input stringlengths 14 315 | answer stringlengths 9 2.16k | gold_ctxs listlengths 1 15 | ctxs listlengths 11 186 |
|---|---|---|---|
What is Attention Function?How is it calculated? | Attention function relates different positions of the sequence to get the overall representation of the sequence [12]. It can be computer by additive attention method or the dot-product method [16]. Attention have been successfully applied in various NLP tasks such as reading comprehensions and summarizations [5]. | [
12,
16,
5
] | [
{
"id": "1706.03762_all_0",
"text": " Recurrent neural networks, long short-term memory and gated recurrent neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation (35, 2, 5)... |
How did we get 8732 default bounding box for 300x300 input resolution? | The 8732 default bounding boxes are obtained for 300x300 input resolution by stacking more scales of boxes on remaining layers and adjusting scales of boxes if needed [1]. | [
1
] | [
{
"id": "1512.02325_all_0",
"text": " Current state-of-the-art object detection systems are variants of the following approach: hypothesize bounding boxes, resample pixels or features for each box, and apply a high-quality classifier. This pipeline has prevailed on detection benchmarks since the Selective S... |
Why is running information on resource-constrained environments an important concern when GPUs are becoming more commonplace today? | Running information on resource-constrained environments is important concern, because many existing state-of-the-art models require high computational resources that makes them not usable in mobile and embedded applications [0]. | [
0
] | [
{
"id": "1801.04381_all_0",
"text": " Neural networks have revolutionized many areas of machine intelligence, enabling superhuman accuracy for challenging image recognition tasks. However, the drive to improve accuracy often comes at a cost: modern state of the art networks require high computational resour... |
Only a small number of examples (32) are randomly selected to be unlearned. Have the authors tried unlearning much larger portions of the training data and observing the effect on the resulting model? | Results show that forgetting 128 samples at once results in a severe degradation of general LM performance while forgetting 32 samples does not [38]. | [
38
] | [
{
"id": "2210.01504_all_0",
"text": " Recent work has shown that an adversary can extract training data from Pretrained Language Models (LMs) including Personally Identifiable Information (PII) such as names, phone numbers, and email addresses, and other information such as licensed code, private clinical n... |
Can we use Faster R-CNN for human pose estimation ? | Yes we can use Faster R-CNN for human pose estimation [52]. | [
52
] | [
{
"id": "1703.06870_all_0",
"text": " The vision community has rapidly improved object detection and semantic segmentation results over a short period of time. In large part, these advances have been driven by powerful baseline systems, such as the Fast/Faster R-CNN (12, 36) and Fully Convolutional Network ... |
How is it better to decrease the depth by 1 over other values? | Decreasing the depth by 1 requires fewer calls as compared to other values [56]. | [
56
] | [
{
"id": "2212.13894_all_0",
"text": " Automated reasoning, the ability to draw valid conclusions from explicitly provided knowledge, has been a fundamental goal for AI since its early days McCarthy (1959); Hewitt (1969). Furthermore, logical reasoning, especially reasoning with unstructured, natural text is... |
Why does the proposed method introduced EM framework to optimize the model (instead of directly optimizing the loss)? | EM guarantees convergence [14]. | [
14
] | [
{
"id": "2202.02519_all_0",
"text": " Recommender systems have been widely used in many scenarios to provide personalized items to users over massive vocabularies of items. The core of an effective recommender system is to accurately predict users’ interests toward items based on their historical interactio... |
SegNet architecture is inspired from which domain? | SegNet architecture is inspired from generative models and unsupervised learning [1]. | [
1
] | [
{
"id": "1505.07293_all_0",
"text": " Semantic segmentation is an important step towards understanding and inferring different objects and their arrangements observed in a scene. This has wide array of applications ranging from estimating scene geometry, inferring support-relationships among objects to auto... |
How is the "relevance" defined in TREC-COVID dataset? | relevance" is defined as judgements in TREC-COVID dataset [5]. | [
5
] | [
{
"id": "2104.08663_all_0",
"text": " Major natural language processing (NLP) problems rely on a practical and efficient retrieval component as a first step to find relevant information. Challenging problems include open-domain question-answering , claim-verification , duplicate question detection , and man... |
How were the number of tasks from each type of continuous control chosen to create the 31-task benchmark? | They attempt to address this problem and present a benchmark consisting of 31 continuous control tasks [2]. These tasks range from simple tasks, such as cart-pole balancing, to challenging tasks such as high-DOF locomotion, tasks with partial observations, and hierarchically structured tasks [8]. | [
2,
8
] | [
{
"id": "1604.06778_all_0",
"text": " Reinforcement learning addresses the problem of how agents should learn to take actions to maximize cumulative reward through interactions with the environment. The traditional approach for reinforcement learning algorithms requires carefully chosen feature representati... |
They said that cross-encoder make mistakes sometimes. Give an example. This is provided on the paper. | In the third example of Table 8, the cross-encoder linked the mention to the wrong entity "Ancient Greek philosophy", which is likely because of a word "philosophers" in the context [46]. | [
46
] | [
{
"id": "1911.03814_all_0",
"text": " Scale is a key challenge for entity linking; there are millions of possible entities to consider for each mention. To efficiently filter or rank the candidates, existing methods use different sources of external information, including manually curated mention tables Gan... |
How does CNN model learns to ignore areas that appear in both healthy and diseased lungs? | The model learns very small weights in the filters for such areas [17]. | [
17
] | [
{
"id": "1602.03409_all_0",
"text": " Tremendous progress has been made in image recognition, primarily due to the availability of large-scale annotated datasets (i.e. ImageNet (1, 2)) and the recent revival of deep convolutional neural networks (CNN) (3, 4). For data-driven learning, large-scale well-annot... |
How can we define text-video paired dataset? For example, how long each video should be and how long should be the text description? | There is no alinged text and only the videos are used [1]. The authors use only public datasets (and no paired text for videos) [2]. A text description describes an image frame in video so it has limitations to associate between text and phenomenon in video [26]. It needs to depict more detailed stories, is left for future work [27]. Moreover, for all of experiments they applied extrapolation network↑F with frame skip 5 to upsample a 16 frame video to 76 frames [35]. | [
1,
2,
26,
27,
35
] | [
{
"id": "2209.14792_all_0",
"text": " The Internet has fueled collecting billions of (alt-text, image) pairs from HTML pages (Schuhmann et al., 2022), enabling the recent breakthroughs in Text-to-Image (T2I) modeling. However, replicating this success for videos is limited since a similarly sized (text, vid... |
Sometimes the posterior distribution is intractable in nature. What do we mean by intractable? | In this paper, authors want to solve several inference tasks including image denoising, inpainting, and super-resolution by approximating marginal inference of the variable \mathbf{x} [0]. To solve these problems, they introduce a recognition model q_{\boldsymbol{\phi}}(\mathbf{z}|\mathbf{x}): an approximation to the intractable true posterior p_{\boldsymbol{\theta}}(\mathbf{z}|\mathbf{x}) [18]. The posterior distribution is intractable, thus in this paper, authors introduce the stragegy that can be used to derive a lower bound estimator (a stochastic objective function) for a variety of directed graphical models with continuous latent variables [2]. | [
0,
18,
2
] | [
{
"id": "1312.6114_all_0",
"text": " How can we perform efficient approximate inference and learning with directed probabilistic models whose continuous latent variables and/or parameters have intractable posterior distributions? The variational Bayesian (VB) approach involves the optimization of an approxi... |
How does the use of identity shortcuts in this paper, which are parameter-free and always pass information through, differ from gated shortcuts used in highway networks that can be closed and block information? | The identity shortcuts presented in this paper differ from gated shortcuts because it don’t depend on data and always learn residual functions [21]. | [
21
] | [
{
"id": "1512.03385_all_0",
"text": " Deep convolutional neural networks (22, 21) have led to a series of breakthroughs for image classification (21, 50, 40). Deep networks naturally integrate low/mid/high-level features and classifiers in an end-to-end multi-layer fashion, and the “levels” of features can... |
What is the unit of local regions (the meaning of 'local') | [A local region of a pixel consists of pixels with array indices within certain Manhattan distance (kernel size) of the pixel [0]. In a point set sampled from a metric space, the neighborhood of a point is defined by metric distance [17]. | [
0,
17
] | [
{
"id": "1706.02413_all_0",
"text": " We are interested in analyzing geometric point sets which are collections of points in a Euclidean space. A particularly important type of geometric point set is point cloud captured by 3D scanners, e.g., from appropriately equipped autonomous vehicles. As a set, such d... |
Why did the authors measure the perturbations using the L`2 norm? | The authors claim that the DeepFool algorithm is a well-founded baseline for finding adversarial perturbations for state-of-the-art models [13]. Although the use of the l-2 norm is not explicitly justified within the paper, it is a reasonable choice taking into account the scarcity of baseline methods [2]. Also, the method can be easily adapted to any l-p norm and the claims of the paper seem to hold for the l-infinity norm [20]. | [
13,
2,
20
] | [
{
"id": "1511.04599_all_0",
"text": " Deep neural networks are powerful learning models that achieve state-of-the-art pattern recognition performance in many research areas such as bioinformatics (1, 16), speech (12, 6), and computer vision (10, 8). Though deep networks have exhibited very good performance ... |
How are mixture of expert gating functions designed? | In a mixture of experts, the gating function is a network that is learned to choose which experts to assign to each example through the relative discriminative performance of the experts on the sample [38]. | [
38
] | [
{
"id": "1503.02531_all_0",
"text": " Many insects have a larval form that is optimized for extracting energy and nutrients from the environment and a completely different adult form that is optimized for the very different requirements of traveling and reproduction. In large-scale machine learning, we typi... |
What does "temperature" mean in context of the authors' proposed model? | Temperature is a value used in the softmax output layer [7]. The softmax layer converts the logit computed for each class into a probability by comparing with other logits and increasing the temperature produces a softer probability distribution over classes [6]. | [
7,
6
] | [
{
"id": "1503.02531_all_0",
"text": " Many insects have a larval form that is optimized for extracting energy and nutrients from the environment and a completely different adult form that is optimized for the very different requirements of traveling and reproduction. In large-scale machine learning, we typi... |
Why does Yolo outperform R-CNN in other categories such as cat and train ? | The paper does not specifically discuss why YOLO is better for cat and train categories in VOC 2012 dataset and worse for the bottle, sheep, and tv/monitor [6]. Thus, it is difficult to answer this question with only the contents of the paper [68]. | [
6,
68
] | [
{
"id": "1506.02640_all_0",
"text": " Humans glance at an image and instantly know what objects are in the image, where they are, and how they interact. The human visual system is fast and accurate, allowing us to perform complex tasks like driving with little conscious thought. Fast, accurate algorithms fo... |
What is the total number of filters in squeeze convolution layer? | s1x1 is the number of filters in the squeeze layer and it is set s1x1 to be less than (e1x1 + e3x3) -the total number of filters in expand layer of the fire module- to limit the number of input channels to the 3x3 filters [17]. | [
17
] | [
{
"id": "1602.07360_all_0",
"text": " Much of the recent research on deep convolutional neural networks (CNNs) has focused on increasing accuracy on computer vision datasets. For a given accuracy level, there typically exist multiple CNN architectures that achieve that accuracy level. Given equivalent accur... |
Why is each attention head equipped with a 2-layer MLP in particular? | The node embeddings are obtained by processing each query and key through the 2-layer MLP, mapping token representations into the node representation space [14]. | [
14
] | [
{
"id": "2210.15541_all_0",
"text": " The Transformer architecture has been the go-to method for encoding sequential data, due to its superior performance in various tasks such as machine translation , image classification , and protein language modeling . Its key strength stems from the multi-head attenti... |
What is bicubic interpolation ? | Bicubic interpolation is a method to estimate missing pixels in a grid [15]. | [
15
] | [
{
"id": "1505.04597_all_0",
"text": " In the last two years, deep convolutional networks have outperformed the state of the art in many visual recognition tasks, e.g. (7, 3). While convolutional networks have already existed for a long time , their success was limited due to the size of the available traini... |
The complexity of the NAS architecture is independent of the depth of the network and the size of input images. How does it scale to produce better models? | To scale for the better models, authors searched for the best convolutional architectures by searching for the best cell structure [1]. | [
1
] | [
{
"id": "1707.07012_all_0",
"text": " Developing neural network image classification models often requires significant architecture engineering. Starting from the seminal work of on using convolutional architectures (17, 34) for ImageNet classification, successive advancements through architecture enginee... |
Why is there decrease of the performance of the zeor-shot fusion without ATOMIC? | In both framework, the slightly degraded performance of the combination of KGs without ATOMIC could be due to the strong alignment between ATOMIC and SIQA [34]. | [
34
] | [
{
"id": "2206.03715_all_0",
"text": " The ability to understand natural language through commonsense reasoning is one of the core focuses in the field of natural language processing. To measure and study the different aspects of commonsense reasoning, several datasets are developed, such as SocialIQA (Sap e... |
What do the seven input pixel channels represent? | There are seven (7) channels that worth of features to be extracted, giving 24x24x7 = 4032 input features [42]. The first three channels are the image in YUV color space, used because it represents image intensity and color separately [65]. | [
42,
65
] | [
{
"id": "1301.3592_all_0",
"text": " Robotic grasping is a challenging problem involving perception, planning, and control. Some recent works (54, 56, 28, 67) address the perception aspect of this problem by converting it into a detection problem in which, given a noisy, partial view of the object from a ca... |
Why P needs initialization at the start of the training? | Since P is one of the parameters in the loss function, we have to initialise P with some value before training [22]. | [
22
] | [
{
"id": "1708.02002_all_0",
"text": " Current state-of-the-art object detectors are based on a two-stage, proposal-driven mechanism. As popularized in the R-CNN framework , the first stage generates a sparse set of candidate object locations and the second stage classifies each candidate location as one of ... |
Does the paper report empirical benchmarks for performance on non-GPU devices (eg. edge devices such as mobile phones)? | While the paper reports the experimental result on GPU device (Nvidia GTX980 card), the result on non-GPU devices is not included in the paper [25]. | [
25
] | [
{
"id": "1512.01274_all_0",
"text": " The scale and complexity of machine learning (ML) algorithms are becoming increasingly large. Almost all recent ImageNet challenge winners employ neural networks with very deep layers, requiring billions of floating-point operations to process one single sample. The ri... |
What are the range of the number of parameters that for the models used in the study? | The number of parameters range from 5 thousand to 160 million for the models in this study [4]. | [
4
] | [
{
"id": "1602.03409_all_0",
"text": " Tremendous progress has been made in image recognition, primarily due to the availability of large-scale annotated datasets (i.e. ImageNet (1, 2)) and the recent revival of deep convolutional neural networks (CNN) (3, 4). For data-driven learning, large-scale well-annot... |
What does "speed" mean in retrieval contexts? | Index are important as speed in retrieval system [34]. | [
34
] | [
{
"id": "2104.08663_all_0",
"text": " Major natural language processing (NLP) problems rely on a practical and efficient retrieval component as a first step to find relevant information. Challenging problems include open-domain question-answering , claim-verification , duplicate question detection , and man... |
Is it true that "a geometric view synthesis system ONLY performs consistently well when its intermediate predictions of the scene geometry and the camera poses correspond to the physical ground-truth"? | Yes view synthesis system needs to get good geometry otherwise the model would fail on the scenes with more diverse layout and appearance structure [2]. | [
2
] | [
{
"id": "1704.07813_all_0",
"text": " Humans are remarkably capable of inferring ego-motion and the 3D structure of a scene even over short timescales. For instance, in navigating along a street, we can easily locate obstacles and react quickly to avoid them. Years of research in geometric computer vision h... |
How does RELU6 differ from vanilla RELU? | RELU6 is more robust compared to vanilla RELU when used with low-precision computation [22]. | [
22
] | [
{
"id": "1801.04381_all_0",
"text": " Neural networks have revolutionized many areas of machine intelligence, enabling superhuman accuracy for challenging image recognition tasks. However, the drive to improve accuracy often comes at a cost: modern state of the art networks require high computational resour... |
How to set Cs(critical point set) | [The second part of theorem 2 in the paper implies that \mathcal{C}_{S} only contains a bounded number of points, determined by K in Equation(1) [39]. | [
39
] | [
{
"id": "1612.00593_all_0",
"text": " In this paper we explore deep learning architectures capable of reasoning about 3D geometric data such as point clouds or meshes. Typical convolutional architectures require highly regular input data formats, like those of image grids or 3D voxels, in order to perform w... |
How does the graph of the average IOU vs. number of clusters imply the claim that k = 5 is the optimal choice for the complexity/recall tradeoff? | A graph is shown between average IOU vs number of clusters [17]. Number of anchar boxes are then hand-picked by comparing the average IOU closest to the prior [19]. K=5 is choosen because At only 5 priors the centroids perform similarly to 9 anchor boxes with an average IOU of 610 compared to 609 [20]. | [
17,
19,
20
] | [
{
"id": "1612.08242_all_0",
"text": " General purpose object detection should be fast, accurate, and able to recognize a wide variety of objects. Since the introduction of neural networks, detection frameworks have become increasingly fast and accurate. However, most detection methods are still constrained ... |
What does "long-tailed app" mean? | Long-tailed application is a type of software or service that is not installed in a number of users [3]. | [
3
] | [
{
"id": "2005.13303_all_0",
"text": " Personalized mobile business, e.g., recommendations, and advertising, often require effective user representations. For better performance, user modeling in industrial applications often considers as much information as possible, including but not limited to gender, loc... |
Why it is possible to say that multiple group convolutional layers works efficiently without weakening representation? | It is clearly stated in the paper that having group convolutions is a trade-off between representative capability and the computational cost of the model [10]. The ShuffleNet allows stacking multiple group convolutions with an appropriate number of groups because of channel shuffle and it is empirically shown in the paper [21]. However, it is also noted that having too many groups might sometimes damage the performance [22]. Thus, multiple group convolutions work efficiently only when the number of groups is chosen carefully and channel shuffle is used [23]. | [
10,
21,
22,
23
] | [
{
"id": "1707.01083_all_0",
"text": " Building deeper and larger convolutional neural networks (CNNs) is a primary trend for solving major visual recognition tasks (21, 9, 33, 5, 28, 24). The most accurate CNNs usually have hundreds of layers and thousands of channels (9, 34, 32, 40), thus requiring computa... |
What was the value of maximum length T used for the experiment and how was the ratio of sequences that longer than length T? | The value of T is not mentioned, and neither is the ratio of sequences that exceed T in length [8]. | [
8
] | [
{
"id": "2202.02519_all_0",
"text": " Recommender systems have been widely used in many scenarios to provide personalized items to users over massive vocabularies of items. The core of an effective recommender system is to accurately predict users’ interests toward items based on their historical interactio... |
What datasets did this paper used for? | They used three datasets, the original TACRED Zhang et al [11]. (2017), TACREV Alt et al [18]. | [
11,
18
] | [
{
"id": "2102.01373_all_0",
"text": " As one of the fundamental information extraction (IE) tasks, relation extraction (RE) aims at identifying the relationship(s) between two entities in a given piece of text from a pre-defined set of relationships of interest. For example, given the sentence “Bill Gates f... |
What is the difference between IsTopVoice and PositionInChord? | IsTopVoice is different from PositionInChord in that an index 1 of IsTopVoice represents the uppermost voice while that of PositionInChord represents the lowermost voice [10]. | [
10
] | [
{
"id": "2208.14867_all_0",
"text": " Computational modeling of expressive music performance focuses on mimicking human behaviors that convey the music (1, 2). For piano performance, one common task is to render an expressive performance from a quantized musical score. It aims to reproduce the loudness and ... |
Does the graph structure include the proportion of triangles and clustering coefficient of a node ? | Yes it is [38]. | [
38
] | [
{
"id": "1706.02216_all_0",
"text": " Low-dimensional vector embeddings of nodes in large graphs111While it is common to refer to these data structures as social or biological networks, we use the term graph to avoid ambiguity with neural network terminology. have proved extremely useful as feature inputs f... |
Is it true that YOLO is highly generalizable and performs well in new unseen data ? | The generalizability of YOLO to unseen data is evaluated by training it on natural images and testing with artwork from Picasso and People-Art datasets [51]. Since YOLO can reason about the entire image and learn the contextual information about the class and its appearance, it shows much better generalizability compared to other state-of-the-art techniques [6]. Generalizability to other domains besides artwork is not mentioned in the paper [7]. | [
51,
6,
7
] | [
{
"id": "1506.02640_all_0",
"text": " Humans glance at an image and instantly know what objects are in the image, where they are, and how they interact. The human visual system is fast and accurate, allowing us to perform complex tasks like driving with little conscious thought. Fast, accurate algorithms fo... |
What is the reason the standard deviation is not shown in the table? | The standard deviation is not shown in the table because it is shown in the Appendix [31]. | [
31
] | [
{
"id": "2210.01504_all_0",
"text": " Recent work has shown that an adversary can extract training data from Pretrained Language Models (LMs) including Personally Identifiable Information (PII) such as names, phone numbers, and email addresses, and other information such as licensed code, private clinical n... |
How many generators have to be trained if you train the cross-domain model for a 5-domain image translation task using the previous approach? | it takes 20 generators in order to do 5-domain image translation task [3]. | [
3
] | [
{
"id": "1711.09020_all_0",
"text": " The task of image-to-image translation is to change a particular aspect of a given image to another, e.g., changing the facial expression of a person from smiling to frowning (see Fig. StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Tra... |
Explain Mode task in Decimal & word operation with examples. | Decimal & word operation is task of subtracting or dividing two numbers [9]. | [
9
] | [
{
"id": "2210.12302_all_0",
"text": " Pretrained Language Models (LMs) have shown singular succcess on a range of natural language understandings tasks, to the extent that they have become foundational for contemporary NLP systems. Several works have investigated why pretraining works so well Warstadt et al... |
Which deep neural network architectures were used for experimental comparison of DeepFool algorithm with existing methods? | Although the conclusion of the paper claims that 8 different classifiers were used, we can only see 6 classifiers with different datasets: 2-layer fully-connected network (MNIST), 2-layer LeNet (MNIST), 3-layer LeNet (CIFAR-10), NIN (CIFAR-10), CaffeNet (ILSVRC 2012), and GoogLeNet (ILSVRC 2012) [14]. | [
14
] | [
{
"id": "1511.04599_all_0",
"text": " Deep neural networks are powerful learning models that achieve state-of-the-art pattern recognition performance in many research areas such as bioinformatics (1, 16), speech (12, 6), and computer vision (10, 8). Though deep networks have exhibited very good performance ... |
What does model transferability mean? | Applying NAS to a large dataset is computationally expensive [1]. So the authors find the good architecture on a proxy dataset and then transfer the learned architecture to ImageNet [35]. | [
1,
35
] | [
{
"id": "1707.07012_all_0",
"text": " Developing neural network image classification models often requires significant architecture engineering. Starting from the seminal work of on using convolutional architectures (17, 34) for ImageNet classification, successive advancements through architecture enginee... |
Why was the IOU metric used and not other segmentation metrics such as the Dice coefficient? | The paper does not include explicit discussion regarding using the IOU metric or Dice coefficient, so it is difficult to answer the question just by the information in the paper [61]. In general, the IOU method is used for object detection, while the Dice coefficient is used for image segmentation [60]. | [
61,
60
] | [
{
"id": "1506.02640_all_0",
"text": " Humans glance at an image and instantly know what objects are in the image, where they are, and how they interact. The human visual system is fast and accurate, allowing us to perform complex tasks like driving with little conscious thought. Fast, accurate algorithms fo... |
How the MobileNet model was trained and why it was different than training the large networks? | Authors use less regularization and data augmentation for MobileNets because of less overfitting [29]. | [
29
] | [
{
"id": "1704.04861_all_0",
"text": " Convolutional neural networks have become ubiquitous in computer vision ever since AlexNet popularized deep convolutional neural networks by winning the ImageNet Challenge: ILSVRC 2012 . The general trend has been to make deeper and more complicated networks in order t... |
What does non-differentiable mean here? If the problem with previous metrics is that they are not per-token differentiable then why are they looking for a way to optimize non-differentiable objectives? | A formal definition of non-differentiability has not been provided by the authors [0]. | [
0
] | [
{
"id": "2210.01241_all_0",
"text": " The ultimate aim of language technology is to interact with humans. However, most language models are trained without direct signals of human preference, with supervised target strings serving as (a sometimes crude) proxy. One option to incorporate user feedback is via ... |
What is the definition of 'meta-learning'? | Meta-learning is a learning method that enables a model to quickly adapt to tasks during learning [0]. Few-shot meta learning means learning a good model within a small number of iterations using only a few training images [1]. Reinforcement learning means that a good policy can be learned using a small number of experiences [18]. | [
0,
1,
18
] | [
{
"id": "1703.03400_all_0",
"text": " Learning quickly is a hallmark of human intelligence, whether it involves recognizing objects from a few examples or quickly learning new skills after just minutes of experience. Our artificial agents should be able to do the same, learning and adapting quickly from onl... |
What is the difference between the learning rate of the SRCNN and the learning rate of the proposed model? | Not only does SRCNN uses different learning rates for its layers, while the proposed model uses the same learning rate for all of its layers, but the proposed model's initial learning rate is also 10000 times greater than SRCNN's [17]. | [
17
] | [
{
"id": "1511.04587_all_0",
"text": " We address the problem of generating a high-resolution (HR) image given a low-resolution (LR) image, commonly referred as single image super-resolution (SISR) , , . SISR is widely used in computer vision applications ranging from security and surveillance imaging to med... |
What are the two steps specified to generate data? | The process consists of two steps: (1) a value \mathbf{z}^{(i)} is generated from some prior distribution p_{\boldsymbol{\theta}^{*}}(\mathbf{z}); (2) a value \mathbf{x}^{(i)} is generated from some conditional distribution p_{\boldsymbol{\theta}^{*}}(\mathbf{x}|\mathbf{z}) [12]. | [
12
] | [
{
"id": "1312.6114_all_0",
"text": " How can we perform efficient approximate inference and learning with directed probabilistic models whose continuous latent variables and/or parameters have intractable posterior distributions? The variational Bayesian (VB) approach involves the optimization of an approxi... |
What is the difference of RocketQAv1 and RocketQAv2 model? | RocketQAv1 trains dual-encoder and cross-encoder in a cascade manner, which leverages the powerful cross-encoder to empower the dual-encoder [10]. While it inherits the parameters from RocketQAv1, RocketQAv2 extends the first version through a novel approach that jointly trains the dense passage retriever and passage re-ranker, and by using a large PLM for data augmentation and denoising (ie, a distillation procedure) [42]. | [
10,
42
] | [
{
"id": "2204.11673_all_0",
"text": " Passage Re-ranking is a crucial stage in modern information retrieval systems, which aims to reorder a small set of candidate passages to be presented to users. To put the most relevant passages on top of a ranking list, a re-ranker is usually designed with powerful cap... |
How does it guess if it is an object or background? (Since using K classes and not K+1, 1 extra for background class) | To decide whether there is an object in the image or not, IoU is used [28]. | [
28
] | [
{
"id": "1708.02002_all_0",
"text": " Current state-of-the-art object detectors are based on a two-stage, proposal-driven mechanism. As popularized in the R-CNN framework , the first stage generates a sparse set of candidate object locations and the second stage classifies each candidate location as one of ... |
What is the effect of changing stride of the convolution? | By reducing the stride of the convolution with the hole algorithm, the authors were able to improve mAP by 26 points [17]. | [
17
] | [
{
"id": "1605.06409_all_0",
"text": " A prevalent family (8, 6, 18) of deep networks for object detection can be divided into two subnetworks by the Region-of-Interest (RoI) pooling layer : (i) a shared, “fully convolutional” subnetwork independent of RoIs, and (ii) an RoI-wise subnetwork that does not shar... |
Is joint model achieved state-of-the-art? | Yes, Joint model achieve superior performance [2]. | [
2
] | [
{
"id": "2108.13530_all_0",
"text": " In this paper we explore a principled approach to solve entity linking (EL) jointly with coreference resolution (coref). Concretely, we formulate coref+EL as a single structured task over directed trees that conceives EL and coref as two complementary components: a core... |
How do the authors come to the conclusion that the thinness is what is causing regions to be masked improperly? | The simultaneous estimation of structure and motion is a well studied problem where traditional methods suffer from thin structures [3]. | [
3
] | [
{
"id": "1704.07813_all_0",
"text": " Humans are remarkably capable of inferring ego-motion and the 3D structure of a scene even over short timescales. For instance, in navigating along a street, we can easily locate obstacles and react quickly to avoid them. Years of research in geometric computer vision h... |
Why is the assumption -- that the predicted box overlaps with the ground truth label by 0.3 IOU -- necessary? | To backpropagate classification loss highest probability bounding box class is used [65]. | [
65
] | [
{
"id": "1612.08242_all_0",
"text": " General purpose object detection should be fast, accurate, and able to recognize a wide variety of objects. Since the introduction of neural networks, detection frameworks have become increasingly fast and accurate. However, most detection methods are still constrained ... |
How can the SGVB be optimised? | Authors form Monte Carlo estimates of expectations of some function f(\mathbf{z}) wrt [0]. q_{\boldsymbol{\phi}}(\mathbf{z}|\mathbf{x}) as Eq [10]. (5) [27]. They use this technique to the variational lower bound (eq [28]. (2)), yielding generic Stochastic Gradient Variational Bayes (SGVB) estimator [32]. | [
0,
10,
27,
28,
32
] | [
{
"id": "1312.6114_all_0",
"text": " How can we perform efficient approximate inference and learning with directed probabilistic models whose continuous latent variables and/or parameters have intractable posterior distributions? The variational Bayesian (VB) approach involves the optimization of an approxi... |
Are the softmax values of different sets of co-hyponyms compared? | Classification approaches use a softmax layer across all categories to predict the final probability of all classes [49]. This technique would fail for models which combine datasets having similar classes [54]. To overcome the proposed model also use a softmax over all sysnsets that are co-hyponyms [59]. | [
49,
54,
59
] | [
{
"id": "1612.08242_all_0",
"text": " General purpose object detection should be fast, accurate, and able to recognize a wide variety of objects. Since the introduction of neural networks, detection frameworks have become increasingly fast and accurate. However, most detection methods are still constrained ... |
What is space meant in "space-efficient in practice" ?Does it mean space of search for solutions? or space in memory? | Dot product attention is much faster and more space-efficient than the additive attention [16]. | [
16
] | [
{
"id": "1706.03762_all_0",
"text": " Recurrent neural networks, long short-term memory and gated recurrent neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation (35, 2, 5)... |
What do you mean by NMS(Non-Maximal Suppression) ?
| NMS is used in SSD to get the final predictions [4]. | [
4
] | [
{
"id": "1512.02325_all_0",
"text": " Current state-of-the-art object detection systems are variants of the following approach: hypothesize bounding boxes, resample pixels or features for each box, and apply a high-quality classifier. This pipeline has prevailed on detection benchmarks since the Selective S... |
What is the bias responsible for the system performing poorly without mask-based scaling? | As shown in Table III our mask-based scaling technique at the visible layer improves grasping results by over 25% for both metrics [50]. | [
50
] | [
{
"id": "1301.3592_all_0",
"text": " Robotic grasping is a challenging problem involving perception, planning, and control. Some recent works (54, 56, 28, 67) address the perception aspect of this problem by converting it into a detection problem in which, given a noisy, partial view of the object from a ca... |
Why is neural integration of different KGs better than symbolic KG integration? | Rather than such symbolic KG integration with the inevitable loss of knowledge, in this work, we explore the neural KG integration leveraging the multiple KGs without additional processing and alignment information between KG and task [6]. | [
6
] | [
{
"id": "2206.03715_all_0",
"text": " The ability to understand natural language through commonsense reasoning is one of the core focuses in the field of natural language processing. To measure and study the different aspects of commonsense reasoning, several datasets are developed, such as SocialIQA (Sap e... |
Would it be possible to reduce the asymptotic cost of GMPool from cubic to quadratic, yet retain its expressive power? | One future direction to enhance scalability of GMPool is to incorporate faster decomposition modules such as randomized approximation methods [31]. | [
31
] | [
{
"id": "2209.02939_all_0",
"text": " Graph Neural Networks (GNNs) learn representations of individual nodes based on the connectivity structure of an input graph. For graph-level prediction tasks, the standard procedure globally pools all the node features into a single graph representation without weight ... |
What are the examples of the tools that enable understanding of Neural Networks for newcomers in deep learning? | The paper talked about two main tools; the first is a software tool to plot activations of each trained layer of a network, for images or videos [3]. Second is introducing new regularization ways to help with understanding learned features through networkThese tools are supposed to help newcomers in deep learning to have better intuitions for hidden interpretations of well known structures and give motivations for more new ideas [32]. | [
3,
32
] | [
{
"id": "1506.06579_all_0",
"text": " The last several years have produced tremendous progress in training powerful, deep neural network models that are approaching and even surpassing human abilities on a variety of challenging machine learning tasks (Taigman et al., 2014; Schroff et al., 2015; Hannun et a... |
What motivated the authors to choose the Pascal VOC 2007 dataset to compare YOLO with other models ? | Although the paper does not give explicit reasons why Pascal VOC 2007 dataset was chosen for comparison, we can make an educated guess [51]. It seems like Pascal VOC 2007 is one of the popular datasets for object detection [59]. | [
51,
59
] | [
{
"id": "1506.02640_all_0",
"text": " Humans glance at an image and instantly know what objects are in the image, where they are, and how they interact. The human visual system is fast and accurate, allowing us to perform complex tasks like driving with little conscious thought. Fast, accurate algorithms fo... |
Should we need to increase the size of the network to reach considerable efficiency after adding region-wise layers? | No, as seen by the model presented in the paper, performance can be maintained or even improved while keeping computational efficiency high and not adding region-wise layers [3]. | [
3
] | [
{
"id": "1605.06409_all_0",
"text": " A prevalent family (8, 6, 18) of deep networks for object detection can be divided into two subnetworks by the Region-of-Interest (RoI) pooling layer : (i) a shared, “fully convolutional” subnetwork independent of RoIs, and (ii) an RoI-wise subnetwork that does not shar... |
Define KL divergence. | In this work, KL divergence is computed between the probability distribution of a specialist model or generalist full model and the full probability distribution over all classes [31]. | [
31
] | [
{
"id": "1503.02531_all_0",
"text": " Many insects have a larval form that is optimized for extracting energy and nutrients from the environment and a completely different adult form that is optimized for the very different requirements of traveling and reproduction. In large-scale machine learning, we typi... |
Since YOLO sees the entire image during training and testing, doesn't it influence badly on its performance ? | The paper discusses both advantages and disadvantages of looking at the image as a whole [37]. Processing the entire image, let YOLO be end-to-end, thus predicting bounding boxes and class probabilities directly [6]. Also, it shows good generalizability to other domains and it copes with background objects much better compared to Fast R-CNN due to looking at the image as a whole [64]. However, to make the entire image consumable to the model, dividing the image into grids and limiting the number of bounding boxes are performed [7]. Because of these and other design decisions, YOLO shows inferior accuracy compared to state-of-the-art detectors [74]. Especially it struggles with localizing objects, small objects, and objects close to each other [8]. | [
37,
6,
64,
7,
74,
8
] | [
{
"id": "1506.02640_all_0",
"text": " Humans glance at an image and instantly know what objects are in the image, where they are, and how they interact. The human visual system is fast and accurate, allowing us to perform complex tasks like driving with little conscious thought. Fast, accurate algorithms fo... |
Matching network uses BERT. Is this true? | It's not true [29]. | [
29
] | [
{
"id": "1711.04043_all_0",
"text": " Supervised end-to-end learning has been extremely successful in computer vision, speech, or machine translation tasks, thanks to improvements in optimization technology, larger datasets and streamlined designs of deep convolutional or recurrent architectures. Despite th... |
What are tasks to show how well English models tend to be multilingual? | Language composition estimation and POS tagging can measure multilingual performance [14]. | [
14
] | [
{
"id": "2204.08110_all_0",
"text": " Pretrained language models have become an integral part of NLP systems. They come in two flavors: monolingual, where the model is trained on text from a single language, and multilingual, where the model is jointly trained on data from many different languages. Monoling... |
How did the authors compute the contributions of the pixels in order to clip the pixels with smaller contributions? | Calculating absolute difference between some neuron activation of an input and the activation for same input without certain pixel can be considered a way of measuring the contribution of that pixel in the total response of the neuron [28]. | [
28
] | [
{
"id": "1506.06579_all_0",
"text": " The last several years have produced tremendous progress in training powerful, deep neural network models that are approaching and even surpassing human abilities on a variety of challenging machine learning tasks (Taigman et al., 2014; Schroff et al., 2015; Hannun et a... |
Can 512x512 be considered high resolution? | Information about Wether 512*512 is high resolution or not is not explicitly provided in the paper [22]. | [
22
] | [
{
"id": "1809.11096_all_0",
"text": " The state of generative image modeling has advanced dramatically in recent years, with Generative Adversarial Networks (GANs, Goodfellow et al. (2014)) at the forefront of efforts to generate high-fidelity, diverse images with models learned directly from data. GAN trai... |
How were the two tasks from each category chosen? | For the other tasks, we try both of the best hyperparameters found in the same category, and report the better performance of the two [34]. | [
34
] | [
{
"id": "1604.06778_all_0",
"text": " Reinforcement learning addresses the problem of how agents should learn to take actions to maximize cumulative reward through interactions with the environment. The traditional approach for reinforcement learning algorithms requires carefully chosen feature representati... |
Give two examples of conventional image descriptors that can be used for object detection and segmentation in medical image analysis. | Two examples of conventional image descriptors for object detection and segmentation in the medical image field are scale-invariant feature transform (SIFT) and histogram of oriented gradients (HOG) [3]. | [
3
] | [
{
"id": "1602.03409_all_0",
"text": " Tremendous progress has been made in image recognition, primarily due to the availability of large-scale annotated datasets (i.e. ImageNet (1, 2)) and the recent revival of deep convolutional neural networks (CNN) (3, 4). For data-driven learning, large-scale well-annot... |
What properties of costrastive self-supervised learning have attracted attention from researchers in the recommendation field? | Sequences of user behavior can be maximally separated or brought together by means of contrastive SSL [12]. | [
12
] | [
{
"id": "2202.02519_all_0",
"text": " Recommender systems have been widely used in many scenarios to provide personalized items to users over massive vocabularies of items. The core of an effective recommender system is to accurately predict users’ interests toward items based on their historical interactio... |
How many entities and relations does ConceptNet has? | ConceptNet is a general knowledge graph and, in this work, they merged relation types in the graph to construct a multi-relational graph with 17 relation types [41]. | [
41
] | [
{
"id": "2204.11673_all_0",
"text": " Passage Re-ranking is a crucial stage in modern information retrieval systems, which aims to reorder a small set of candidate passages to be presented to users. To put the most relevant passages on top of a ranking list, a re-ranker is usually designed with powerful cap... |
How is next sentence prediction (NSP) different from sentence relation prediction (SRP)? | Compared to conventional Next Sentence Prediction (NSP), Sentence Relation Prediction (SRP) aims to predict whether a given sentence is the next sentence, previous sentence relation, or no relation with another sentence [38]. | [
38
] | [
{
"id": "2204.11673_all_0",
"text": " Passage Re-ranking is a crucial stage in modern information retrieval systems, which aims to reorder a small set of candidate passages to be presented to users. To put the most relevant passages on top of a ranking list, a re-ranker is usually designed with powerful cap... |
What is the contribution of this paper? | They modify the supernet architecture by varying the number of blocks in stages, and adds MixConv to the search space [26]. This enables more diverse combinations of kernel sizes and expansion ratios than original MixConv [36]. Moreover, they eases the search process [50]. As a result, they could find a better network than existing network models [56]. Note that their method can be used to any type of NPU [77]. | [
26,
36,
50,
56,
77
] | [
{
"id": "2009.02009_all_0",
"text": " As there are growing needs of deep learning applications based on convolutional neural network(CNN) in embedded systems, improving the accuracy of CNNs under a given set of constraints on latency and energy consumption has brought keen interest to researchers as a chall... |
What is meant by "linear sweep" in hyperparameter space? | linear sweep can be seen as a regular increment in the values of some regularization hyperparameter (from leftmost where there is no regularization to rightmost where strong regularization occur ) to see the variation of their effects on the corresponding activations [18]. | [
18
] | [
{
"id": "1506.06579_all_0",
"text": " The last several years have produced tremendous progress in training powerful, deep neural network models that are approaching and even surpassing human abilities on a variety of challenging machine learning tasks (Taigman et al., 2014; Schroff et al., 2015; Hannun et a... |
Authors used a modified version of DropPath regularization named ScheduledDropPath. What is modified? | In ScheduledDropPath, each path in the cell is dropped out with a probability that is linearly increased over the course of training [22]. | [
22
] | [
{
"id": "1707.07012_all_0",
"text": " Developing neural network image classification models often requires significant architecture engineering. Starting from the seminal work of on using convolutional architectures (17, 34) for ImageNet classification, successive advancements through architecture enginee... |
For creating feature maps, did SSD extracted features from single or multiple layers of the network? | SSD uses features from multiple layers of the network for creating the feature maps [11]. | [
11
] | [
{
"id": "1512.02325_all_0",
"text": " Current state-of-the-art object detection systems are variants of the following approach: hypothesize bounding boxes, resample pixels or features for each box, and apply a high-quality classifier. This pipeline has prevailed on detection benchmarks since the Selective S... |
Why do the authors claim that human feedback may be less important when their experiments showed that InstructGPT, which had human-generated data, outperformed their model without human-generated data? | The authors claim that human feedback might not be essential since their model is able to almost meet the performance of InstructGPT despite not having access to private human-generated training data or manual annotations [2]. They claim that their model's success, of almost reaching InstructGPT performance with only a 5% gap is a strong indication that human data, while useful is not necessarily essential for teaching models how to follow instructions [22]. Additionally, they point out that their work is merely a beginning step in research in this field - while numerous studies have successfully used human annotations to improve performance, studies that attempt to remove the human requirement have not been as explored [25]. Also, the authors do acknowledge that the truth is somewhere in between the two extremes of (1) human instructional data is essential, or (2) such data is largely optional, and similar results can be achieved without it [29]. | [
2,
22,
25,
29
] | [
{
"id": "2212.10560_all_0",
"text": " The recent NLP literature has witnessed a tremendous amount of activity in building models that can follow natural language instructions (Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022; Wang et al., 2022; Ouyang et al., 2022; Chung et al., 2022, i.a.). These d... |
Why can’t we use sampling based solutions instead of this algorithm in case of large datasets? | It is hard to use sampling based solutions because batch optimization with so much data is too expensive [0]. If you want to inference in almost any model with continuous latent variables and/or parameters, sampling based solution is not applicable [1]. For very low-dimensional latent space it is possible to estimate the marginal likelihood of the learned generative models using an MCMC estimator which is one of sampling based solution [30]. But we need to deal with high dimensional data and the AEVB algorithm is useful [4]. | [
0,
1,
30,
4
] | [
{
"id": "1312.6114_all_0",
"text": " How can we perform efficient approximate inference and learning with directed probabilistic models whose continuous latent variables and/or parameters have intractable posterior distributions? The variational Bayesian (VB) approach involves the optimization of an approxi... |
Out of conditional and unconditional decoder blocks, which one is better? | The author talk both advantage and disadvantage of conditional and unconditional decoder blocks [17]. | [
17
] | [
{
"id": "1502.04681_all_0",
"text": " Understanding temporal sequences is important for solving many problems in the AI-set. Recently, recurrent neural networks using the Long Short Term Memory (LSTM) architecture (Hochreiter & Schmidhuber, 1997) have been used successfully to perform various supervised seq... |
How do feature-based methods work in face recognition? | The only feature-based method that is mentioned in the paper is the local-feature-based methods from the 2000s of Gabor and LBP that tried local filtering to extract invariant properties [0]. | [
0
] | [
{
"id": "1804.06655_all_0",
"text": " Face recognition (FR) has been the prominent biometric technique for identity authentication and has been widely used in many areas, such as military, finance, public security and daily life. FR has been a long-standing research topic in the CVPR community. In the early... |
Why did the authors chose to train YOLO using VGG-16 and not other neural network architecture ? | In fact, the base YOLO model and Fast YOLO have used GoogLeNet-inspired architecture to VGG-16 [54]. The authors claim that they have trained it with VGG-16 and it had better accuracy, however, it was too slow to be real-time [18]. The YOLO model is first pretrained on the ImageNet 1000-class competition dataset and later trained on training and validation data of the Pascal VOC 2007 dataset [21]. | [
54,
18,
21
] | [
{
"id": "1506.02640_all_0",
"text": " Humans glance at an image and instantly know what objects are in the image, where they are, and how they interact. The human visual system is fast and accurate, allowing us to perform complex tasks like driving with little conscious thought. Fast, accurate algorithms fo... |
What is maximum spanning tree problem? | They lead to a maximum spanning tree problem in their global approach by proposing bidirectional connections between mentions [7]. | [
7
] | [
{
"id": "2108.13530_all_0",
"text": " In this paper we explore a principled approach to solve entity linking (EL) jointly with coreference resolution (coref). Concretely, we formulate coref+EL as a single structured task over directed trees that conceives EL and coref as two complementary components: a core... |
How many categories used in non-English text classifier? | Non-English text classifier uses six categories [11]. | [
11
] | [
{
"id": "2204.08110_all_0",
"text": " Pretrained language models have become an integral part of NLP systems. They come in two flavors: monolingual, where the model is trained on text from a single language, and multilingual, where the model is jointly trained on data from many different languages. Monoling... |
What is the key difference between inductive and transductive learning? | General inductive transfer learning is, given a static source task and any target task where the source task and target task are not equal, to improve the performance of the target task [0]. This occurs by fine-tuning a model that has been pretrained on other datasets [11]. What transductive transfer learning is and the difference between transductive and inductive transfer learning cannot be answered from this paper [2]. | [
0,
11,
2
] | [
{
"id": "1801.06146_all_0",
"text": " Inductive transfer learning has had a large impact on computer vision (CV). Applied CV models (including object detection, classification, and segmentation) are rarely trained from scratch, but instead are fine-tuned from models that have been pretrained on ImageNet, MS... |
Does β-VAE's objective control overlap ? | The \beta-vae objective to show that its contribution to disentangling is primarily through direct control of the level of overlap between encodings of the data, expressed by maximising the entropy of the encoding distribution [25]. \beta-vae’s ability to encourage disentanglement is predominantly through direct control over the level of overlap [29]. Hence, β-VAE's objective control overlap [31]. | [
25,
29,
31
] | [
{
"id": "1812.02833_all_0",
"text": " An oft-stated motivation for learning disentangled representations of data with deep generative models is a desire to achieve interpretability (5, 10)—particularly the decomposability (see §3.2.1 in 33) of latent representations to admit intuitive explanations. Most wor... |
What is the final improved architecture used by authors for experiments in this paper? | They use variable width with 2 residual blocks per resolution, multiple heads with 64 channels per head, attention at 32, 16 and 8 resolutions, BigGAN residual blocks for up and downsampling, and adaptive group normalization for injecting timestep and class embeddings into residual blocks [18]. | [
18
] | [
{
"id": "2105.05233_all_0",
"text": " Over the past few years, generative models have gained the ability to generate human-like natural language Brown et al. (2020), infinite high-quality synthetic images Brock et al. (2018); Karras et al. (2019b); Razavi et al. (2019) and highly diverse human speech and mu... |
How does YOLO9000 achieve the feat of predicting detections for classes despite not having labelled data for them? | YOLO9000 can perform well on classes which it has not seen during training is because of its WordTree based data combination method from various sources [66]. For example it can learn animal categories which it has not seen because objectness properties in case of such objects can be generalized well [67]. | [
66,
67
] | [
{
"id": "1612.08242_all_0",
"text": " General purpose object detection should be fast, accurate, and able to recognize a wide variety of objects. Since the introduction of neural networks, detection frameworks have become increasingly fast and accurate. However, most detection methods are still constrained ... |
What kinds of distribution shifts are considered for evaluating retrievers on out-of-distribution datasets? | BioASQ, or task-shifts like in Touché-2020 distribution shifts are considered for evaluating retrievers on out-of-distribution datasets [29]. | [
29
] | [
{
"id": "2104.08663_all_0",
"text": " Major natural language processing (NLP) problems rely on a practical and efficient retrieval component as a first step to find relevant information. Challenging problems include open-domain question-answering , claim-verification , duplicate question detection , and man... |
Are the results similar for other variants of values, given test set has only 1000 examples? | Yes, the results are similar for other variants of values [41]. | [
41
] | [
{
"id": "2212.13894_all_0",
"text": " Automated reasoning, the ability to draw valid conclusions from explicitly provided knowledge, has been a fundamental goal for AI since its early days McCarthy (1959); Hewitt (1969). Furthermore, logical reasoning, especially reasoning with unstructured, natural text is... |
How could this vastness be defined or quantitatively measured? | They employed annotators to make prompts and filtered out them correctly [28]. Evaluation is done about video quality and faithfulness [29]. | [
28,
29
] | [
{
"id": "2209.14792_all_0",
"text": " The Internet has fueled collecting billions of (alt-text, image) pairs from HTML pages (Schuhmann et al., 2022), enabling the recent breakthroughs in Text-to-Image (T2I) modeling. However, replicating this success for videos is limited since a similarly sized (text, vid... |
How Inception-ResNet v1 compare with Inception-ResNet v2 in terms of structure, stem and settings? | Both Inception-ResNet-v1 and Inception-ResNet-v2 are Inception style networksthat utilize residual connections instead of filter concatenation [20]. | [
20
] | [
{
"id": "1602.07261_all_0",
"text": " Since the 2012 ImageNet competition winning entry by Krizhevsky et al , their network “AlexNet” has been successfully applied to a larger variety of computer vision tasks, for example to object-detection , segmentation , human pose estimation , video classification , o... |
How did they finetune GPT3? | The authors fine-tuned the GPT3 model via an OpenAI API [13]. The model, which is also called "da vinci", was dine-tuned by the authors for 2 epochs with a prompt loss weight set to zero [19]. | [
13,
19
] | [
{
"id": "2212.10560_all_0",
"text": " The recent NLP literature has witnessed a tremendous amount of activity in building models that can follow natural language instructions (Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022; Wang et al., 2022; Ouyang et al., 2022; Chung et al., 2022, i.a.). These d... |
What is the purpose of using shortcut connections in residual learning? | [The purpose of using shortcut connections in residual learning is to ease the comparison between plain and residual networks] [17]. | [
17
] | [
{
"id": "1512.03385_all_0",
"text": " Deep convolutional neural networks (22, 21) have led to a series of breakthroughs for image classification (21, 50, 40). Deep networks naturally integrate low/mid/high-level features and classifiers in an end-to-end multi-layer fashion, and the “levels” of features can... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.