大规模多标签文本分类

There are 2000 rows in the dataset.

There are 9 duplicate titles.

There are 1991 rows in the deduplicated dataset.
208
275

(1783, 3)

array([list(['cs.CV', 'cs.LG']), list(['cs.CV', 'cs.AI', 'cs.LG']),
       list(['cs.CV', 'cs.AI']), list(['cs.CV']),
       list(['cs.CV', 'cs.LG'])], dtype=object)

Number of rows in training set: 1604
Number of rows in validation set: 90
Number of rows in test set: 89

Vocabulary:

['[UNK]', 'cs.CV', 'cs.LG', 'cs.AI', 'stat.ML', 'eess.IV', 'cs.NE', 'cs.RO', 'cs.CL', 'cs.SI', 'cs.MM', 'math.NA', 'cs.CG', 'cs.CR', 'I.4.6', 'math.OC', 'cs.GR', 'cs.NA', 'cs.HC', 'cs.DS', '68U10', 'stat.ME', 'q-bio.NC', 'math.AP', 'eess.SP', 'cs.DM', '62H30']

Original label: ['cs.CV']

An NVIDIA GPU may be present on this machine, but a CUDA-enabled jaxlib is not installed. Falling back to cpu.

Label-binarized representation: [[0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]]

count    1604.000000
mean      158.151496
std        41.543130
min        25.000000
25%       130.000000
50%       156.000000
75%       184.250000
max       283.000000
Name: summaries, dtype: float64

Abstract: b"For the Domain Generalization (DG) problem where the hypotheses are composed\nof a common representation function followed by a labeling function, we point\nout a shortcoming in existing approaches that fail to explicitly optimize for a\nterm, appearing in a well-known and widely adopted upper bound to the risk on\nthe unseen domain, that is dependent on the representation to be learned. To\nthis end, we first derive a novel upper bound to the prediction risk. We show\nthat imposing a mild assumption on the representation to be learned, namely\nmanifold restricted invertibility, is sufficient to deal with this issue.\nFurther, unlike existing approaches, our novel upper bound doesn't require the\nassumption of Lipschitzness of the loss function. In addition, the\ndistributional discrepancy in the representation space is handled via the\nWasserstein-2 barycenter cost. In this context, we creatively leverage old and\nrecent transport inequalities, which link various optimal transport metrics, in\nparticular the $L^1$ distance (also known as the total variation distance) and\nthe Wasserstein-2 distances, with the Kullback-Liebler divergence. These\nanalyses and insights motivate a new representation learning cost for DG that\nadditively balances three competing objectives: 1) minimizing classification\nerror across seen domains via cross-entropy, 2) enforcing domain-invariance in\nthe representation space via the Wasserstein-2 barycenter cost, and 3)\npromoting non-degenerate, nearly-invertible representation via one of two\nmechanisms, viz., an autoencoder-based reconstruction loss or a mutual\ninformation loss. It is to be noted that the proposed algorithms completely\nbypass the use of any adversarial training mechanism that is typical of many\ncurrent domain generalization approaches. Simulation results on several\nstandard datasets demonstrate superior performance compared to several\nwell-known DG algorithms."
Label(s): ['cs.LG' 'stat.ML']

Abstract: b'Image segmentation of touching objects plays a key role in providing accurate\nclassification for computer vision technologies. A new line profile based\nimaging segmentation algorithm has been developed to provide a robust and\naccurate segmentation of a group of touching corns. The performance of the line\nprofile based algorithm has been compared to a watershed based imaging\nsegmentation algorithm. Both algorithms are tested on three different patterns\nof images, which are isolated corns, single-lines, and random distributed\nformations. The experimental results show that the algorithm can segment a\nlarge number of touching corn kernels efficiently and accurately.'
Label(s): ['cs.CV']

Abstract: b'Semantic image segmentation is a principal problem in computer vision, where\nthe aim is to correctly classify each individual pixel of an image into a\nsemantic label. Its widespread use in many areas, including medical imaging and\nautonomous driving, has fostered extensive research in recent years. Empirical\nimprovements in tackling this task have primarily been motivated by successful\nexploitation of Convolutional Neural Networks (CNNs) pre-trained for image\nclassification and object recognition. However, the pixel-wise labelling with\nCNNs has its own unique challenges: (1) an accurate deconvolution, or\nupsampling, of low-resolution output into a higher-resolution segmentation mask\nand (2) an inclusion of global information, or context, within locally\nextracted features. To address these issues, we propose a novel architecture to\nconduct the equivalent of the deconvolution operation globally and acquire\ndense predictions. We demonstrate that it leads to improved performance of\nstate-of-the-art semantic segmentation models on the PASCAL VOC 2012 benchmark,\nreaching 74.0% mean IU accuracy on the test set.'
Label(s): ['cs.CV']

Abstract: b'Modern deep learning models have revolutionized the field of computer vision.\nBut, a significant drawback of most of these models is that they require a\nlarge number of labelled examples to generalize properly. Recent developments\nin few-shot learning aim to alleviate this requirement. In this paper, we\npropose a novel lightweight CNN architecture for 1-shot image segmentation. The\nproposed model is created by taking inspiration from well-performing\narchitectures for semantic segmentation and adapting it to the 1-shot domain.\nWe train our model using 4 meta-learning algorithms that have worked well for\nimage classification and compare the results. For the chosen dataset, our\nproposed model has a 70% lower parameter count than the benchmark, while having\nbetter or comparable mean IoU scores using all 4 of the meta-learning\nalgorithms.'
Label(s): ['cs.CV' 'cs.LG' 'eess.IV']

Abstract: b'In this work, we propose CARLS, a novel framework for augmenting the capacity\nof existing deep learning frameworks by enabling multiple components -- model\ntrainers, knowledge makers and knowledge banks -- to concertedly work together\nin an asynchronous fashion across hardware platforms. The proposed CARLS is\nparticularly suitable for learning paradigms where model training benefits from\nadditional knowledge inferred or discovered during training, such as node\nembeddings for graph neural networks or reliable pseudo labels from model\npredictions. We also describe three learning paradigms -- semi-supervised\nlearning, curriculum learning and multimodal learning -- as examples that can\nbe scaled up efficiently by CARLS. One version of CARLS has been open-sourced\nand available for download at:\nhttps://github.com/tensorflow/neural-structured-learning/tree/master/research/carls'
Label(s): ['cs.LG']

20498





























Categorical accuracy on the test set: 97.34%.



Abstract: b'High-resolution image segmentation remains challenging and error-prone due to\nthe enormous size of intermediate feature maps. Conventional methods avoid this\nproblem by using patch based approaches where each patch is segmented\nindependently. However, independent patch segmentation induces errors,\nparticularly at the patch boundary due to the lack of contextual information in\nvery high-resolution images where the patch size is much smaller compared to\nthe full image. To overcome these limitations, in this paper, we propose a\nnovel framework to segment a particular patch by incorporating contextual\ninformation from its neighboring patches. This allows the segmentation network\nto see the target patch with a wider field of view without the need of larger\nfeature maps. Comparative analysis from a number of experiments shows that our\nproposed framework is able to segment high resolution images with significantly\nimproved mean Intersection over Union and overall accuracy.'
Label(s): ['cs.CV']
Predicted Label(s): (cs.CV, eess.IV, cs.LG)

Abstract: b"Convolutional neural networks for visual recognition require large amounts of\ntraining samples and usually benefit from data augmentation. This paper\nproposes PatchMix, a data augmentation method that creates new samples by\ncomposing patches from pairs of images in a grid-like pattern. These new\nsamples' ground truth labels are set as proportional to the number of patches\nfrom each image. We then add a set of additional losses at the patch-level to\nregularize and to encourage good representations at both the patch and image\nlevels. A ResNet-50 model trained on ImageNet using PatchMix exhibits superior\ntransfer learning capabilities across a wide array of benchmarks. Although\nPatchMix can rely on random pairings and random grid-like patterns for mixing,\nwe explore evolutionary search as a guiding strategy to discover optimal\ngrid-like patterns and image pairing jointly. For this purpose, we conceive a\nfitness function that bypasses the need to re-train a model to evaluate each\nchoice. In this way, PatchMix outperforms a base model on CIFAR-10 (+1.91),\nCIFAR-100 (+5.31), Tiny Imagenet (+3.52), and ImageNet (+1.16) by significant\nmargins, also outperforming previous state-of-the-art pairwise augmentation\nstrategies."
Label(s): ['cs.CV' 'cs.LG' 'cs.NE']
Predicted Label(s): (cs.CV, cs.LG, stat.ML)


/home/humbulani/tensorflow-env/env/lib/python3.11/site-packages/keras/src/models/functional.py:252: UserWarning: The structure of `inputs` doesn't match the expected structure.
Expected: ['keras_tensor_2']
Received: inputs=Tensor(shape=(2, 20498))
  warnings.warn(msg)