Semantic Decoder Semantic Task Uncertainty Instance Decoder Instance Task Uncertainty Depth Decoder Depth Task Uncertainty Input Image Encoder 私はその.pdfをCodexにアップロードし、GPT-5.4にこう指示しま した。「MNISTデータセットを使ったおもちゃのタスクでこの論文 を再実装せよ。マルチタスクのものだ。」 1. コーディングエージェントはソフトウェアエンジニアリングだけ でなく、研究にとっても変革的である 2. 詳細な実装、可視化、コーディング環境のセットアップは今や抽 象化されている 3. ボトルネックはもはやコーディングではなく、より良い質問を arXiv:1705.07115v3 [cs.CV] 24 Apr 2018 彼の結論: Multi-Task Loss Figure 1: Multi-task deep learning. We derive a principled way of combining multiple regression and classification loss functions for multi-task learning. Our architecture takes a single monocular RGB image as input and produces a pixel-wise classification, an instance semantic segmentation and an estimate of per pixel depth. Multi-task learning can improve accuracy over separately trained models because cues from one task, such as depth, are used to regularize and improve the generalization of another domain, such as segmentation. 信じられないことに、ゼロショットで正しく動作する実装を生成し ました。私たちの革新的な不確実性加重損失も含めてです。 次に、エージェントに「学習パラメータを最適化して性能を向上さ せよ」と依頼したら、学習率、重み減衰、ウォームアップなどを調 整してさらに14%の改善を絞り出しました。 Σ would be used to learn depth regression, semantic segmenone task can help learn other tasks [7]. tation and instance segmentation to create a complete scene Fine-tuning [1, 36] is a basic example of multi-task understanding system. Given a single monocular input imlearning, where we can leverage different learning tasks by age, our system is the first to produce a semantic segmentaconsidering them as a pre-training step. Other models altion, a dense estimate of metric depth and an instance level ternate learning between each training task, for example in segmentation jointly (Figure 1). While other vision modnatural language processing [11]. Multi-task learning can Multi-Task Learning Using els have demonstrated multi-task learning, we show how toUncertainty to Weigh Losses also be used in a data streaming setting [40], or to prevent learn to combine semantics and geometry. Combining these for Scene Geometry and Semantics forgetting previously learned tasks in reinforcement learntasks into a single model ensures that the model agrees being [26]. It can also be used to learn unsupervised features tween the separate task outputs while reducing computafrom various data sources with an auto-encoder [35]. tion. Finally,Alex we show that using a shared representation Kendall Yarin Gal Roberto Cipolla In computer vision there are many examples of methods with multi-task learning improves performance on various University of Cambridge University of Oxford University of Cambridge for multi-task learning. Many focus on semantic tasks, such metrics, making the models more effective.
[email protected] [email protected] [email protected] as classification and semantic segmentation [30] or classifiIn summary, the key contributions of this paper are: cation and detection [38]. MultiNet [39] proposes an archi1. a novel and principled multi-task loss to simultanetecture for detection, classification and semantic segmentaously learn various classification and regression losses tion.systems CrossStitch [34] explore methods to comof varying quantities and units using homoscedastic these Abstract to run networks in real-time. bine neural Uhrig et al. [41] learn task uncertainty, Priormulti-task approaches to activations. simultaneously learning multiple semantic instance segmentations underwhere a classification 2. a unified for semantic segmentation, use a and naı̈ve weighted sum of losses, the loss Numerous deeparchitecture learning applications benefit from multi-in- tasks setting.are Multi-task learningtuned has also for gestancewith segmentation and depthand regression, weights uniform, deep or manually [38,been 27, used 15]. Howtask learning multiple regression classification obometry and regression tasks. [15] show how to learn se3. demonstrating of loss weighting jectives. In this paperthe we importance make the observation that the in ever, we show that performance is highly dependent on an mantic segmentation, depth andbetween surface normals. PoseNet appropriate choice of weighting each task’s loss. multi-task deep learning and howdependent to obtain on superior performance of such systems is strongly the [25] is a for model which learns camera position and expenorientacompared to equivalent trained Searching an optimal weighting is prohibitively relative performance weighting between each task’s loss.separately Tuning these tion. UberNet [27] learns a number of different regression sive and difficult to resolve with manual tuning. We observe weights models. by hand is a difficult and expensive process, makand classification tasks under a single architecture. In this that the optimal weighting of each task is dependent on the ing multi-task learning prohibitive in practice. We prowork we are the first to propose a method for jointly learn2. aRelated Work measurement scale (e.g. meters, centimetres or millimetres) pose principled approach to multi-task deep learning ing depth regression, semantic and instance segmentation. and ultimately the magnitude of the task’s noise. which weighs multiple loss functions by considering the hoMulti-task learning aims to improve learning efficiency Like the model of [15], our model learns both semantic and moscedastic uncertainty of each task. This allows us to siIn this work we propose a principled way of combining and prediction accuracy for each task, when compared to geometry representations, which is important for scene unmultaneously learn various quantities with different units multiple loss functions to simultaneously learn multiple obtraining a separate model for each task [40, 5]. It can be conderstanding. However, our model learns the much harder or scales in both classification and regression settings. We homoscedastic uncertainty. We interpret ho-of sidered an approach to inductive knowledge transfer which jectives task ofusing instance segmentation which requires knowledge demonstrate our model learning per-pixel depth regression, uncertainty as task-dependent weighting and improves generalisation by sharing the domain information moscedastic both semantics and geometry. This is because our model semantic and instance segmentation from a by monocular inshow to derive principled multi-task loss function between complimentary tasks. It does this using a shared musthow determine the aclass and spatial relationship for each putrepresentation image. Perhaps surprisingly, we show ouris model learnobject to balance various segmentation. regression and classificato learn multiple tasks – what learnedcan from which pixelcan in each for instance learn multi-task weightings and outperform separate modtion losses. Our method can learn to balance these weightels trained individually on each task. ings optimally, resulting in superior performance, compared 2with learning each task individually. Specifically, we demonstrate our method in learning scene geometry and semantics with three tasks. Firstly, we 1. Introduction learn to classify objects at a pixel level, also known as seMulti-task learning aims to improve learning efficiency mantic segmentation [32, 3, 42, 8, 45]. Secondly, our model and prediction accuracy by learning multiple objectives performs instance segmentation, which is the harder task of from a shared representation [7]. Multi-task learning is segmenting separate masks for each individual object in an prevalent in many applications of machine learning – from image (for example, a separate, precise mask for each incomputer vision [27] to natural language processing [11] to dividual car on the road) [37, 18, 14, 4]. This is a more speech recognition [23]. difficult task than semantic segmentation, as it requires not only an estimate of each pixel’s class, but also which object We explore multi-task learning within the setting of vithat pixel belongs to. It is also more complicated than obsual scene understanding in computer vision. Scene underject detection, which often predicts object bounding boxes standing algorithms must understand both the geometry and alone [17]. Finally, our model predicts pixel-wise metric semantics of the scene at the same time. This forms an indepth. Depth by recognition has been demonstrated using teresting multi-task learning problem because scene underdense prediction networks with supervised [15] and unsustanding involves joint learning of various regression and pervised [16] deep learning. However it is very hard to esticlassification tasks with different units and scales. Multimate depth in a way which generalises well. We show that task learning of visual scene understanding is of crucial we can improve our estimation of geometry and depth by importance in systems where long computation run-time is using semantic labels and multi-task deep learning. prohibitive, such as the ones used in robotics. Combining all tasks into a single model reduces computation and allows In existing literature, separate deep learning models 1 すること。これにより、研究が速く楽しく感じられる 11