242 research outputs found
Smooth loss functions for deep top-k classification
Get PDFThe top- error is a common measure of performance in machine learning and computer vision. In practice, top- classification is typically performed with deep neural networks trained with the cross-entropy loss. Theoretical results indeed suggest that cross-entropy is an optimal learning objective for such a task in the limit of infinite data. In the context of limited and noisy data however, the use of a loss function that is specifically designed for top- classification can bring significant improvements.
Our empirical evidence suggests that the loss function must be smooth and have non-sparse gradients in order to work well with deep neural networks. Consequently, we introduce a family of smoothed loss functions that are suited to top- optimization via deep learning. The widely used cross-entropy is a special case of our family. Evaluating our smooth loss functions is computationally challenging: a na{\"i}ve algorithm would require operations, where is the number of classes. Thanks to a connection to polynomial algebra and a divide-and-conquer approach, we provide an algorithm with a time complexity of . Furthermore, we present a novel approximation to obtain fast and stable algorithms on GPUs with single floating point precision. We compare the performance of the cross-entropy loss and our margin-based losses in various regimes of noise and data size, for the predominant use case of . Our investigation reveals that our loss is more robust to noise and overfitting than cross-entropy
Smooth loss functions for deep top-k classification
Get PDFThe top- error is a common measure of performance in machine learning and computer vision. In practice, top- classification is typically performed with deep neural networks trained with the cross-entropy loss. Theoretical results indeed suggest that cross-entropy is an optimal learning objective for such a task in the limit of infinite data. In the context of limited and noisy data however, the use of a loss function that is specifically designed for top- classification can bring significant improvements. Our empirical evidence suggests that the loss function must be smooth and have non-sparse gradients in order to work well with deep neural networks. Consequently, we introduce a family of smoothed loss functions that are suited to top- optimization via deep learning. The widely used cross-entropy is a special case of our family. Evaluating our smooth loss functions is computationally challenging: a naïve algorithm would require operations, where is the number of classes. Thanks to a connection to polynomial algebra and a divide-and-conquer approach, we provide an algorithm with a time complexity of . Furthermore, we present a novel approximation to obtain fast and stable algorithms on GPUs with single floating point precision. We compare the performance of the cross-entropy loss and our margin-based losses in various regimes of noise and data size, for the predominant use case of . Our investigation reveals that our loss is more robust to noise and overfitting than cross-entropy
Low cost manufacturing of light trapping features on multi-crystalline silicon solar cells : jet etching method and cost analysis
Get PDFThesis (S.M.)--Massachusetts Institute of Technology, Dept. of Mechanical Engineering; and, (S.M. in Technology and Policy)--Massachusetts Institute of Technology, Engineering Systems Division, Technology and Policy Program, 2010.This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.Cataloged from student-submitted PDF version of thesis.Includes bibliographical references (p. 127-128).An experimental study was conducted in order to determine low cost methods to improve the light trapping ability of multi-crystalline solar cells. We focused our work on improving current wet etching methods to achieve the desired light trapping features which consists in micro-scale trenches with parabolic cross-sectional profiles with a target aspect ratio of 1.0. The jet etching with a hard mask method, which consists in impinging a liquid mixture of hydrofluoric, nitric and acetic acids through the opening of hard mask, was developed. First, a computational fluid dynamics simulation was conducted to determine the desired jet velocity and angle to be used in our experiments. We find that using a jet velocity of 3 m/s and a jetting angle of 45° yields the necessary flow characteristics for etching high aspect ratio features. Second, we performed experiments to determine the effect of jet etching using a photo-resist mask and thermally grown silicon oxide mask on multiple silicon substrates : , , and multi-crystalline silicon. Compared to a baseline of etching with no jet, we find that the jet etching process can improve the light trapping ability of the baseline features by improving their aspect ratio up to 65.2% and their light trapping ability up to 38.1%. The highest aspect ratio achieved using the jet etching process was 0.62. However, it must be noted that the repeatability of the results was not consistent: significant variations in the results of the same experiment occurred, making the jet etching process promising but difficult to control. Finally, we performed a cost analysis in order to determine the minimum efficiency that a jet etching process would have to achieve to be cost competitive and its corresponding features aspect ratio. We find that a minimum cell efficiency of 16.63% and feature aspect ratios of 0.57 are necessary for cost competitiveness with current solar cell manufacturing technology.by Amine Berrada Sounni.S.M.in Technology and PolicyS.M
Migrant Necropolitics at the Table: Civilized Cannibalism in Mahi Binebine\u27s \u3cem\u3eCannibales\u3c/em\u3e
Get PDFIn Cannibales, the Maghrebi Francophone author Mahi Binebine revisits the encounter between the so-called “cannibals” and the European colonizer in the context of illegal immigration where bodies become commodities exchangeable for social improvements creating a different form of cannibalism. It is no longer the usual dichotomy between the civilized and the savage that is at work but rather a “civilized” European imperialist who feeds himself on a migrant’s flesh. This article argues that this representation works as a “colonial fragment” from the past but contextualized in today’s globalization. Binebine’s morbid depiction of an ambivalent postcolonial cannibalistic encounter translates as a representation of migrants in terms of cannibalistic necropolitics. The illegal migrant has no choice but to be swallowed by a narcissistic exocannibalism which seeks to incorporate what it feeds on to a total unity suggesting a bleak future not only for illegal migrants but for globalization as possibly devouring itself
NK and T cells constitute two major, functionally distinct intestinal epithelial lymphocyte subsets in the chicken
Get PDFNon-mammalian NK cells have not been characterized in detail; however, their analysis is essential for the understanding of the NK cell receptor phylogeny. As a first step towards defining chicken NK cells, several tissues were screened for the presence of NK cells, phenotypically defined as CD8(+) cells lacking T- or B-lineage specific markers. By this criteria, approximately 30% of CD8(+) intestinal intraepithelial lymphocytes (IEL), but <1% of splenocytes or peripheral blood lymphocytes were defined as NK cells. These CD8(+)CD3(-) IEL were used for the generation of the 28-4 mAb, immunoprecipitating a 35-kDa glycoprotein with a 28-kDa protein core. The CD3 and 28-4 mAb were used to separate IEL into CD3(+) IEL T cells and 28-4(+) cells, both co-expressing the CD8 antigen. During ontogeny, 28-4(+) cells were abundant in the IEL and in the embryonic spleen, where two subsets could be distinguished according to their CD8 and c-kit expression. Most importantly, 28-4(+) IEL lysed NK-sensitive targets, whereas intestinal T cells did not have any spontaneous cytolytic activity. These results define two major, phenotypically and functionally distinct IEL subpopulations, and imply an important role of NK cells in the mucosal immune system
Ligophorus hamzati Hafidi & Diamanka & Rkhami & Pariselle 2013, n. sp.
No full textLigophorus hamzati n. sp. (Fig. 5) TYPE MATERIAL. — Holotype MNHN HEL315; paratypes MNHN (9) HEL316, BMNH (10) 2012.12.17.4. MATERIAL EXAMINED. — 30 specimens mounted in ammonium picrate-glycerol. TYPE HOST. — Liza grandisquamis (Mugilidae). SITE OF INFECTION. — Gills, between secondary gill lamellae. TYPE LOCALITY. — Grand Lahou Lagoon, Ivory Coast (5°08’11”N, 5°01’33”E). ETYMOLOGY. — Hamzati is given for Hamzat, name of the son of the first author of this article. FIG. 5. – Ligophorus hamzati n. sp.: morphological structures as in Fig. 1. Abbreviations: see Material & Methods. Scale bar: 30 µm. DESCRIPTION Flattened adult, 580 ± 33 (410-537) [30] long and 80 ± 10 (65-116) [30] wide at gonad level, pharynx: 26 (20-33) [28] larger diameter. Haptor well demarcated with 14 marginal hooks: 13 ± 0.9 (8-17) [360] long. Dorsal anchor with guard two time longer than shaft: a = 43 ± 4 (34-50) [60]; b = 34 ± 3 (28-39) [60]; c = 6 ± 0.7 (4-7) [60]; d = 13 ± 1.5 (9-16) [60]; e = 7 ± 0.8 (6-9) [60]. Dorsal bar V-shaped: 20 ± 2 (17-24) [30] long, 4 ± 0.7 (3-6) [30] wide and 7 ± 1.1 (5-10) [30] high. Ventral anchor: a = 42 ± 2 (36-45) [60]; b = 34 ± 2 (31-37) [60]; c = 7 ± 0.8 (5-10) [60]; d = 11 ± 1.3 (8-14) [60]; e = 7 ± 0.6 (6-9) [60]. Ventral bar, 39 ± 3 (32-47) [30] long and 9 ± 2 (6-13) [30] wide, with small antero-median protuberance and two lateral and symmetrical expansions: 10 ± 3 (4-18) [30] apart. MCO as copulatory tube: 92 ± 6 (81-100) [30] long, passes through a tubular accessory piece, 48 ± 7 (38-59) [30] long, with bifurcated distal extremity, each branches bifurcated also. Vagina: 51 ± 10 (21-67) [30] long. REMARKS This species is distinguished from all Ligophorus species by the shape of the accessory piece of the copulatory organ, which is bifurcated two times at its distal extremities. DISCUSSION The co-existence of fish hosts with a great difference in Ligophorus species richness is not exceptional. Euzet & Suriano (1977) observed only one species (Lig. angustus Euzet & Suriano, 1977) from Chelon labrosus (Risso, 1827) in the Mediterranean Sea, whereas the other mullet species may be parasitized by at least two species. Six species were reported from Liza carinata by Dmitrieva et al. (2012) and from Liza subviridis by Soo and Lim (2012), and at least 14 species were reported from M. cephalus by Dmitrieva et al. (2012). Similar differences in monogenean species richness were reported from cichlid hosts in West Africa by Pariselle et al. (2003). These authors drew a parallel between parasite species richness and host genetic diversity, which were both shaped by fluctuations of host populations through bottleneck or vicariant events. Therefore, in the case of Ligophorus from studied African mugilids, only the population of Liza bandialensis, which is endemic to a very limited area in Senegal, may have suffered numerous bottleneck events, leading to reductions in its size, and in turn to the loss of all its monogenean parasites. Among the three other widely distributed Liza species in Africa, Liza falcipinnis, which was infected by only one Ligophorus species, may have had a more stable history (and thus, a lower genetic and parasitic diversity) when compared to Liza grandisquamis, which was infected by three Ligophorus species and Liza dumerili, which according to Berrada Rkhami et al. (1993) was infected by over ten Ligophorus species.Published as part of Hafidi, Fouzia El, Diamanka, Arfang, Rkhami, Ouafae Berrada & Pariselle, Antoine, 2013, New species of Ligophorus (Monogenea, Ancyrocephalidae), parasite of Liza spp. (Teleostei, Mugilidae) off the Northwestern African coast, pp. 215-225 in Zoosystema 35 (2) on pages 222-224, DOI: 10.5252/z2013n2a6, http://zenodo.org/record/516044
Fostering Graduates’ Critical Thinking with University-Business Collaboration: The Think4Jobs Project
Get PDFLiterature highlights a lack of Higher Education curricula that promotes graduates’ soft skills. Critical Thinking (CT) is considered one of the soft skills associated with higher employment levels. The European-funded project “Critical Thinking for Successful Jobs” (Think4Jobs), currently in progress, aims at strengthening the collaboration between Higher Education Institutions (HEIs) and Labor Market Organizations (LMOs) to design, develop, implement and evaluate CT blended apprenticeships curricula in five disciplines (i.e., Veterinary Medicine, Teacher Education, Business and Economics, Business Informatics, English as a Foreign Language). The curricula are implemented for apprenticeships. We aim to summarize the findings and milestones achieved so far in the project. First, a focus group approach, document analysis, and observation of CT instruction in HEI and LMOs were conducted to assess the state of the art of CT teaching in HEI and the needs of the stakeholders (i.e., HE and LMO) regarding the instruction of CT in HEI apprenticeships and LMO internships. Our results revealed that there is not necessarily a “gap” between HEIs and LMOs concerning CT instruction but rather a difference in understanding and a need to develop a common language between stakeholders. Therefore, as the next step, intensive training for HE instructors and LMO tutors was conducted to establish a common understanding of CT. The results showed no statistical differences in participants’ conceptual understanding of CT, but still drew attention to several misconceptions. Finally, the CT blended apprenticeships curricula were designed as a byproduct of the University-Business Collaboration. Currently, the implementation and evaluation of the effectiveness of the CT blended apprenticeships curricula are carried out.
Research Contribution: Our research within the project is deemed a relative interface that links HEIs and LMOs to establish a sustainable collaboration for developing graduates’ CT
Emerg Infect Dis
No full textWe analyzed sera from diverse mammals of Martha's Vineyard, Massachusetts, for evidence of Francisella tularensis exposure. Skunks and raccoons were frequently seroreactive, whereas white-footed mice, cottontail rabbits, deer, rats, and dogs were not. Tularemia surveillance may be facilitated by focusing on skunks and raccoons
Leveraging structure for optimization in deep learning
Get PDFIn the past decade, neural networks have demonstrated impressive performance in supervised learning. They now power many applications ranging from real- time medical diagnosis to human-sounding virtual assistants through wild animal monitoring. Despite their increasing importance however, they remain difficult to train due to a complex interplay between the learning objective, the optimization algorithm and generalization performance. Indeed, using different loss functions and optimization algorithms lead to trained models with significantly different performances on unseen data.
In this thesis, we focus first on the loss function, for which using a task-specific approach can improve the generalization performance in the small or noisy data setting. Specifically, we consider the top-k classification setting. We show that traditional piecewise-linear top-k loss functions require smoothing to work well with neural networks. However, it is computationally challenging to evaluate the resulting smoothed loss function and its gradient. Indeed, using a naive approach would result in a runtime proportional to the number of combinations of possible k-predictions. Thanks to a connection to polynomial algebra, we develop computationally efficient algorithms to evaluate the smoothed loss function and its gradient. This allows us to train models with stochastic gradient descent (SGD) using the smooth top-k loss function. We show that doing so is more robust to over-fitting than using the standard cross-entropy loss function.
Second, we turn our attention to optimization algorithms. Indeed, while SGD empirically provides good generalization, it requires a manually tuned learning-rate schedule. Obtaining a suitable learning-rate schedule for a given network and data set is a time-consuming and computationally expensive task. In this thesis, we propose novel optimization algorithms to alleviate this issue. In particular, we propose to exploit the structure in three different ways – each one leading to a new optimization algorithm. First, we exploit the piecewise linearity of the activation and loss functions, which results in a difference-of-convex programming approach. Second, we use the compositionality of the model and the loss function with the help of a proximal approach. Third, we exploit the property of interpolating models to derive an adaptive learning-rate for SGD. Empirically, we compare the performance of the three algorithms on various deep learning tasks, and we demonstrate their advantages over state-of-the-art methods while avoiding the need for manual learning rate schedules
Entanglement, nonclassical properties and geometric phase of Raman photon pairs in the presence of time-dependent coupling
No full textAbstractIn this paper, we develop the model of the four-level double Raman pairs by exploiting the required optimal conditions for this system that are feasible with real experimental realization. We investigate qualitatively the entanglement, statistical properties, and geometric phase for the pair of Stokes and anti-Stokes photons in the presence of the time-dependent coupling effect. We show that these quantifiers are very sensitive to the change of the Rabi frequency and time, exhibiting substantial phenomena that are depending on the kind of coupling between the atom and photons. Finally, we explore the relationship between the quantum quantifiers in terms of the physical parameters with and without time-dependent coupling effect
- …
