark these machine vision cameras, we propose a parameterized model generator that generates end to end models of Convolutional Neural Networks(CNN). Using these generated models we benchmark latency and throughput of two machine vision cameras, JeVois A33 and Sipeed Maix Bit. Our experiments demonstrate that the selected machine vision camera and the deep learning models can robustly verify grasp with 97% per frame accuracy.
In optimization methods that return diverse solution sets, three interpretations of diversity can be distinguished: multi-objective optimization which searches diversity in objective space, multimodal optimization which tries spreading out the solutions in genetic space, and quality diversity which performs diversity maintenance in phenotypic space. We introduce niching methods that provide more flexibility to the analysis of diversity and a simple domain to compare and provide insights about the paradigms. We show that multiobjective optimization does not always produce much diversity, quality diversity is not sensitive to genetic neutrality and creates the most diverse set of solutions, and multimodal optimization produces higher fitness solutions. An autoencoder is used to discover phenotypic features automatically, producing an even more diverse solution set. Finally, we make recommendations about when to use which approach.
The way solutions are represented, or encoded, is usually the result of domain knowledge and experience. In this work, we combine MAP-Elites with Variational Autoencoders to learn a Data-Driven Encoding (DDE) that captures the essence of the highest-performing solutions while still able to encode a wide array of solutions. Our approach learns this data-driven encoding during optimization by balancing between exploiting the DDE to generalize the knowledge contained in the current archive of elites and exploring new representations that are not yet captured by the DDE. Learning representation during optimization allows the algorithm to solve high-dimensional problems, and provides a low-dimensional representation which can be then be re-used. We evaluate the DDE approach by evolving solutions for inverse kinematics of a planar arm (200 joint angles) and for gaits of a 6-legged robot in action space (a sequence of 60 positions for each of the 12 joints). We show that the DDE approach not only accelerates and improves optimization, but produces a powerful encoding that captures a bias for high performance while expressing a variety of solutions.
Modern Monte-Carlo-based rendering systems still suffer from the computational complexity involved in the generation of noise-free images, making it challenging to synthesize interactive previews. We present a framework suited for rendering such previews of static scenes using a caching technique that builds upon a linkless octree. Our approach allows for memory-efficient storage and constant-time lookup to cache diffuse illumination at multiple hitpoints along the traced paths. Non-diffuse surfaces are dealt with in a hybrid way in order to reconstruct view-dependent illumination while maintaining interactive frame rates. By evaluating the visual fidelity against ground truth sequences and by benchmarking, we show that our approach compares well to low-noise path traced results, but with a greatly reduced computational complexity allowing for interactive frame rates. This way, our caching technique provides a useful tool for global illumination previews and multi-view rendering.