7 research outputs found
Guided Zoom: Zooming into Network Evidence to Refine Fine-Grained Model Decisions
No full textIn state-of-the-art deep single-label classification models, the top-k (k = 2,3,4,....) accuracy is usually significantly higher than the top-1 accuracy. This is more evident in fine-grained datasets, where differences between classes are quite subtle. Exploiting the information provided in the top k predicted classes boosts the final prediction of a model. We propose Guided Zoom, a novel way in which explainabitity could be used to improve model performance. We do so by making sure the model has "the right reasons" fora prediction. The reason/evidence upon which a deep neural network makes a prediction is defined to be the grounding, in the pixel space, for a specific class conditional probability in the model output. Guided Zoom examines how reasonable the evidence used to make each of the top-k predictions is. Test time evidence is deemed reasonable if it is coherent with evidence used to make similar correct decisions at training time. This leads to better informed predictions. We explore a variety of grounding techniques and study their complementarity for computing evidence. We show that Guided Zoom results in an improvement of a model's classification accuracy and achieves state-of-the-art classification performance on four fine-grained classification datasets
Guided Zoom: Questioning Network Evidence for Fine-Grained Classification
No full textWe propose Guided Zoom, an approach that utilizes spatial grounding of a model’s decision to make more informed predictions. It does so by making sure the model has “the right reasons” for a prediction, defined as reasons that are coherent with those used to make similar correct decisions at training time. The reason/evidence upon which a deep convolutional neural network makes a prediction is defined to be the spatial grounding, in the pixel space, for a specific class conditional probability in the model output. Guided Zoom examines how reasonable such evidence is for each of the top-k predicted classes, rather than solely trusting the top-1 prediction. We show that Guided Zoom improves the classification accuracy of a deep convolutional neural network model and obtains state-of-the-art results on three fine-grained classification benchmark datasets
Guided Zoom: Zooming into Network Evidence to Refine Fine-Grained Model Decisions
No full textIn state-of-the-art deep single-label classification models, the top-kk (k=2,3,4, \dots)(k=2,3,4,⋯) accuracy is usually significantly higher than the top-1 accuracy. This is more evident in fine-grained datasets, where differences between classes are quite subtle. Exploiting the information provided in the top kk predicted classes boosts the final prediction of a model. We propose Guided Zoom, a novel way in which explainability could be used to improve model performance. We do so by making sure the model has 'the right reasons' for a prediction. The reason/evidence upon which a deep neural network makes a prediction is defined to be the grounding, in the pixel space, for a specific class conditional probability in the model output. Guided Zoom examines how reasonable the evidence used to make each of the top-kk predictions is. Test time evidence is deemed reasonable if it is coherent with evidence used to make similar correct decisions at training time. This leads to better informed predictions. We explore a variety of grounding techniques and study their complementarity for computing evidence. We show that Guided Zoom results in an improvement of a model's classification accuracy and achieves state-of-the-art classification performance on four fine-grained classification datasets. Our code is available at https://github.com/andreazuna89/Guided-Zoom
Untersuchung des Ladungstransportes in den Metal Halogenid Perowskiten durch THz Spektroskopie
No full textHalide perovskites are a class of novel photovoltaic materials that have recently attracted much attention in the photovoltaics research community due to their highly promising optoelectronic properties, including large absorption coefficients and long carrier lifetimes. The charge carrier mobility of halide perovskites is investigated in this thesis by THz spectroscopy, which is a contact-free technique that yields the intra-grain sum mobility of electrons and holes
in a thin film.
The polycrystalline halide perovskite thin films, provided from Potsdam University, show moderate mobilities in the range from 21.5 to 33.5 cm2V-1s-1. It is shown in this work that the room temperature mobility is limited by charge carrier scattering at polar optical phonons. The mobility at low temperature is likely to be limited by scattering at charged and neutral impurities at impurity concentration N=1017-1018 cm-3. Furthermore, it is shown that exciton formation
may decrease the mobility at low temperatures. Scattering at acoustic phonons can be neglected at both low and room temperatures. The analysis of mobility spectra over a broad range of temperatures for perovskites with various cation compounds shows that cations have a minor impact on charge carrier mobility.
The low-dimensional thin films of quasi-2D perovskite with different numbers of [PbI6]4−sheets (n=2-4) alternating with long organic spacer molecules were provided by S. Zhang from Potsdam University. They exhibit mobilities in the range from 3.7 to 8 cm2V-1s-1. A clear
decrease of mobility is observed with decrease in number of metal-halide sheets n, which likely arises from charge carrier confinement within metal-halide layers. Modelling the measured THz mobility with the modified Drude-Smith model yields localization length from 0.9 to 3.7 nm, which agrees well on the thicknesses of the metal-halide layers. Additionally, the mobilities are found to be dependent on the orientation of the layers. The charge carrier dynamics is also
dependent on the number of metal-halide sheets n. For the thin films with n =3-4 the dynamics is similar to the 3D MHPs. However, the thin film with n = 2 shows clearly different dynamics, where the signs of exciton formation are observed within 390 fs timeframe after
photoexcitation.
Also, the charge carrier dynamics of CsPbI3 perovskite nanocrystals was investigated, in particular the effect of post treatments on the charge carrier transport.Metall-Halogenid Perowskite sind eine Klasse von photovoltaischen Materialien, welche in letzter Zeit sehr viel Aufmerksamkeit von Forschern bekommen haben. Der Grund dafür liegt in ihren vielversprechenden optoelektronischen Eigenschaften, wie beispielsweise hohe Absorptionskoeffizienten, lange Lebenszeiten der Ladungsträger und moderate Beweglichkeiten.
Die Beweglichkeit der Ladungsträger und deren Kinetik wurde in dieser Dissertation mit Hilfe von Teraherzspektroskopie in verschiedenen Metall-Halogenide Perowskiten untersucht.
Die polykristallinen Halogenide Perowskit-Dünnschichten, bereitgestellt von Dr. M. Stolterfoht von der Universität Potsdam, haben bei Raumtemperatur moderate Ladungsträgerbeweglichkeiten in einem Bereich von 21.5 bis 33.5 cm²V-1s-1. Die Analyse dieser Beweglichkeiten in Abhängigkeit der Temperatur zeigt, dass die Beweglichkeit bei Raumtemperatur durch die Interaktion mit polaren optischen Phononen limitiert wird. Bei niedrigeren Temperaturen sind die Beweglichkeiten durch Streuung an geladenen und neutralen Störstellen limitiert, wobei die Störstellenkonzentration bei ca. N =1017-1018 cm-3 liegt. Weiterhin wird es gezeigt, dass die Reduktion der Anzahl beweglicher Ladungsträger durch Exzitonenbildung ebenfalls bei niedrigen Temperaturen berücksichtigt werden muss. Streuung an akustischen Phononen kann sowohl bei Raum- als auch bei niedrigen Temperaturen vernachlässigt werden. Die Analyse der Beweglichkeitsspektren von Perowskiten mit unterschiedlichen Kationen und bei verschiedenen Temperaturspannen zeigt, dass diese Kationen einen sehr geringen Einfluss auf die Ladungsträgerbeweglichkeit haben.
Niederdimensionale Perowskit-Dünnschichten aus alternierenden quasi-2D [PbI6]4− Schichten n (n=3-4) und organischen Trennschichten wurde von S. Zhang von der Universität Potsdam bereitgestellt. Diese zeigen Beweglichkeiten zwischen 3.7 und 8 cm²V-1s-1. Der signifikante Rückgang der beobachteten Beweglichkeit lässt sich auf die Anzahl der Metall-Halogeniden Schichten n zurückführen, in welcher die Ladungsträger räumlich eingeschränkt sind. Die Lokalisationslänge reicht von 0.9 bis 3.7 nm und ist vergleichbar mit der Dicke der einzelnen quasi-2D-Schichten. Ebenfalls ist die Beweglichkeit abhängig von der Schichtenorientierung. Zusätzlich ist die Ladungsträgerdynamik abhängig von der Anzahl der Metall-Halogeniden [PbI6]4−Schichten n. Dicke quasi-2D-Schichten (n = 3-4) zeigen ähnliche Dynamik wie drei dimensionale Perowskite, wogegen die dünnen quasi-2D-Schichten (n = 2) schnelle Exzitonbildung innerhalb 390 fs nach der Ladungsträgeranregung zeigen.
Des weiteren wurde die Ladungsträgerdynamik von CsPbI3 Perovskite-Nanokristallen untersucht, insbesondere die Auswirkung von Ligandenaustausch und Temperierung auf die Ladungsträgerbeweglichkeit
The role of bulk and interface recombination in high-efficiency low-dimensional Perovskite solar cells
No full text2D Ruddlesden–Popper perovskite (RPP) solar cells have excellent environmental stability. However, the power conversion efficiency (PCE) of RPP cells remains inferior to 3D perovskite-based cells. Herein, 2D (CH(CH)NH)(CHNH)PbI perovskite cells with different numbers of [PbI] sheets (n = 2–4) are analyzed. Photoluminescence quantum yield (PLQY) measurements show that nonradiative open-circuit voltage (V) losses outweigh radiative losses in materials with n > 2. The n = 3 and n = 4 films exhibit a higher PLQY than the standard 3D methylammonium lead iodide perovskite although this is accompanied by increased interfacial recombination at the top perovskite/C interface. This tradeoff results in a similar PLQY in all devices, including the n = 2 system where the perovskite bulk dominates the recombination properties of the cell. In most cases the quasi-Fermi level splitting matches the device V within 20 meV, which indicates minimal recombination losses at the metal contacts. The results show that poor charge transport rather than exciton dissociation is the primary reason for the reduction in fill factor of the RPP devices. Optimized n = 4 RPP solar cells had PCEs of 13% with significant potential for further improvements
