paper_id
stringlengths
9
12
venue
stringclasses
139 values
year
stringclasses
7 values
paper_title
stringlengths
0
181
paper_authors
stringlengths
4
925
paper_abstract
stringlengths
1
5k
paper_keywords
stringlengths
2
436
paper_content
stringlengths
0
100k
review_id
stringlengths
9
12
review_title
stringlengths
0
500
review_rating
stringclasses
61 values
review_text
stringlengths
2
28.3k
review_confidence
stringclasses
13 values
text
stringlengths
402
130k
gpsfGAOUs58
MIDL.io/2023/Short_Paper_Track
2023
Caption generation from histopathology whole-slide images using pre-trained transformers
["Bryan Cardenas Guevara", "Niccol\u00f2 Marini", "Stefano Marchesin", "Witali Aswolinskiy", "Robert-Jan Schlimbach", "Damian Podareanu", "Francesco Ciompi"]
The recent advent of foundation models and large language models has enabled scientists to leverage large-scale knowledge of pretrained (vision) transformers and efficiently tailor it to downstream tasks. This technology can potentially automate multiple aspects of cancer diagnosis in digital pathology, from whole-slid...
["Whole slide images", "histopathology", "multi-modal training", "caption generation", "transformers"]
Medical Imaging with Deep Learning – Under Review 2023 Short Paper – MIDL 2023 submissionCaption generation from histopathology whole-slide imagesusing pre-trained transformersBryan Cardenas Guevara1bryan.cardenasguevara@surf.nl1SURF, Amsterdam, The NetherlandsNiccol` o Marini2niccolo.marini@hevs.ch2University of Appli...
bY3uqo26nwt
Review
6: Marginally above acceptance threshold
# Summary The paper applies a transformer-based caption model to generate pathology reports of whole slide images. The generated capt ions are then used to (manually) predict a diagnosis, which is comparable in performance to a fully supervised classi fication model. # Strengths I believe this is interesting work, a...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Caption generation from histopathology whole-slide images using pre-trained transformers ### Paper Abstract The recent advent of foundation models and large language mod...
HeEzgm-f4g1
ICLR.cc/2021/Conference
2021
On Batch-size Selection for Stochastic Training for Graph Neural Networks
["Yaochen Hu", "Amit Levi", "Ishaan Kumar", "Yingxue Zhang", "Mark Coates"]
Batch size is an important hyper-parameter for training deep learning models with stochastic gradient decent (SGD) method, and it has great influence on the training time and model performance. We study the batch size selection problem for training graph neural network (GNN) with SGD method. To reduce the training tim...
["gradients", "gnn", "selection", "stochastic training", "sgd", "training time", "metric", "graph neural networks", "important"]
ABSTRACTBatch size is an important hyper-parameter for training deep learning models withstochastic gradient decent (SGD) method, and it has great influence on the trainingtime and model performance. We study the batch size selection problem for train-ing graph neural network (GNN) with SGD method. To reduce the traini...
s7USH2rCCP
Official Blind Review #4
5: Marginally below acceptance threshold
In this paper, the authors study an important problem on how the choice of batch size (i.e., the number of sampled nodes) affects the training efficiency and accuracy of graph neural networks (GNN). Focusing on the layer-wise and graph-wise sampling for training, the authors theoretically characterize the impact of bat...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title On Batch-size Selection for Stochastic Training for Graph Neural Networks ### Paper Abstract Batch size is an important hyper-parameter for training deep learning models...
NsMLjcFaO8O
ICLR.cc/2021/Conference
2021
WaveGrad: Estimating Gradients for Waveform Generation
["Nanxin Chen", "Yu Zhang", "Heiga Zen", "Ron J Weiss", "Mohammad Norouzi", "William Chan"]
This paper introduces WaveGrad, a conditional model for waveform generation which estimates gradients of the data density. The model is built on prior work on score matching and diffusion probabilistic models. It starts from a Gaussian white noise signal and iteratively refines the signal via a gradient-based sampler c...
["vocoder", "diffusion", "score matching", "text-to-speech", "gradient estimation", "waveform generation"]
ABSTRACTThis paper introduces WaveGrad , a conditional model for waveform generationwhich estimates gradients of the data density. The model is built on prior workon score matching and diffusion probabilistic models. It starts from a Gaussianwhite noise signal and iteratively refines the signal via a gradient-based sam...
FzxTmyUQtPd
Good paper, incline to accept
7: Good paper, accept
Motivated by recent works on score matching and diffusion probabilistic models, the authors presented a conditional, non-autoregressive model for waveform generation. The paper is well motivated and easy to follow. The model starts from Gaussian white noise signal and iteratively refines the signal via Langevin dynami...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title WaveGrad: Estimating Gradients for Waveform Generation ### Paper Abstract This paper introduces WaveGrad, a conditional model for waveform generation which estimates gra...
aKt7FHPQxVV
ICLR.cc/2021/Conference
2021
Efficient Differentiable Neural Architecture Search with Model Parallelism
["Yi-Wei Chen", "Qingquan Song", "Xia Hu"]
Neural architecture search (NAS) automatically designs effective network architectures. Differentiable NAS with supernets that encompass all potential architectures in a large graph cuts down search overhead to few GPU days or less. However, these algorithms consume massive GPU memory, which will restrain NAS from larg...
["Neural Architecture Search", "Model Parallel"]
ABSTRACTNeural architecture search (NAS) automatically designs effective network archi-tectures. Differentiable NAS with supernets that encompass all potential archi-tectures in a large graph cuts down search overhead to few GPU days or less.However, these algorithms consume massive GPU memory, which will restrainNAS f...
ydF9NDBTKa
I believe the subject is important and I like the idea, but more experiments to demonstrate the effectiveness of the proposed approaches are needed.
5: Marginally below acceptance threshold
############################################################## Summary: This paper provides the interesting method that leverages GPU memory resources more efficiently for supernet (meta-graph) of differentiable NAS. For this, this paper proposes binary neural architecture search and consecutive model parallel (CMP)....
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Efficient Differentiable Neural Architecture Search with Model Parallelism ### Paper Abstract Neural architecture search (NAS) automatically designs effective network ar...
HklmoRVYvr
ICLR.cc/2020/Conference
2020
Long History Short-Term Memory for Long-Term Video Prediction
["Wonmin Byeon", "Jan Kautz"]
While video prediction approaches have advanced considerably in recent years, learning to predict long-term future is challenging — ambiguous future or error propagation over time yield blurry predictions. To address this challenge, existing algorithms rely on extra supervision (e.g., action or object pose), motion flo...
["LSTM", "video", "long-term prediction"]
ABSTRACTWhile video prediction approaches have advanced considerably in recent years,learning to predict long-term future is challenging — ambiguous future or errorpropagation over time yield blurry predictions. To address this challenge, existingalgorithms rely on extra supervision (e.g., action or object pose), motio...
r1xjMAW0tH
Official Blind Review #1
3: Weak Reject
Summary: This paper proposes a new LSTM architecture called LH-STM (and Double LH-STM). The main idea deals with having a history selection mechanism to directly extract what information from the past. The authors also propose to decompose the history and update in LH-STM into two networks called Double LH-STM. In expe...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Long History Short-Term Memory for Long-Term Video Prediction ### Paper Abstract While video prediction approaches have advanced considerably in recent years, learning t...
BygKZkBtDH
ICLR.cc/2020/Conference
2020
Balancing Cost and Benefit with Tied-Multi Transformers
["Raj Dabre", "Raphael Rubino", "Atsushi Fujita"]
This paper proposes a novel procedure for training multiple Transformers with tied parameters which compresses multiple models into one enabling the dynamic choice of the number of encoder and decoder layers during decoding. In sequence-to-sequence modeling, typically, the output of the last layer of the N-layer encode...
["tied models", "encoder-decoder", "multi-layer softmaxing", "depth prediction", "model compression"]
ABSTRACTThis paper proposes a novel procedure for training multiple Transformers withtied parameters which compresses multiple models into one enabling the dynamicchoice of the number of encoder and decoder layers during decoding. In sequence-to-sequence modeling, typically, the output of the last layer of the N-layer ...
B1liTs2TYS
Official Blind Review #3
6: Weak Accept
This paper proposes a novel procedure for training multiple Transformers with tied parameters which compresses multiple models into one enabling the dynamic choice of the number of encoder and decoder layers during decoding. The idea is simple and reasonable and the results are promising. I have several questions abo...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Balancing Cost and Benefit with Tied-Multi Transformers ### Paper Abstract This paper proposes a novel procedure for training multiple Transformers with tied parameters ...
K3qa-sMHpQX
ICLR.cc/2021/Conference
2021
ForceNet: A Graph Neural Network for Large-Scale Quantum Chemistry Simulation
["Weihua Hu", "Muhammed Shuaibi", "Abhishek Das", "Siddharth Goyal", "Anuroop Sriram", "Jure Leskovec", "Devi Parikh", "Larry Zitnick"]
Machine Learning (ML) has a potential to dramatically accelerate large-scale physics-based simulations. However, practical models for real large-scale and complex problems remain out of reach. Here we present ForceNet, a model for accurate and fast quantum chemistry simulations to accelerate catalyst discovery for rene...
["Graph Neural Networks", "Physical simulation", "Quantum chemistry", "Catalysis"]
ABSTRACTMachine Learning (ML) has a potential to dramatically accelerate large-scalephysics-based simulations. However, practical models for real large-scale and com-plex problems remain out of reach. Here we present ForceNet, a model for accurateand fast quantum chemistry simulations to accelerate catalyst discovery f...
c3-FpyPD-Lu
Relatively new application domain + nice results + ok technical contribution.
7: Good paper, accept
This paper presents ForceNet, a graph neural network, for estimating per-atom forces based on the 3D molecular structure. The authors argue that force-centric learning could be better in terms of prediction accuracy compared to energy-centric learning. As such, they focused on force-centric simulation and did several s...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title ForceNet: A Graph Neural Network for Large-Scale Quantum Chemistry Simulation ### Paper Abstract Machine Learning (ML) has a potential to dramatically accelerate large-s...
rJxbJeHFPS
ICLR.cc/2020/Conference
2020
What Can Neural Networks Reason About?
["Keyulu Xu", "Jingling Li", "Mozhi Zhang", "Simon S. Du", "Ken-ichi Kawarabayashi", "Stefanie Jegelka"]
Neural networks have succeeded in many reasoning tasks. Empirically, these tasks require specialized network structures, e.g., Graph Neural Networks (GNNs) perform well on many such tasks, but less structured networks fail. Theoretically, there is limited understanding of why and when a network structure generalizes be...
["reasoning", "deep learning theory", "algorithmic alignment", "graph neural networks"]
ABSTRACTNeural networks have succeeded in many reasoning tasks. Empirically, thesetasks require specialized network structures, e.g., Graph Neural Networks (GNNs)perform well on many such tasks, but less structured networks fail. Theoretically,there is limited understanding of why and when a network structure generaliz...
rJeOFDKjFr
Official Blind Review #1
6: Weak Accept
This work seeks theoretical and empirical proof of the reasoning capacity of neural networks. The authors build on a body of research that demonstrates the usefulness of different neural network architectures for different reasoning problems. For example, Deep Sets have been proposed to answer questions about sets (e.g...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title What Can Neural Networks Reason About? ### Paper Abstract Neural networks have succeeded in many reasoning tasks. Empirically, these tasks require specialized network st...
MG8Zde0ip6u
ICLR.cc/2021/Conference
2021
A Siamese Neural Network for Behavioral Biometrics Authentication
["Jes\u00fas Solano", "Esteban Rivera", "Alejandra Castelblanco", "Lizzy Tengana", "Christian Lopez", "Martin Ochoa"]
The raise in popularity of personalized web and mobile applications brings about a need of robust authentication systems. Although password authentication is the most popular authentication mechanism, it has also several drawbacks. Behavioral Biometrics Authentication has emerged as a complementary risk-based authentic...
["Deep Learning", "Few-shot Learning", "Behavioral Biometrics", "Biometric Authentication"]
ABSTRACTThe raise in popularity of web and mobile applications brings about a need ofrobust authentication systems. Although password authentication is the most pop-ular authentication mechanism, it has several drawbacks. Behavioral Biomet-rics Authentication has emerged as a complementary risk-based authenticationappr...
0vTcET67eH
Straightforward approach but limited ML contribution
4: Ok but not good enough - rejection
Summary: The paper proposes a siamese neural network to perform few-shot verification of user behaviour. ########################################################################## Reasons for score: Overall, I vote for rejection. I like the idea of applying the siamese paradigm to authentication however the ...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title A Siamese Neural Network for Behavioral Biometrics Authentication ### Paper Abstract The raise in popularity of personalized web and mobile applications brings about a n...
cT0jK5VvFuS
ICLR.cc/2021/Conference
2021
Uncertainty in Neural Processes
["Saeid Naderiparizi", "Kenny Chiu", "Benjamin Bloem-Reddy", "Frank Wood"]
We explore the effects of architecture and training objective choice on amortized posterior predictive inference in probabilistic conditional generative models. We aim this work to be a counterpoint to a recent trend in the literature that stresses achieving good samples when the amount of conditioning data is large. ...
["Neural Processes", "Meta-learning", "Variational Inference"]
ABSTRACTWe explore the effects of architecture and training objective choice on amortizedposterior predictive inference in probabilistic conditional generative models. Weaim this work to be a counterpoint to a recent trend in the literature that stressesachieving good samples when the amount of conditioning data is lar...
57Le8gn2iQj
contributions not clear
5: Marginally below acceptance threshold
Summary: This paper is an empirical investigation into the role of architecture and objective choices in Neural Process (NP) models when the amount of conditioning data is limited. Specifically, they investigate the question of well-calibrated uncertainty. Clarity: The overall quality of the writing is clear, but mis...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Uncertainty in Neural Processes ### Paper Abstract We explore the effects of architecture and training objective choice on amortized posterior predictive inference in pr...
Hyxsl2AqKm
ICLR.cc/2019/Conference
2019
ON THE EFFECTIVENESS OF TASK GRANULARITY FOR TRANSFER LEARNING
["Farzaneh Mahdisoltani", "Guillaume Berger", "Waseem Gharbieh", "David Fleet", "Roland Memisevic"]
We describe a DNN for video classification and captioning, trained end-to-end, with shared features, to solve tasks at different levels of granularity, exploring the link between granularity in a source task and the quality of learned features for transfer learning. For solving the new task domain in transfer learning,...
["Transfer Learning", "Video Understanding", "Fine-grained Video Classification", "Video Captioning", "Common Sense", "Something-Something Dataset."]
ABSTRACTWe describe a DNN for video classification and captioning, trained end-to-end,with shared features, to solve tasks at different levels of granularity, exploring thelink between granularity in a source task and the quality of learned features fortransfer learning. For solving the new task domain in transfer lear...
B1xz8jHq2X
Interesting observation and dataset, but more analysis would be necessary
5: Marginally below acceptance threshold
Summary This paper studied video classification and video caption generation problem. Especially, the paper tried few baseline architectures datasets used from pretraining features on recently proposed Something-something-V2 dataset and another newly proposed dataset. This paper argues that fine-grained annotation help...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title ON THE EFFECTIVENESS OF TASK GRANULARITY FOR TRANSFER LEARNING ### Paper Abstract We describe a DNN for video classification and captioning, trained end-to-end, with sha...
Bke6vTVYwH
ICLR.cc/2020/Conference
2020
Graph convolutional networks for learning with few clean and many noisy labels
["Ahmet Iscen", "Giorgos Tolias", "Yannis Avrithis", "Ondrej Chum", "Cordelia Schmid"]
In this work we consider the problem of learning a classifier from noisy labels when a few clean labeled examples are given. The structure of clean and noisy data is modeled by a graph per class and Graph Convolutional Networks (GCN) are used to predict class relevance of noisy examples. For each class, the GCN is trea...
["clean", "graph convolutional networks", "many noisy labels", "classifier", "noisy data", "class", "gcn", "noisy examples", "clean examples", "convolutional networks"]
ABSTRACTIn this work we consider the problem of learning a classifier from noisy labels whena few clean labeled examples are given. The structure of clean and noisy data ismodeled by a graph per class and Graph Convolutional Networks (GCN) are usedto predict class relevance of noisy examples. For each class, the GCN is...
rkgIzF3Zcr
Official Blind Review #1
6: Weak Accept
This paper presents a classification method when the data consists of few clean labels and many noisy labels. The authors propose to construct a graph structure within each class and use graph convolutional network to determine the clean/noisy labels of samples in each class. The model is based on a binary cross entrop...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Graph convolutional networks for learning with few clean and many noisy labels ### Paper Abstract In this work we consider the problem of learning a classifier from nois...
rkegcC4YvS
ICLR.cc/2020/Conference
2020
Removing the Representation Error of GAN Image Priors Using the Deep Decoder
["Max Daniels", "Reinhard Heckel", "Paul Hand"]
Generative models, such as GANs, have demonstrated impressive performance as natural image priors for solving inverse problems such as image restoration and compressive sensing. Despite this performance, they can exhibit substantial representation error for both in-distribution and out-of-distribution images, because t...
["deep decoder", "deep image prior", "GAN", "inverse problems"]
ABSTRACTGenerative models, such as GANs, learn an explicit low-dimensional represen-tation of a particular class of images, and so they may be used as natural imagepriors for solving inverse problems such as image restoration and compressivesensing. GAN priors have demonstrated impressive performance on these tasks,but...
BJgwdLAaKr
Official Blind Review #1
1: Reject
[Update after rebuttal period] I am sorry that the response cannot address my confusion. I still doubt the motivation of this paper and the actual experimental performance compared with state-of-the-art methods are still ignored. Thus I decrease my score. [Original reviews] This paper proposed to modeling image as the...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Removing the Representation Error of GAN Image Priors Using the Deep Decoder ### Paper Abstract Generative models, such as GANs, have demonstrated impressive performance...
1sPcfSScGWO
logconference.io/LOG/2022/Conference
2022
Graph Reinforcement Learning for Network Control via Bi-Level Optimization
["Daniele Gammelli", "James Harrison", "Kaidi Yang", "Marco Pavone", "Filipe Rodrigues", "Francisco C. Pereira"]
Dynamic network flow models have been extensively studied and widely used in the past decades to formulate many problems with great real-world impact, such as transportation, supply chain management, power grid control, and more. Within this context, time-expansion techniques currently represent a generic approach for ...
["Graph reinforcement learning", "Network control", "Bi-level optimization"]
Graph Reinforcement Learning for Network Controlvia Bi-Level OptimizationDaniele GammelliStanford Universitygammelli@stanford.eduJames HarrisonGoogle Research, Brain Teamjamesharrison@google.comKaidi YangNational University of Singaporekaidi.yang@nus.edu.sgMarco PavoneStanford Universitypavone@stanford.eduFilipe Rodrig...
7Fnw-0L1eyd
The paper formulates solving a dynamic network flow problem as a reinforcement learning problem where GNNs are used to represent both the policies and the value functions. In this formulation, the (infinite) action space is the joint set of flows and exchange weights. Instead of sampling from this space, authors propo...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Graph Reinforcement Learning for Network Control via Bi-Level Optimization ### Paper Abstract Dynamic network flow models have been extensively studied and widely used i...
HCa8gC_COVk
ICLR.cc/2021/Conference
2021
Mutual Calibration between Explicit and Implicit Deep Generative Models
["Qitian Wu", "Rui Gao", "Hongyuan Zha"]
Deep generative models are generally categorized into explicit models and implicit models. The former defines an explicit density form that allows likelihood inference; while the latter targets a flexible transformation from random noise to generated samples. To take full advantages of both models, we propose Stein Bri...
["deep generative models", "generative adversarial networks", "density estimation"]
ABSTRACTDeep generative models are generally categorized into explicit models and im-plicit models. The former defines an explicit density form that allows likelihoodinference; while the latter targets a flexible transformation from random noise togenerated samples. To take full advantages of both models, we propose St...
c_W11bDa70T
Very similar to a NeurIPS-19 paper
3: Clear rejection
This paper adopted Stein's method to connect an explicit density estimator and an implicit sample generator to propose an objective function for deep generative learning. This paper is written well and the organization is very clear. However, this paper is very similar to a NeurIPS-19 paper (Two Generator Game: Learnin...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Mutual Calibration between Explicit and Implicit Deep Generative Models ### Paper Abstract Deep generative models are generally categorized into explicit models and impl...
r1eVX0EFvH
ICLR.cc/2020/Conference
2020
Exploiting Excessive Invariance caused by Norm-Bounded Adversarial Robustness
["J\u00f6rn-Henrik Jacobsen", "Jens Behrmann", "Nicholas Carlini", "Florian Tram\u00e8r", "Nicolas Papernot"]
Adversarial examples are malicious inputs crafted to cause a model to misclassify them. In their most common instantiation, "perturbation-based" adversarial examples introduce changes to the input that leave its true label unchanged, yet result in a different model prediction. Conversely, "invariance-based" adversari...
["Invariance", "Robustness", "Adversarial Examples"]
ABSTRACTAdversarial examples are malicious inputs crafted to cause a model to misclassifythem. In their most common instantiation, “perturbation-based” adversarial exam-ples introduce changes to the input that leave its true label unchanged, yet result ina different model prediction. Conversely, “invariance-based” adve...
rJlJAnX3YB
Official Blind Review #2
6: Weak Accept
Thanks for your response. It has addressed most of my concerns. I'd like to increase my score. ----------------- This paper examines the perturbation-based and invariance-based robustness of deep neural networks. They found that models trained to be robust under l_p threat model are more vulnerable to invariance-base...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Exploiting Excessive Invariance caused by Norm-Bounded Adversarial Robustness ### Paper Abstract Adversarial examples are malicious inputs crafted to cause a model to mi...
SJlzvTNtvS
ICLR.cc/2020/Conference
2020
Better Optimization for Neural Architecture Search with Mixed-Level Reformulation
["Chaoyang He", "Haishan Ye", "Tong Zhang"]
Many recently proposed methods for Neural Architecture Search (NAS) can be formulated as bilevel optimization. For efficient implementation, its solution requires approximations of second-order methods. In this paper, we demonstrate that gradient errors caused by such approximation leads to suboptimality, in that such ...
["neural architecture search", "reformulation", "methods", "bilevel optimization", "milenas", "better optimization", "optimization", "reformulation many", "nas", "efficient implementation"]
ABSTRACTMany recently proposed methods for Neural Architecture Search (NAS) can beformulated as bilevel optimization. For efficient implementation, its solution re-quires approximations of second-order methods. In this paper, we demonstratethat gradient errors caused by such approximation leads to suboptimality, in tha...
BJezuiV0qS
Official Blind Review #4
1: Reject
This paper proposes a new objective for one-shot neural architecture search (NAS) using gradient-based methods. The authors hypothesize that the finite differences approximation in the second-order approximation in DARTS [1] directs the gradients towards bad local minimas in the architecture space. To this end, the aut...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Better Optimization for Neural Architecture Search with Mixed-Level Reformulation ### Paper Abstract Many recently proposed methods for Neural Architecture Search (NAS) ...
a5KvtsZ14ev
ICLR.cc/2021/Conference
2021
SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks
["Bahare Fatemi", "Seyed Mehran Kazemi", "Layla El Asri"]
Graph neural networks (GNNs) work well when the graph structure is provided. However, this structure may not always be available in real-world applications. One solution to this problem is to infer the latent structure and then apply a GNN to the inferred graph. Unfortunately, the space of possible graph structures gro...
["Graph Neural Networks", "Graph Representation Learning", "Graph Structure Learning", "Self-supervision"]
ABSTRACTGraph neural networks (GNNs) work well when the graph structure is provided.However, this structure may not always be available in real-world applications.One solution to this problem is to infer the latent structure and then apply a GNNto the inferred graph. Unfortunately, the space of possible graph structure...
KX2uDF2HrLn
Interesting work with convincing results.
7: Good paper, accept
The paper proposes a way to use self-supervision with denoising autoencoders to improve learning of the graph structure for GNNs. The approach is compared with a number of recent approaches from the literature. The approach addresses the highly relevant problem of learning graph structure with GNNs. The paper is well ...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks ### Paper Abstract Graph neural networks (GNNs) work well when the graph structure is provi...
Skf-oo0qt7
ICLR.cc/2019/Conference
2019
On Generalization Bounds of a Family of Recurrent Neural Networks
["Minshuo Chen", "Xingguo Li", "Tuo Zhao"]
Recurrent Neural Networks (RNNs) have been widely applied to sequential data analysis. Due to their complicated modeling structures, however, the theory behind is still largely missing. To connect theory and practice, we study the generalization properties of vanilla RNNs as well as their variants, including Minimal Ga...
["Recurrent Neural Networks", "MGU", "LSTM", "Generalization Bound", "PAC-Learning"]
ABSTRACTRecurrent Neural Networks (RNNs) have been widely applied to sequential dataanalysis. Due to their complicated modeling structures, however, the theory be-hind is still largely missing. To connect theory and practice, we study the gener-alization properties of vanilla RNNs as well as their variants, including M...
SJgzW2Vqh7
Marginal improvement over existing bounds; Incomplete comparison with existing works
4: Ok but not good enough - rejection
This paper studies the generalization properties of Recurrent Neural Networks (RNN) and its variants for the sequence to sequence multiclass prediction problem. The problem is important to understand in the theoretical machine learning community. The paper is written well overall, clearly explaining the results obtaine...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title On Generalization Bounds of a Family of Recurrent Neural Networks ### Paper Abstract Recurrent Neural Networks (RNNs) have been widely applied to sequential data analysi...
BkVf1AeAZ
ICLR.cc/2018/Conference
2018
Label Embedding Network: Learning Label Representation for Soft Training of Deep Networks
["Xu Sun", "Bingzhen Wei", "Xuancheng Ren", "Shuming Ma"]
We propose a method, called Label Embedding Network, which can learn label representation (label embedding) during the training process of deep networks. With the proposed method, the label embedding is adaptively and automatically learned through back propagation. The original one-hot represented loss function is conv...
["label embedding", "deep learning", "label representation", "computer vision", "natural language processing"]
ABSTRACTWe propose a method, called Label Embedding Network, which can learn labelrepresentation (label embedding) during the training process of deep networks.With the proposed method, the label embedding is adaptively and automaticallylearned through back propagation. The original one-hot represented loss func-tion i...
r1zEZ9ief
Technique not properly justified; not enough insights can be learned from the work.
3: Clear rejection
The paper proposes a method which jointly learns the label embedding (in the form of class similarity) and a classification model. While the motivation of the paper makes sense, the model is not properly justified, and I learned very little after reading the paper. There are 5 terms in the proposed objective function....
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Label Embedding Network: Learning Label Representation for Soft Training of Deep Networks ### Paper Abstract We propose a method, called Label Embedding Network, which c...
Byaj2S98f
ICLR.cc/2018/Workshop
2018
Neural Network Quine
["Oscar Chang", "Hod Lipson"]
We describe a method for building a self-replicating neural network. We also test it on an image classification network for MNIST.
["self-replication", "quine"]
ABSTRACTWe describe a method for building a self-replicating neural network. We also testit on an image classification network for MNIST.1 S ELF-REPLICATIONJohn von Neumann pioneered the concept of an artificial self-replicating machine before the discov-ery of DNA. He described a set of initial states and transformati...
SJbqcbxKG
Interesting idea but no real results and not developed enough theoretically
2: Strong rejection
The paper proposes to train neural networks to reproduce their own parameter vector. While I find the motivations based on biological systems a little bit vague, the idea to train a neural network to reproduce its own weights is certainly original. Unfortunately it's not that clear to me why that would be a useful th...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Neural Network Quine ### Paper Abstract We describe a method for building a self-replicating neural network. We also test it on an image classification network for MNIST...
BklhAj09K7
ICLR.cc/2019/Conference
2019
Unsupervised Domain Adaptation for Distance Metric Learning
["Kihyuk Sohn", "Wenling Shang", "Xiang Yu", "Manmohan Chandraker"]
Unsupervised domain adaptation is a promising avenue to enhance the performance of deep neural networks on a target domain, using labels only from a source domain. However, the two predominant methods, domain discrepancy reduction learning and semi-supervised learning, are not readily applicable when source and target ...
["domain adaptation", "distance metric learning", "face recognition"]
ABSTRACTUnsupervised domain adaptation is a promising avenue to enhance the performanceof deep neural networks on a target domain, using labels only from a sourcedomain. However, the two predominant methods, domain discrepancy reductionlearning and semi-supervised learning, are not readily applicable when source andtar...
ByemonjVnm
Interesting paper addressing a difficult problem. Good formalization and reasonable evaluation
8: Top 50% of accepted papers, clear accept
I like the idea of the paper and I believe it addressing a very relevant problem. While the authors provide a good formalization of the problem and convincing demonstration of the generalization bound, the evaluation could have been better by including some more challenging experiments to really prove the point of the ...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Unsupervised Domain Adaptation for Distance Metric Learning ### Paper Abstract Unsupervised domain adaptation is a promising avenue to enhance the performance of deep ne...
67q9f8gChCF
ICLR.cc/2021/Conference
2021
Learning Efficient Planning-based Rewards for Imitation Learning
["Xingrui Yu", "Yueming Lyu", "Ivor Tsang"]
Imitation learning from limited demonstrations is challenging. Most inverse reinforcement learning (IRL) methods are unable to perform as good as the demonstrator, especially in a high-dimensional environment, e.g, the Atari domain. To address this challenge, we propose a novel reward learning method, which streamlines...
["imitation", "efficient", "rewards", "limited demonstrations", "inverse reinforcement learning", "irl", "methods", "unable", "good", "demonstrator"]
ABSTRACTImitation learning from limited demonstrations is challenging. Most inverse rein-forcement learning (IRL) methods are unable to perform as good as the demon-strator, especially in a high-dimensional environment, e.g, the Atari domain. Toaddress this challenge, we propose a novel reward learning method, which st...
OXKjo6y8msG
Learning a policy from one demonstration without rewards
5: Marginally below acceptance threshold
This paper assumes no access to the reward values and attempts to learn a policy by starting just with one demonstration to define the reward. For obtaining the reward, the authors rely on the ideas from Value Iteration Networks (VIN) method and they add the modules that help to deal with cases with complex transition ...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Learning Efficient Planning-based Rewards for Imitation Learning ### Paper Abstract Imitation learning from limited demonstrations is challenging. Most inverse reinforce...
Bke61krFvS
ICLR.cc/2020/Conference
2020
Learning representations for binary-classification without backpropagation
["Mathias Lechner"]
The family of feedback alignment (FA) algorithms aims to provide a more biologically motivated alternative to backpropagation (BP), by substituting the computations that are unrealistic to be implemented in physical brains. While FA algorithms have been shown to work well in practice, there is a lack of rigorous theory...
["feedback alignment", "alternatives to backpropagation", "biologically motivated learning algorithms"]
ABSTRACTThe family of feedback alignment (FA) algorithms aims to provide a more bio-logically motivated alternative to backpropagation (BP), by substituting the com-putations that are unrealistic to be implemented in physical brains. While FAalgorithms have been shown to work well in practice, there is a lack of rigoro...
rJxE1TVAFH
Official Blind Review #2
8: Accept
This paper examines the question of learning in neural networks with random, fixed feedback weights, a technique known as “feedback alignment”. Feedback alignment was originally discovered by Lillicrap et al. (2016; Nature Communications, 7, 13276) when they were exploring potential means of solving the “weight transpo...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Learning representations for binary-classification without backpropagation ### Paper Abstract The family of feedback alignment (FA) algorithms aims to provide a more bio...
B1x3EgHtwB
ICLR.cc/2020/Conference
2020
ExpandNets: Linear Over-parameterization to Train Compact Convolutional Networks
["Shuxuan Guo", "Jose M. Alvarez", "Mathieu Salzmann"]
In this paper, we introduce a novel approach to training a given compact network. To this end, we build upon over-parameterization, which typically improves both optimization and generalization in neural network training, while being unnecessary at inference time. We propose to expand each linear layer of the compact n...
["Compact Network Training", "Linear Expansion", "Over-parameterization", "Knowledge Transfer"]
ABSTRACTIn this paper, we introduce a novel approach to training a given compact network.To this end, we build upon over-parameterization, which typically improves bothoptimization and generalization in neural network training, while being unneces-sary at inference time. We propose to expand each linear layer of the co...
SyxThYwDYS
Official Blind Review #3
3: Weak Reject
This paper is extremely interesting and quite surprising. In fact, the major claim is that using a cascade of linear layers instead of a single layer can lead to better performance in deep neural networks. As the title reports, expanding layers seems to be the key to obtain extremely interesting results. Moreover, the ...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title ExpandNets: Linear Over-parameterization to Train Compact Convolutional Networks ### Paper Abstract In this paper, we introduce a novel approach to training a given comp...
HklRKpEKDr
ICLR.cc/2020/Conference
2020
Deep Coordination Graphs
["Wendelin Boehmer", "Vitaly Kurin", "Shimon Whiteson"]
This paper introduces the deep coordination graph (DCG) for collaborative multi-agent reinforcement learning. DCG strikes a flexible trade-off between representational capacity and generalization by factorizing the joint value function of all agents according to a coordination graph into payoffs between pairs of agents...
["multi-agent reinforcement learning", "coordination graph", "deep Q-learning", "value factorization", "relative overgeneralization"]
ABSTRACTThis paper introduces the deep coordination graph (DCG) for collaborative multi-agent reinforcement learning. DCG strikes a flexible trade-off between represen-tational capacity and generalization by factorizing the joint value function of allagents according to a coordination graph into payoffs between pairs o...
HkxPtjf2FS
Official Blind Review #1
6: Weak Accept
This paper introduces the deep coordination graph for collaborative multi-agent reinforcement learning aimed to solve predator-prey tasks by preventing relative overgeneralization during the exploration of agents. In general, this paper gives a detailed and comprehensible depiction of the Introduction, Related work a...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Deep Coordination Graphs ### Paper Abstract This paper introduces the deep coordination graph (DCG) for collaborative multi-agent reinforcement learning. DCG strikes a f...
H1g2NhC5KQ
ICLR.cc/2019/Conference
2019
Multiple-Attribute Text Rewriting
["Guillaume Lample", "Sandeep Subramanian", "Eric Smith", "Ludovic Denoyer", "Marc'Aurelio Ranzato", "Y-Lan Boureau"]
The dominant approach to unsupervised "style transfer'' in text is based on the idea of learning a latent representation, which is independent of the attributes specifying its "style''. In this paper, we show that this condition is not necessary and is not always met in practice, even with domain adversarial training t...
["controllable text generation", "generative models", "conditional generative models", "style transfer"]
ABSTRACTThe dominant approach to unsupervised “style transfer” in text is based on the ideaof learning a latent representation, which is independent of the attributes specify-ing its “style”. In this paper, we show that this condition is not necessary and isnot always met in practice, even with domain adversarial train...
H1lXb9okiX
Impressive experiments, but hard to determine how much is methodologically new here
6: Marginally above acceptance threshold
The paper proposes "style transfer" approaches for text rewriting that allow for controllable attributes. For example, given one piece of text (and the conditional attributes associated with the user who generated it, such as their age and gender), these attributes can be changed so as to generate equivalent text in a ...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Multiple-Attribute Text Rewriting ### Paper Abstract The dominant approach to unsupervised "style transfer'' in text is based on the idea of learning a latent representa...
S4nlj2Vdw-9
ICLR.cc/2022/Workshop/DGM4HSD
2022
Register Allocation using Graph Neural Networks
["Anonymous"]
The execution lifecycle of a program goes through various optimizations and code generation. Register allocation is critical part of optimization phase of compiler. Graph coloring is one of the NP hard problem and it applies to various applications such as register allocation. Deep Learning methods has various state-of...
["Register allocation", "graph coloring", "graph neural networks"]
ABSTRACTThe execution lifecycle of a program goes through various optimizations and codegeneration. Register allocation is critical part of optimization phase of compiler.Graph coloring is one of the NP hard problem and it applies to various applicationssuch as register allocation. Deep Learning methods has various sta...
BRzeNr7BkX9
Somehow incomplete
3: Clear rejection
This paper uses GNNs to solve register allocation, which is equivalent to a graph coloring problem. As I understand, (in practice) the authors solve the decision version of the problem. The GNN problem boils down to graph classification, taking as input a bipartite graph of variables and registers. The method's descri...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Register Allocation using Graph Neural Networks ### Paper Abstract The execution lifecycle of a program goes through various optimizations and code generation. Register ...
snOgiCYZgJ7
ICLR.cc/2021/Conference
2021
Neural representation and generation for RNA secondary structures
["Zichao Yan", "William L. Hamilton", "Mathieu Blanchette"]
Our work is concerned with the generation and targeted design of RNA, a type of genetic macromolecule that can adopt complex structures which influence their cellular activities and functions. The design of large scale and complex biological structures spurs dedicated graph-based deep generative modeling techniques, wh...
["Graph neural network", "Deep generative modeling", "Machine learning", "Drug discovery", "RNA structure", "RNA structure embedding", "RNA-protein interaction prediction"]
ABSTRACTOur work is concerned with the generation and targeted design of RNA, a type ofgenetic macromolecule that can adopt complex structures which influence theircellular activities and functions. The design of large scale and complex biologicalstructures spurs dedicated graph-based deep generative modeling technique...
u1mN0RVn0oV
Evaluating generative models for RNA secondary structure is a challenging problem
6: Marginally above acceptance threshold
Summary: This paper proposes 3 deep generative models based on VAEs (with different encoding schemes for RNA secondary structure) for the generation of RNA secondary structures. They test each model on 3 benchmark tasks: unsupervised generation, semi-supervised learning and targeted generation. This paper has many int...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Neural representation and generation for RNA secondary structures ### Paper Abstract Our work is concerned with the generation and targeted design of RNA, a type of gene...
HyRVBzap-
ICLR.cc/2018/Conference
2018
Cascade Adversarial Machine Learning Regularized with a Unified Embedding
["Taesik Na", "Jong Hwan Ko", "Saibal Mukhopadhyay"]
Injecting adversarial examples during training, known as adversarial training, can improve robustness against one-step attacks, but not for unknown iterative attacks. To address this challenge, we first show iteratively generated adversarial images easily transfer between networks trained with the same strategy. Inspir...
["adversarial machine learning", "embedding", "regularization", "adversarial attack"]
ABSTRACTInjecting adversarial examples during training, known as adversarial training, canimprove robustness against one-step attacks, but not for unknown iterative at-tacks. To address this challenge, we first show iteratively generated adversarialimages easily transfer between networks trained with the same strategy ...
HyPfhQ2ez
Review
6: Marginally above acceptance threshold
This paper improves adversarial training by adding to its traditional objective a regularization term forcing a clean example and its adversarial version to be close in the embedding space. This is an interesting idea which, from a robustness point of view (Xu et al, 2013) makes sense. Note that a similar strategy has ...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Cascade Adversarial Machine Learning Regularized with a Unified Embedding ### Paper Abstract Injecting adversarial examples during training, known as adversarial trainin...
BkgVx3A9Km
ICLR.cc/2019/Conference
2019
A More Globally Accurate Dimensionality Reduction Method Using Triplets
["Ehsan Amid", "Manfred K. Warmuth"]
We first show that the commonly used dimensionality reduction (DR) methods such as t-SNE and LargeVis poorly capture the global structure of the data in the low dimensional embedding. We show this via a number of tests for the DR methods that can be easily applied by any practitioner to the dataset at hand. Surprisingl...
["Dimensionality Reduction", "Visualization", "Triplets", "t-SNE", "LargeVis"]
ABSTRACTWe first show that the commonly used dimensionality reduction (DR) methods suchas t-SNE and LargeVis poorly capture the global structure of the data in the lowdimensional embedding. We show this via a number of tests for the DR methodsthat can be easily applied by any practitioner to the dataset at hand. Surpri...
rJxzobq32m
Novel loss function but experiments are lacking
6: Marginally above acceptance threshold
Motivated by the observation that most of previous dimensionality reduction methods focus on preserving local pairwise neighboring probabilities and lack in preserving global properties, this paper proposes a method called TriMap to optimize a loss function preserving similarities among triplets of data points. A lar...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title A More Globally Accurate Dimensionality Reduction Method Using Triplets ### Paper Abstract We first show that the commonly used dimensionality reduction (DR) methods suc...
B1MbDj0ctQ
ICLR.cc/2019/Conference
2019
Switching Linear Dynamics for Variational Bayes Filtering
["Philip Becker-Ehmck", "Jan Peters", "Patrick van der Smagt"]
System identification of complex and nonlinear systems is a central problem for model predictive control and model-based reinforcement learning. Despite their complexity, such systems can often be approximated well by a set of linear dynamical systems if broken into appropriate subsequences. This mechanism not only hel...
["sequence model", "switching linear dynamical systems", "variational bayes", "filter", "variational inference", "stochastic recurrent neural network"]
ABSTRACTSystem identification of complex and nonlinear systems is a central problemfor model predictive control and model-based reinforcement learning. Despitetheir complexity, such systems can often be approximated well by a set of lineardynamical systems if broken into appropriate subsequences. This mechanism notonly...
S1eYvPS0om
Interesting paper showing how to use switching variables in deep probabilistic temporal models
7: Good paper, accept
This paper proposes a deep probabilistic model for temporal data that leverages latent variables to switch between different learned linear dynamics. The probability distributions are parameterized by deep neural networks and learning is performed end-to-end with amortized variational inference using inference networks...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Switching Linear Dynamics for Variational Bayes Filtering ### Paper Abstract System identification of complex and nonlinear systems is a central problem for model predic...
SkeYUkStPr
ICLR.cc/2020/Conference
2020
Deep Lifetime Clustering
["S Chandra Mouli", "Leonardo Teixeira", "Jennifer Neville", "Bruno Ribeiro"]
The goal of lifetime clustering is to develop an inductive model that maps subjects into $K$ clusters according to their underlying (unobserved) lifetime distribution. We introduce a neural-network based lifetime clustering model that can find cluster assignments by directly maximizing the divergence between the empiri...
["Lifetime Clustering", "Deep Learning", "Survival Distributions", "Kuiper two-sample test"]
ABSTRACTThe goal of lifetime clustering is to develop an inductive model that maps subjectsintoKclusters according to their underlying (unobserved) lifetime distribution.We introduce a neural-network based lifetime clustering model that can find clusterassignments by directly maximizing the divergence between the empir...
H1gI2HOCYB
Official Blind Review #2
6: Weak Accept
This paper proposes a method to cluster subjects based on the latent lifetime distribution. The proposed model clusters by maximizing the empirical divergence between different clusters. The problem setting of this paper is not clear to me. I am not sure which variables are observed and which are not. For example, i...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Deep Lifetime Clustering ### Paper Abstract The goal of lifetime clustering is to develop an inductive model that maps subjects into $K$ clusters according to their unde...
SylUzpNFDS
ICLR.cc/2020/Conference
2020
SoftLoc: Robust Temporal Localization under Label Misalignment
["Julien Schroeter", "Kirill Sidorov", "Dave Marshall"]
This work addresses the long-standing problem of robust event localization in the presence of temporally of misaligned labels in the training data. We propose a novel versatile loss function that generalizes a number of training regimes from standard fully-supervised cross-entropy to count-based weakly-supervised learn...
["deep learning", "temporal localization", "robustness", "label misalignment", "music", "time series"]
ABSTRACTThis work addresses the long-standing problem of robust event localization in thepresence of temporally misaligned labels in the training data. We propose a novelversatile loss function that generalizes a number of training regimes from standardfully-supervised cross-entropy to count-based weakly-supervised lea...
B1lBTZiAKS
Official Blind Review #2
6: Weak Accept
This paper deals with temporal label noise, or label misalignment problem in localization, which is the long-standing problem of robust event localization regarding the temporally misaligned labels in sequential data. Some existing works are constrained to hardly fit the annotations of data during training. The work m...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title SoftLoc: Robust Temporal Localization under Label Misalignment ### Paper Abstract This work addresses the long-standing problem of robust event localization in the prese...
HJePXkHtvS
ICLR.cc/2020/Conference
2020
Deep Generative Classifier for Out-of-distribution Sample Detection
["Dongha Lee", "Sehun Yu", "Hwanjo Yu"]
The capability of reliably detecting out-of-distribution samples is one of the key factors in deploying a good classifier, as the test distribution always does not match with the training distribution in most real-world applications. In this work, we propose a deep generative classifier which is effective to detect out...
["Out-of-distribution Detection", "Generative Classifier", "Deep Neural Networks", "Multi-class Classification", "Gaussian Discriminant Analysis"]
ABSTRACTThe capability of reliably detecting out-of-distribution samples is one of the keyfactors in deploying a good classifier, as the test distribution always does not matchwith the training distribution in most real-world applications. In this work, wepropose a deep generative classifier which is effective to detec...
SJgAzmZUYr
Official Blind Review #1
3: Weak Reject
Summary: Unlike the softmax classifier, the authors considered the generative classifier based on Gaussian discriminative analysis and showed that such deep generative classifiers can be useful for detecting out-of-distribution samples. For various benchmark tasks, the proposed method outperforms baselines based on th...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Deep Generative Classifier for Out-of-distribution Sample Detection ### Paper Abstract The capability of reliably detecting out-of-distribution samples is one of the key...
EUUp9nWXsop
ICLR.cc/2021/Conference
2021
IALE: Imitating Active Learner Ensembles
["Christoffer L\u00f6ffler", "Christopher Mutschler"]
Active learning (AL) prioritizes the labeling of the most informative data samples. However, the performance of AL heuristics depends on the structure of the underlying classifier model and the data. We propose an imitation learning scheme that imitates the selection of the best expert heuristic at each stage of the AL...
["active learning", "imitating learning", "ensembles"]
ABSTRACTActive learning (AL) prioritizes the labeling of the most informative data samples.However, the performance of AL heuristics depends on the structure of the under-lying classifier model and the data. We propose an imitation learning scheme thatimitates the selection of the best expert heuristic at each stage of...
NWi9-MRiKbp
The proposed framework makes sense, but exact contributions need to be clarified.
6: Marginally above acceptance threshold
The paper proposes a batch-mode active learning approach via imitation learning to derive a sample selection strategy for active learning. The proposed method makes use of a pool of existing heuristics, learning a sample selection strategy that rely on different heuristics in the pool depending on the different stage o...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title IALE: Imitating Active Learner Ensembles ### Paper Abstract Active learning (AL) prioritizes the labeling of the most informative data samples. However, the performance ...
ShGxRxFV6Mq
graphicsinterface.org/Graphics_Interface/2022/Conference
2022
Structured Shape-Patterns from a Sketch: A Multi-Scale Approach
["Pauline Olivier", "Pooran Memari", "Marie-Paule Cani"]
Structured 2D patterns formed by the anisotropic distribution of arbitrary shapes are ubiquitous in nature and man-made environments. They may include both bounded and unbounded (extended fiber-like) shapes. In this work, we address the problem of interactively generating such patterns from a single exemplar sketched b...
["Sketch-based modeling", "distribution synthesis", "perceptual user study", "expressive rendering."]
ABSTRACTStructured 2D patterns formed by the anisotropic distribution ofarbitrary shapes are ubiquitous in nature and man-made environ-ments. They may include both bounded and unbounded (extendedfiber-like) shapes. In this work, we address the problem of inter-actively generating such patterns from a single exemplar sk...
YlxJCi29OaF
Not much technical contribution, but otherwise good paper
6: Marginally above acceptance threshold
The paper presents a new vector example-based texture synthesis method. Having two types of input strokes, bounded and unbounded, they group them, fit straight fibers/fiber medians with their 'ribbons'. Ribbons are then extended while bending them slightly to avoid introducing new overlaps. Within each ribbon, strokes ...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Structured Shape-Patterns from a Sketch: A Multi-Scale Approach ### Paper Abstract Structured 2D patterns formed by the anisotropic distribution of arbitrary shapes are ...
BJgqqsAct7
ICLR.cc/2019/Conference
2019
Non-vacuous Generalization Bounds at the ImageNet Scale: a PAC-Bayesian Compression Approach
["Wenda Zhou", "Victor Veitch", "Morgane Austern", "Ryan P. Adams", "Peter Orbanz"]
Modern neural networks are highly overparameterized, with capacity to substantially overfit to training data. Nevertheless, these networks often generalize well in practice. It has also been observed that trained networks can often be ``compressed to much smaller representations. The purpose of this paper is to connect...
["generalization", "deep-learning", "pac-bayes"]
ABSTRACTModern neural networks are highly overparameterized, with capacity to substan-tially overfit to training data. Nevertheless, these networks often generalize well inpractice. It has also been observed that trained networks can often be “compressed”to much smaller representations. The purpose of this paper is to ...
SkxzLGFojm
Good paper provided the authors have answers to some technical questions.
8: Top 50% of accepted papers, clear accept
This paper gives the first nonvacuous generalization bounds for meaningful Imagenet models. These bounds are given in terms of the bit length of compressions of learned models together with a method for taking into account symmetries of the uncompressed parameters. These bounds are nonvacuous only when the compressed...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Non-vacuous Generalization Bounds at the ImageNet Scale: a PAC-Bayesian Compression Approach ### Paper Abstract Modern neural networks are highly overparameterized, with...
NqPW1ZJjXDJ
ICLR.cc/2021/Conference
2021
NASOA: Towards Faster Task-oriented Online Fine-tuning
["Hang Xu", "Ning Kang", "Gengwei Zhang", "Xiaodan Liang", "Zhenguo Li"]
Fine-tuning from pre-trained ImageNet models has been a simple, effective, and popular approach for various computer vision tasks. The common practice of fine-tuning is to adopt a default hyperparameter setting with a fixed pre-trained model, while both of them are not optimized for specific tasks and time constraints....
["Fine-tuning", "AutoML", "NAS"]
ABSTRACTFine-tuning from pre-trained ImageNet models has been a simple, effective, andpopular approach for various computer vision tasks. The common practice of fine-tuning is to adopt a default hyperparameter setting with a fixed pre-trained model,while both of them are not optimized for specific tasks and time constr...
YJjkmxoyHVv
Official Blind Review #4 - post rebuttal update
6: Marginally above acceptance threshold
SUMMARY: The paper presents a joint Neural Architecture Search (NAS) and Online Adaption (OA) framework named NASOA, aiming at providing faster task-oriented fine-tuning system. The first step of the approach is an offline NAS to form a pretrained training-efficient model zoo, which is followed by an online scheduler w...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title NASOA: Towards Faster Task-oriented Online Fine-tuning ### Paper Abstract Fine-tuning from pre-trained ImageNet models has been a simple, effective, and popular approach...
g1KmTQhOhag
ICLR.cc/2021/Conference
2021
Memory Representation in Transformer
["Mikhail Burtsev", "Yurii Kuratov", "Anton Peganov", "Grigory V. Sapunov"]
Transformer-based models have achieved state-of-the-art results in many natural language processing tasks. The self-attention architecture allows transformer to combine information from all elements of a sequence into context-aware representations. However, information about the context is stored mostly in the same ele...
["transformer", "memory augmented networks"]
ABSTRACTTransformer-based models have achieved state-of-the-art results in many naturallanguage processing tasks. The self-attention architecture allows transformer tocombine information from all elements of a sequence into context-aware repre-sentations. However, information about the context is stored mostly in the s...
_0S1DsO1_kU
Review from R5: Interesting exploration but the method and empirical studies are relatively weak
4: Ok but not good enough - rejection
> Summary: The paper proposes to study three formulations (MemTransformer, MemCtrl and MemBottleneck) of memory-augmented self-attention transformers, and investigate the influence of adding memory tokens to the model to its overall performance. The authors claim via some experiments on MT and LM that memory augmentati...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Memory Representation in Transformer ### Paper Abstract Transformer-based models have achieved state-of-the-art results in many natural language processing tasks. The se...
5vShUEyjmm
ICLR.cc/2021/Conference
2021
Dual Contradistinctive Generative Autoencoder
["Gaurav Parmar", "Dacheng Li", "Kwonjoon Lee", "Zhuowen Tu"]
We present a new generative autoencoder model with dual contradistinctive losses to improve generative autoencoder that performs simultaneous inference (reconstruction) and synthesis (generation). We name our model dual contradistinctive generative autoencoder (DC-VAE) that integrates an instance-level discriminative l...
["fidelity", "dual contradistinctive losses", "generative autoencoder", "simultaneous inference", "reconstruction", "synthesis", "generation"]
ABSTRACTWe present a new generative autoencoder model with dual contradistinctive lossesto improve generative autoencoder that performs simultaneous inference (recon-struction) and synthesis (generation). We name our model dual contradistinctivegenerative autoencoder (DC-V AE) that integrates an instance-level discrimi...
_Ns3sRK8EB2
Official Review for DC-VAE
3: Clear rejection
This paper proposed to replace the ELBO loss in VAE-GAN with an instance-level contrastive loss to improve the performance of image synthesis and reconstruction. Experiments on multiple benchmarks of images demonstrate the effectiveness of DC-VAE on image synthesis and reconstruction, and latent representation learning...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Dual Contradistinctive Generative Autoencoder ### Paper Abstract We present a new generative autoencoder model with dual contradistinctive losses to improve generative a...
U2adryQpew5
AKBC.ws/2021/Conference
2021
Modelling Monotonic and Non-Monotonic Attribute Dependencies with Embeddings: A Theoretical Analysis
["Steven Schockaert"]
During the last decade, entity embeddings have become ubiquitous in Artificial Intelligence. Such embeddings essentially serve as compact but semantically meaningful representations of the entities of interest. In most approaches, vectors are used for representing the entities themselves, as well as for representing th...
["Entity embeddings", "learning & reasoning", "theoretical limitations"]
Automated Knowledge Base Construction (2021) Conference paperModelling Monotonic and Non-Monotonic AttributeDependencies with Embeddings: A Theoretical AnalysisSteven Schockaert schockaerts1@cardif.ac.ukCardi University, UKAbstractDuring the last decade, entity embeddings have become ubiquitous in Articial Intel-ligenc...
p9j1FGA_ojY
Interesting Theoretical Analysis on Different Models.
7: Good paper, accept
**Paper summary**: This paper tackles the problem of modeling monotonic and non-monotonic reasoning rules in different embeddings spaces, mostly vectors, but under different score functions used by the vector embeddings. The paper first provides the definition of monotonic reasoning, then iterates through a few vector-...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Modelling Monotonic and Non-Monotonic Attribute Dependencies with Embeddings: A Theoretical Analysis ### Paper Abstract During the last decade, entity embeddings have be...
2m0g1wEafh
ICLR.cc/2021/Conference
2021
Benefit of deep learning with non-convex noisy gradient descent: Provable excess risk bound and superiority to kernel methods
["Taiji Suzuki", "Shunta Akiyama"]
Establishing a theoretical analysis that explains why deep learning can outperform shallow learning such as kernel methods is one of the biggest issues in the deep learning literature. Towards answering this question, we evaluate excess risk of a deep learning estimator trained by a noisy gradient descent with ridge re...
["Excess risk", "minimax optimal rate", "local Rademacher complexity", "fast learning rate", "kernel method", "linear estimator"]
ABSTRACTEstablishing a theoretical analysis that explains why deep learning can outperformshallow learning such as kernel methods is one of the biggest issues in the deeplearning literature. Towards answering this question, we evaluate excess risk ofa deep learning estimator trained by a noisy gradient descent with rid...
_HVFOHI0U1
Interesting work but requires more refined explanations and discussions
7: Good paper, accept
This paper is theoretical sound and well organized. This paper shows that the Bayes estimator with Gaussian prior can outperform the linear estimators (including kernel regression and k-NN), which I believe is indeed interesting and important. Besides, I would like to raise the following comments and questions. 1. Th...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Benefit of deep learning with non-convex noisy gradient descent: Provable excess risk bound and superiority to kernel methods ### Paper Abstract Establishing a theoretic...
SW4eu2MmnRY
ML_Reproducibility_Challenge/2021/Fall
2021
Reproducing Results for Crossing the Line: Where do Demographic Variables Fit into Humor Detection?
["Anonymous"]
Reproducibility Summary Scope of Reproducibility Within the original experiment two groups of annotators of size 10 each in age groups 18-25 and 55-70 were relied on to generate the metrics displayed as part of the results. The scope of our study was limited to instances from The Short Jokes dataset on Kaggle with to...
["humor detection"]
Reproducing Results for Crossing the Line: Where doDemographic Variables Fit into Humor Detection?Anonymous Author(s)AffiliationAddressemailReproducibility Summary 12Scope of Reproducibility 3Within the original experiment two groups of annotators of size 10 each in age groups 18-25 and 55-70 were relied on 4to generat...
BGGxo86HIZc
Demographic Variable in Humor Dataset
6: Marginally above acceptance threshold
This paper seems to follow the guidelines for reproducibility, notation, code. The authors propose a method for counterfactual explanation on GNN which uses an extraction decision boundaries shared by multiple samples. They use empirical analysis to justify their theoretical proposal. The author is working in the comp...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Reproducing Results for Crossing the Line: Where do Demographic Variables Fit into Humor Detection? ### Paper Abstract Reproducibility Summary Scope of Reproducibility ...
IHLZ1qPzGbt
eswc-conferences.org/ESWC/2021/Workshop/KGCW
2021
Demo: Knowledge Graph-Based Housing Market Analysis
["Ziping Hu", "Zepei Zhao", "Mohammad Rostami", "Filip Ilievski", "Basel Shbita"]
The housing market is complex and multi-faceted, which makes its analysis challenging for users and professionals. We develop a four-step knowledge graph-based knowledge extraction approach for the housing market for efficient and accurate data analysis, consisting of data acquisition and cleaning, entity linking, onto...
["Knowledge graph", "DBpedia", "COVID-19", "Recommendation"]
Demo: Knowledge Graph-BasedHousing Market AnalysisAbstract. The housing market is complex and multi-faceted, whichmakes its analysis challenging for users and professionals. We developa four-step knowledge graph-based knowledge extraction approach forthe housing market for ecient and accurate data analysis, consistingo...
6L-TkzHYG_r
Complete demo of a knowledge graph for housing market analysis
7: Good paper, accept
The demo paper describes a complete pipeline demonstrating the construction and exploitation of a knowledge graph for housing market analysis. The discussed demo is relevant for the workshop and describes a comprehensive implementation: (i) the pipeline to build the graph involves data crawling from different data sour...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Demo: Knowledge Graph-Based Housing Market Analysis ### Paper Abstract The housing market is complex and multi-faceted, which makes its analysis challenging for users an...
HJMXus0ct7
ICLR.cc/2019/Conference
2019
iRDA Method for Sparse Convolutional Neural Networks
["Xiaodong Jia", "Liang Zhao", "Lian Zhang", "Juncai He", "Jinchao Xu"]
We propose a new approach, known as the iterative regularized dual averaging (iRDA), to improve the efficiency of convolutional neural networks (CNN) by significantly reducing the redundancy of the model without reducing its accuracy. The method has been tested for various data sets, and proven to be significantly mor...
["sparse convolutional neural networks", "regularized dual averaging"]
ABSTRACTWe propose a new approach, known as the iterative regularized dual averaging(iRDA), to improve the efficiency of convolutional neural networks (CNN) bysignificantly reducing the redundancy of the model without reducing its accuracy.The method has been tested for various data sets, and proven to be significantly...
SklGAR9ah7
Algorithm and presentation are flawed
3: Clear rejection
The submission made a few modifications to the RDA (regularized dual averaging) optimization solver to form the proposed "iterative RDA (iRDA)" algorithm, and shows that empirically the proposed algorithm could reduce the number of non-zero parameters by an order of magnitude on CIFAR10 for a number of benchmark netwo...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title iRDA Method for Sparse Convolutional Neural Networks ### Paper Abstract We propose a new approach, known as the iterative regularized dual averaging (iRDA), to improve t...
vyY0jnWG-tK
ICLR.cc/2021/Conference
2021
Physics-aware, probabilistic model order reduction with guaranteed stability
["Sebastian Kaltenbach", "Phaedon Stelios Koutsourelakis"]
Given (small amounts of) time-series' data from a high-dimensional, fine-grained, multiscale dynamical system, we propose a generative framework for learning an effective, lower-dimensional, coarse-grained dynamical model that is predictive of the fine-grained system's long-term evolution but also of its behavior unde...
["inductive bias", "probabilistic generative models", "state-space models", "model order reduction", "slowness", "long-term stability"]
ABSTRACTGiven (small amounts of) time-series’ data from a high-dimensional, fine-grained,multiscale dynamical system, we propose a generative framework for learning aneffective, lower-dimensional, coarse-grained dynamical model that is predictiveof the fine-grained system’s long-term evolution but also of its behavior ...
iQRRZKI1aW6
Interesting idea with encouraging results but lacks necessary motivation and ablations (updated)
6: Marginally above acceptance threshold
Summary: The paper presents a generative approach to modeling physical systems with high-dimensional, nonlinear dynamical systems such as those found in fluid mechanics. The authors provide a physics-motivated hierarchical model for high-dimensional time series and a variational inference method for inferring latent v...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Physics-aware, probabilistic model order reduction with guaranteed stability ### Paper Abstract Given (small amounts of) time-series' data from a high-dimensional, fine...
Sked_0EYwB
ICLR.cc/2020/Conference
2020
Objective Mismatch in Model-based Reinforcement Learning
["Nathan Lambert", "Brandon Amos", "Omry Yadan", "Roberto Calandra"]
Model-based reinforcement learning (MBRL) has been shown to be a powerful framework for data-efficiently learning control of continuous tasks. Recent work in MBRL has mostly focused on using more advanced function approximators and planning schemes, leaving the general framework virtually unchanged since its conception...
["Model-based Reinforcement learning", "dynamics model", "reinforcement learning"]
ABSTRACTModel-based reinforcement learning (MBRL) has been shown to be a powerfulframework for data-efficiently learning control of continuous tasks. Recent workin MBRL has mostly focused on using more advanced function approximators andplanning schemes, leaving the general framework virtually unchanged since itsconcep...
S1eiQv52tB
Official Blind Review #1
3: Weak Reject
The paper "OBJECTIVE MISMATCH IN MODEL-BASED REINFORCEMENT LEARNING" explores the relationships between model optimization and control improvement in model-based reinforcement learning. While it is an interesting problem, the paper fails at demonstrating really useful effects, and the writting needs to be greatly impro...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Objective Mismatch in Model-based Reinforcement Learning ### Paper Abstract Model-based reinforcement learning (MBRL) has been shown to be a powerful framework for data-...
SJx63jRqFm
ICLR.cc/2019/Conference
2019
Diversity is All You Need: Learning Skills without a Reward Function
["Benjamin Eysenbach", "Abhishek Gupta", "Julian Ibarz", "Sergey Levine"]
Intelligent creatures can explore their environments and learn useful skills without supervision. In this paper, we propose ``Diversity is All You Need''(DIAYN), a method for learning useful skills without a reward function. Our proposed method learns skills by maximizing an information theoretic objective using a maxi...
["reinforcement learning", "unsupervised learning", "skill discovery"]
ABSTRACTIntelligent creatures can explore their environments and learn useful skills withoutsupervision. In this paper, we propose “Diversity is All You Need”(DIAYN), amethod for learning useful skills without a reward function. Our proposed methodlearns skills by maximizing an information theoretic objective using a m...
HJxA5jvqnQ
Strong paper with interesting contributions, but with slightly over-stated claims
7: Good paper, accept
This paper proposes a method for learning skills in absence of a reward function. These skills are learned so that the diversity of the trajectories produced by each skill is maximised. This is achieved by having a discriminator attempting to tell these skills apart. The agent is rewarded for visiting states that are e...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Diversity is All You Need: Learning Skills without a Reward Function ### Paper Abstract Intelligent creatures can explore their environments and learn useful skills with...
BkejpaVFDH
ICLR.cc/2020/Conference
2020
Affine Self Convolution
["Nichita Diaconu", "Daniel E. Worrall"]
Attention mechanisms, and most prominently self-attention, are a powerful building block for processing not only text but also images. These provide a parameter efficient method for aggregating inputs. We focus on self-attention in vision models, and we combine it with convolution, which as far as we know, are the firs...
["Self-attention", "convolution", "equivariant"]
ABSTRACTAttention mechanisms, and most prominently self-attention, are a powerful build-ing block for processing not only text but also images. These provide a parameterefficient method for aggregating inputs. We focus on self-attention in vision mod-els, and we combine it with convolution, which as far as we know, are...
SygohhRA5S
Official Blind Review #4
3: Weak Reject
This paper introduces the Affine Self Convolution that combines local connectivity, relative positional embeddings, and a convolutional filter. This form of local self-attention is spatially equivariant and is further extended to include rotational symmetries. Finally, a generalized form of Squeeze and Excite is derive...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Affine Self Convolution ### Paper Abstract Attention mechanisms, and most prominently self-attention, are a powerful building block for processing not only text but also...
6oVAzFsHLFK
NeurIPS.cc/2022/Workshop/SyntheticData4ML
2022
Distributional Privacy for Data Sharing
["Zinan Lin", "Shuaiqi Wang", "Vyas Sekar", "Giulia Fanti"]
Data sharing between different parties has become an important engine powering modern research and development processes. An important class of privacy concerns in data sharing regards the underlying distribution of data. For example, the total traffic volume of data from a networking company reveals the scale of its b...
["privacy", "distributional privacy", "data release", "data sharing", "synthetic data"]
Distributional Privacy for Data SharingZinan Lin∗Carnegie Mellon UniversityPittsburgh, PA 15213zinanlin@microsoft.comShuaiqi Wang∗Carnegie Mellon UniversityPittsburgh, PA 15213shuaiqiw@andrew.cmu.eduVyas SekarCarnegie Mellon UniversityPittsburgh, PA 15213vsekar@andrew.cmu.eduGiulia FantiCarnegie Mellon UniversityPittsb...
RAzTUusyh7J
Authors present a new privacy framework for defining, analyzing, and protecting distributional privacy concerns while data sharing.
6: Marginally above acceptance threshold
Overall a well-written paper that delves into the very important topic of privacy in the context of data sharing. Few comments regarding the manuscript that should definitely be worked on/ addressed, 1. It was not clear what sort of datasets the proposed framework is best suited for? Is it meant only for the cross-sec...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Distributional Privacy for Data Sharing ### Paper Abstract Data sharing between different parties has become an important engine powering modern research and development...
rJx_b3RqY7
ICLR.cc/2019/Conference
2019
AIM: Adversarial Inference by Matching Priors and Conditionals
["Hanbo Li", "Yaqing Wang", "Changyou Chen", "Jing Gao"]
Effective inference for a generative adversarial model remains an important and challenging problem. We propose a novel approach, Adversarial Inference by Matching priors and conditionals (AIM), which explicitly matches prior and conditional distributions in both data and code spaces, and puts a direct constraint on th...
["Generative adversarial network", "inference", "generative model"]
ABSTRACTEffective inference for a generative adversarial model remains an important andchallenging problem. We propose a novel approach, Adversarial Inference byMatching priors and conditionals (AIM), which explicitly matches prior and condi-tional distributions in both data and code spaces, and puts a direct constrain...
r1l9TGz637
Ok paper, some nice comparisons, but too similar to existing models
4: Ok but not good enough - rejection
This paper presents a variant of the adversarial generative modeling framework, allowing it to incorporate an inference mechanism. As such it is very much in the same spirit as existing methods such as ALI/BiGAN. The authors go through an information theoretic motivation but end up with the standard GAN objective funct...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title AIM: Adversarial Inference by Matching Priors and Conditionals ### Paper Abstract Effective inference for a generative adversarial model remains an important and challen...
r1gxEILFu4
ICLR.cc/2019/Workshop/DeepGenStruct
2019
Correlated Variational Auto-Encoders
["Da Tang", "Dawen Liang", "Tony Jebara", "Nicholas Ruozzi"]
Variational Auto-Encoders (VAEs) are capable of learning latent representations for high dimensional data. However, due to the i.i.d. assumption, VAEs only optimize the singleton variational distributions and fail to account for the correlations between data points, which might be crucial for learning latent representa...
["variational", "latent representations", "vaes", "correlations", "correlation structure", "capable", "high dimensional data", "due", "assumption", "singleton variational distributions"]
ABSTRACTVariational Audo-Encoders ( VAEs) are capable of learning latent representationsfor high dimensional data. However, due to the i.i.d. assumption, VAEs only opti-mize the singleton variational distributions and fail to account for the correlationsbetween data points, which might be crucial for learning latent re...
SkeYRMmW54
Decent paper on extending the VAE framework with derived correlated prior from graph structures
4: Top 50% of accepted papers, clear accept
The paper is based on the idea of extending the VAE framework beyond the simplistic IID assumption that it makes on the data. They do so by using structured priors that explicitly take the correlations between data points into account. The methods can be used to learn meaningful representation directly from a weighted ...
2: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Correlated Variational Auto-Encoders ### Paper Abstract Variational Auto-Encoders (VAEs) are capable of learning latent representations for high dimensional data. Howeve...
opHLcXxYTC_
ICLR.cc/2021/Conference
2021
Influence Estimation for Generative Adversarial Networks
["Naoyuki Terashita", "Hiroki Ohashi", "Yuichi Nonaka", "Takashi Kanemaru"]
Identifying harmful instances, whose absence in a training dataset improves model performance, is important for building better machine learning models. Although previous studies have succeeded in estimating harmful instances under supervised settings, they cannot be trivially extended to generative adversarial networ...
["influence", "generative adversarial networks", "data cleansing"]
ABSTRACTIdentifying harmful instances, whose absence in a training dataset improves modelperformance, is important for building better machine learning models. Althoughprevious studies have succeeded in estimating harmful instances under super-vised settings, they cannot be trivially extended to generative adversarial ...
zA_4aQteX3m
Provides a detailed theoretical and experimental analysis of the proposed approach on influence estimation for GANs, but there are concerns regarding its practical applicability.
7: Good paper, accept
Summary: The paper presents an influence estimation method for GANs. It discusses why previous approaches on influence estimation cannot be easily extended to GANs. It proposes to use Jacobian of the gradient of discriminator’s loss with respect to the generator’s parameters to learn how absence of an instance in the...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Influence Estimation for Generative Adversarial Networks ### Paper Abstract Identifying harmful instances, whose absence in a training dataset improves model performance...
hsdU13XxklX
AAAI.org/2023/Bridge/CCBridge
2023
Treatment Effect Estimation to Guide Model Optimization in Continual Learning
["Jonas Seng", "Florian Peter Busch", "Matej Ze\u010devi\u0107", "Moritz Willig"]
Continual Learning systems are faced with a potentially large numbers of tasks to be learned while the models employed have only limited capacity available, which makes it potentially impossible to learn all required tasks within a single model. In order to detect on when a model might break we propose to use treatment...
["model optimization", "model", "treatment effect estimation", "continual", "continual learning systems", "large numbers", "tasks", "models", "limited capacity available"]
Treatment Effect Estimation to Guide Model Optimization in Continual LearningJonas Seng1†, Florian P. Busch1,2, Matej Ze ˇcevi ́c1, Moritz Willig11Computer Science Department, AIML, TU Darmstadt, Germany,2hessian.AI†correspondence: jonas.seng@tu-darmstadt.deAbstractContinual Learning systems are faced with a potentiall...
wKmvTQQuKBZ
Do Calculus for ATE towards Causal Inference in CL
7: Good paper, accept
Authors propose to address the continual learning issues (i.e., updating the current parameter without losing performance on the previous task and learning the model under the fixed resources) from the perspective of Causal Inference. These issues are well-known in continual learning under the name of catastrophic forg...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Treatment Effect Estimation to Guide Model Optimization in Continual Learning ### Paper Abstract Continual Learning systems are faced with a potentially large numbers of...
S1AG8zYeg
ICLR.cc/2017/conference
2017
Sentence Ordering using Recurrent Neural Networks
["Lajanugen Logeswaran", "Honglak Lee", "Dragomir Radev"]
Modeling the structure of coherent texts is a task of great importance in NLP. The task of organizing a given set of sentences into a coherent order has been commonly used to build and evaluate models that understand such structure. In this work we propose an end-to-end neural approach based on the recently proposed se...
["Natural language processing", "Deep learning", "Applications"]
ABSTRACTModeling the structure of coherent texts is a task of great importance in NLP.The task of organizing a given set of sentences into a coherent order has beencommonly used to build and evaluate models that understand such structure. In thiswork we propose an end-to-end neural approach based on the recently propos...
Skba7frNx
interesting problem and promising results
6: Marginally above acceptance threshold
This paper extends the "order matters" idea in (Vinyals et al., 2015) from the sentence level to an interesting application on discourse level. Experiments in this paper show the capacity of the proposed model on both order discrimination task and sentence ordering. I think the problem is interesting and the results a...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Sentence Ordering using Recurrent Neural Networks ### Paper Abstract Modeling the structure of coherent texts is a task of great importance in NLP. The task of organizin...
rkeNqkBFPB
ICLR.cc/2020/Conference
2020
Deep automodulators
["Ari Heljakka", "Yuxin Hou", "Juho Kannala", "Arno Solin"]
We introduce a novel autoencoder model that deviates from traditional autoencoders by using the full latent vector to independently modulate each layer in the decoder. We demonstrate how such an 'automodulator' allows for a principled approach to enforce latent space disentanglement, mixing of latent codes, and a strai...
["unsupervised learning", "generative models", "autoencoders", "disentanglement", "style transfer"]
ABSTRACTWe introduce a novel autoencoder model that deviates from traditional autoen-coders by using the full latent vector to independently modulate each layer inthe decoder. We demonstrate how such an ‘automodulator’ allows for a princi-pled approach to enforce latent space disentanglement, mixing of latent codes,and...
rkgGbTwr5r
Official Blind Review #3
3: Weak Reject
Deep Automodulators introduces a generative autoencoder architecture that replaces the canonical encoder decoder autoencoder architecture with one inspired by StyleGAN. The encoder interacts with the decoder by modulating layer statistics via Adaptive Instance Normalization (AdaIN) conditioned on the latent. The paper ...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Deep automodulators ### Paper Abstract We introduce a novel autoencoder model that deviates from traditional autoencoders by using the full latent vector to independentl...
BJeguTEKDB
ICLR.cc/2020/Conference
2020
INSTANCE CROSS ENTROPY FOR DEEP METRIC LEARNING
["Xinshao Wang", "Elyor Kodirov", "Yang Hua", "Neil M. Robertson"]
Loss functions play a crucial role in deep metric learning thus a variety of them have been proposed. Some supervise the learning process by pairwise or tripletwise similarity constraints while others take the advantage of structured similarity information among multiple data points. In this work, we approach deep metr...
["Deep Metric Learning", "Instance Cross Entropy", "Sample Mining/Weighting", "Image Retrieval"]
ABSTRACTLoss functions play a crucial role in deep metric learning thus a variety of themhave been proposed. Some supervise the learning process by pairwise or triplet-wise similarity constraints while others take advantage of structured similarityinformation among multiple data points. In this work, we approach deep m...
ryxSBCNAFr
Official Blind Review #3
8: Accept
The paper proposes a method to measure the difference between an estimated instance-level matching distribution and its ground-truth one, based on instance cross entropy (ICE). The goal is to learn an embedding that captures the semantic similarities among samples. In particular, with ICE they try to maximize the matc...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title INSTANCE CROSS ENTROPY FOR DEEP METRIC LEARNING ### Paper Abstract Loss functions play a crucial role in deep metric learning thus a variety of them have been proposed. ...
nJ6e-3M2rx
JSYS/2022/Oct_Papers
2022
Solution: Scalable, Contiguous Sequencing for Building Consistent Services
["Anonymous"]
Some recent services use a sequencer to simplify ordering operations on sharded data. The sequencer assigns each operation a multi-sequence number which explicitly orders the operation on each shard it accesses. Existing sequencers have two shortcomings. First, failures can result in some multi-sequence numbers never b...
["scalable", "contiguous", "operation", "sequencers", "abstraction", "sequencer", "second", "solution", "consistent services solution", "consistent services"]
Solution: Scalable, Contiguous Sequencing for Building ConsistentServicesAbstractSome recent services use a sequencer to simplify orderingoperations on sharded data. The sequencer assigns each op-eration a multi-sequence number which explicitly orders theoperation on each shard it accesses. Existing sequencers havetwo ...
MX1NtPP5LL
An excellent example of how consensus-based system can be used in a large scale service
This paper introduced MASON, a consensus-based replicated proxy service that coordinates between clients, sharding datastore and sequencer. The proxy (or proxies) as a replicated state machine which tolerates up to f failures, provided the guarantee of any sequence number assigned to an operation will not be lost, even...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Solution: Scalable, Contiguous Sequencing for Building Consistent Services ### Paper Abstract Some recent services use a sequencer to simplify ordering operations on sha...
sfgcqgOm2F_
ICLR.cc/2021/Conference
2021
Natural Compression for Distributed Deep Learning
["Samuel Horv\u00e1th", "Chen-Yu Ho", "Ludovit Horv\u00e1th", "Atal Narayan Sahu", "Marco Canini", "Peter Richtarik"]
Modern deep learning models are often trained in parallel over a collection of distributed machines to reduce training time. In such settings, communication of model updates among machines becomes a significant performance bottleneck and various lossy update compression techniques have been proposed to alleviate this p...
["nat", "natural compression", "new", "technique", "natural", "distributed deep", "parallel", "collection", "distributed machines"]
ABSTRACTModern deep learning models are often trained in parallel over a collection ofdistributed machines to reduce training time. In such settings, communicationof model updates among machines becomes a significant performance bottleneckand various lossy update compression techniques have been proposed to alleviateth...
Jn0-JW5uMZK
Recommendation to Reject: Not enough contribution
5: Marginally below acceptance threshold
This paper proposes a data compression method named natural compression. This method can be implemented very fast and could be added to other compression methods for additional compression. Then this is generalized to natural dithering for a more aggressive compression. In addition, it compares different compression me...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Natural Compression for Distributed Deep Learning ### Paper Abstract Modern deep learning models are often trained in parallel over a collection of distributed machines...
B1fysiAqK7
ICLR.cc/2019/Conference
2019
Probabilistic Binary Neural Networks
["Jorn W.T. Peters", "Tim Genewein", "Max Welling"]
Low bit-width weights and activations are an effective way of combating the increasing need for both memory and compute power of Deep Neural Networks. In this work, we present a probabilistic training method for Neural Network with both binary weights and activations, called PBNet. By embracing stochasticity during tra...
["binary neural Network", "efficient deep learning", "stochastic training", "discrete neural network", "efficient inference"]
ABSTRACTLow bit-width weights and activations are an effective way of combating theincreasing need for both memory and compute power of Deep Neural Networks. Inthis work, we present a probabilistic training method for Neural Network with bothbinary weights and activations, called PBNet. By embracing stochasticity durin...
BkeHN6yhnQ
Overall score 6
6: Marginally above acceptance threshold
1. This paper presents a novel approach for training efficient binary networks with both binary weights and binary activations and verifies the capability of the proposed method on benchmark. In addition, related bach normalization and pooling layers are also improved. 2. How long do we need for training BNNs using th...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Probabilistic Binary Neural Networks ### Paper Abstract Low bit-width weights and activations are an effective way of combating the increasing need for both memory and c...
H9ulBifsnMm
ICLR.cc/2023/Workshop/RRL
2023
Co-Imitation Learning without Expert Demonstration
["Kun-Peng Ning", "Hu Xu", "Kun Zhu", "Sheng-Jun Huang"]
Imitation learning is a primary approach to improve the efficiency of reinforcement learning by exploiting the expert demonstrations. However, in many real scenarios, obtaining expert demonstrations could be extremely expensive or even impossible. To overcome this challenge, in this paper, we propose a novel learning f...
["Co-Imitation Learning", "Reinforcement Learning"]
ABSTRACTImitation learning is a primary approach to improve the efficiency of reinforce-ment learning by exploiting the expert demonstrations. However, in many realscenarios, obtaining expert demonstrations could be extremely expensive or evenimpossible. To overcome this challenge, in this paper, we propose a novel lea...
k87ZumWBy7a
The more the merrier - faster and better imitation learning through cooperation between agents
4: Good paper, strong accept
This paper proposes a CoIL framework - Co-Imitation Learning, that trains two agents at the same time and “cooperatively” leverages past good experiences of the agents to learn a policy. This approach avoids the need for potentially expensive expert demonstrations in imitation learning, and is better than related appr...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Co-Imitation Learning without Expert Demonstration ### Paper Abstract Imitation learning is a primary approach to improve the efficiency of reinforcement learning by exp...
b6BdrqTnFs7
ICLR.cc/2021/Conference
2021
Grounded Compositional Generalization with Environment Interactions
["Yuanpeng Li"]
In this paper, we present a compositional generalization approach in grounded agent instruction learning. Compositional generalization is an important part of human intelligence, but current neural network models do not have such ability. This is more complicated in multi-modal problems with grounding. Our proposed app...
["compositional generalization", "grounding"]
ABSTRACTIn this paper, we present a compositional generalization approach in groundedagent instruction learning. Compositional generalization is an important part ofhuman intelligence, but current neural network models do not have such ability.This is more complicated in multi-modal problems with grounding. Our propose...
HbyH9KWl65U
Good results, but many questions about the system and experiments
5: Marginally below acceptance threshold
Summary: This paper proposes a new model for the gSCAN dataset (Ruis et al. 2020) which is a synthetically-generated dataset that challenges models to generalize to new compositions of attributes and objects in an instruction. The paper proposes to use "entropy regularization" as a way to enforce that spurious correla...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Grounded Compositional Generalization with Environment Interactions ### Paper Abstract In this paper, we present a compositional generalization approach in grounded agen...
gW8n0uD6rl
ICLR.cc/2021/Conference
2021
Amortized Causal Discovery: Learning to Infer Causal Graphs from Time-Series Data
["Sindy L\u00f6we", "David Madras", "Richard Zemel", "Max Welling"]
Standard causal discovery methods must fit a new model whenever they encounter samples from a new underlying causal graph. However, these samples often share relevant information - for instance, the dynamics describing the effects of causal relations - which is lost when following this approach. We propose Amortized Ca...
["causal graphs", "samples", "causal discovery", "learning", "data", "dynamics", "causal relations", "new model"]
ABSTRACTStandard causal discovery methods must fit a new model whenever they encountersamples from a new underlying causal graph. However, these samples often sharerelevant information – for instance, the dynamics describing the effects of causalrelations – which is lost when following this approach. We propose Amortiz...
QYQca3YOlC4
Causal discovery based on shared dynamics is exciting but does this method work on real(istic) data in the presence of noise?
5: Marginally below acceptance threshold
The paper makes an observation that signal dynamics common to a class of causal systems may contain strong information to enable the use of the encoder of the Neural Relational Inference (2018) for extracting (Granger) causal graphs. It minimally extends the NRI model with an empty edge type and demonstrates that the o...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Amortized Causal Discovery: Learning to Infer Causal Graphs from Time-Series Data ### Paper Abstract Standard causal discovery methods must fit a new model whenever they...
BNU_N-7EIR
KDD.org/2023/Workshop/epiDAMIK
2023
Pandemic Data Collection, Management, Analysis and Decision Support: A Large Urban University Retrospective
["Namrata Banerji", "Steve Chang", "Andrew Perrault", "Tanya Berger-Wolf", "Mikkel Quam"]
The COVID-19 pandemic has disrupted the world. During this crisis, data has emerged as a critical resource for understanding, monitoring, and mitigating the impact of the disease. We present The Ohio State University's data-driven framework for comprehensive monitoring of the COVID-19 pandemic. We discuss the challeng...
["datasets", "public health", "data management", "ethics"]
Pandemic Data Collection, Management, Analysis andDecision Support:A Large Urban University RetrospectiveNamrata Banerjibanerji.8@osu.eduThe Ohio State UniversityColumbus, Ohio, USASteve Changschang@osc.eduOhio Supercomputer CenterColumbus, Ohio, USAAndrew Perraultperrault.17@osu.eduThe Ohio State UniversityColumbus, O...
LtzEQpWylBm
OSU Covid-19 Data Retrospective
2: Marginally below acceptance threshold
# Quality The paper is well written and provides a comprehensive retrospective of OSU's pandemic response # Clarity The paper seems to offer examples rather than comprehensive descriptions of data. Understandably difficult to cover everything, but in a retrospective like this, comprehensive analysis is going to be mor...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Pandemic Data Collection, Management, Analysis and Decision Support: A Large Urban University Retrospective ### Paper Abstract The COVID-19 pandemic has disrupted the wo...
SyxL2TNtvr
ICLR.cc/2020/Conference
2020
Unsupervised Model Selection for Variational Disentangled Representation Learning
["Sunny Duan", "Loic Matthey", "Andre Saraiva", "Nick Watters", "Chris Burgess", "Alexander Lerchner", "Irina Higgins"]
Disentangled representations have recently been shown to improve fairness, data efficiency and generalisation in simple supervised and reinforcement learning tasks. To extend the benefits of disentangled representations to more complex domains and practical applications, it is important to enable hyperparameter tuning ...
["unsupervised disentanglement metric", "disentangling", "representation learning"]
ABSTRACTDisentangled representations have recently been shown to improve fairness, dataefficiency and generalisation in simple supervised and reinforcement learning tasks.To extend the benefits of disentangled representations to more complex domains andpractical applications, it is important to enable hyperparameter tu...
Byx8ouwAFr
Official Blind Review #1
6: Weak Accept
This paper proposes a criterion called Unsupervised Disentanglement Ranking (UDR) score. The score is computed based on the assumption that good disentangled representations are alike, while the representations can be entangled in multiple possible ways. The UDR score can be used for unsupervised hyperparameter tuning...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Unsupervised Model Selection for Variational Disentangled Representation Learning ### Paper Abstract Disentangled representations have recently been shown to improve fai...
iMIPkk143Rh
icaps-conference.org/ICAPS/2022/Workshop/XAIP
2022
Explaining Soft-Goal Conflicts through Constraint Relaxations
["Rebecca Eifler", "Jeremy Frank", "Joerg Hoffmann"]
Recent work suggests to explain trade-offs between soft goals in terms of their conflicts, i. e., minimal unsolvable soft-goal subsets. But this does not explain the conflicts themselves: Why can a given set of soft goals not be jointly achieved? Here we approach that question in terms of the underlying constraints on ...
["conflicts", "algorithms", "soft goals", "terms", "conflict", "baseline", "relaxed tasks", "constraint relaxations", "minimal unsolvable"]
Explaining Soft-Goal Conflicts through Constraint RelaxationsRebecca Eifler1,Jeremy Frank2and J ̈org Hoffmann1;31Saarland University, Saarland Informatics Campus, Germany2NASA Ames Research Center, Mountain View, CA, USA3German Research Center for Artificial Intelligence (DFKI), Saarbr ̈ucken, Germanyfeifler, hoffmanng...
jHnckf1tExE
Interesting manuscript on conflict explanation
7: Good paper, accept
The manuscript contributes to generating explanations for soft-goal conflicts in planning. The topic is clearly relevant to XAIP. The authors propose to explain soft-goal conflicts by relaxing resource and time constraints until the to-be-explained soft-goal conflict does not occur anymore and report the relaxation as ...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Explaining Soft-Goal Conflicts through Constraint Relaxations ### Paper Abstract Recent work suggests to explain trade-offs between soft goals in terms of their conflict...
HJxDugSFDB
ICLR.cc/2020/Conference
2020
Stochastic Latent Actor-Critic: Deep Reinforcement Learning with a Latent Variable Model
["Alex X. Lee", "Anusha Nagabandi", "Pieter Abbeel", "Sergey Levine"]
Deep reinforcement learning (RL) algorithms can use high-capacity deep networks to learn directly from image observations. However, these kinds of observation spaces present a number of challenges in practice, since the policy must now solve two problems: a representation learning problem, and a task learning problem....
["slac", "stochastic latent", "deep reinforcement", "model", "sample efficiency", "algorithms", "deep networks"]
ABSTRACTDeep reinforcement learning (RL) algorithms can use high-capacity deep networksto learn directly from image observations. However, these kinds of observationspaces present a number of challenges in practice, since the policy must now solvetwo problems: a representation learning problem, and a task learning prob...
SJe5StvycH
Official Blind Review #2
8: Accept
The authors propose SLAC, an important extension of the recently introduced soft actor critic (SAC) algorthm, which operates on a learned latent state, rather than an observed one, and therefore aims to jointly learn to represent high dimensional inputs and execute continuous control based on this representation. SLA...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Stochastic Latent Actor-Critic: Deep Reinforcement Learning with a Latent Variable Model ### Paper Abstract Deep reinforcement learning (RL) algorithms can use high-capa...
kic8cng35wX
ICLR.cc/2021/Conference
2021
Weak NAS Predictor Is All You Need
["Junru Wu", "Xiyang Dai", "Dongdong Chen", "Yinpeng Chen", "Mengchen Liu", "Ye Yu", "Zhangyang Wang", "Zicheng Liu", "Mei Chen", "Lu Yuan"]
Neural Architecture Search (NAS) finds the best network architecture by exploring the architecture-to-performance manifold. It often trains and evaluates a large amount of architectures, causing tremendous computation cost. Recent predictor-based NAS approaches attempt to solve this problem with two key steps: sampling...
["performance predictor", "neural architecture search"]
ABSTRACTNeural Architecture Search (NAS) finds the best network architecture by explor-ing the architecture-to-performance manifold. It often trains and evaluates a largeamount of architectures, causing tremendous computation cost. Recent predictor-based NAS approaches attempt to solve this problem with two key steps: ...
kUFXa0Os92c
This work is more like an incremental improvement to the neural predictor. It gets a small performance improvement by learning in a better subspace.
4: Ok but not good enough - rejection
Pros: - A progressive method for neural predictor-based NAS is proposed, and it shows better performance than previous methods. - Experiments on the NAS-bench-101 and NAS-bench-201, as well as on the NASNet search space to validate the effectiveness of the method. - The proposed method is conceptually simple and effici...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Weak NAS Predictor Is All You Need ### Paper Abstract Neural Architecture Search (NAS) finds the best network architecture by exploring the architecture-to-performance m...
B1kJ6H9ex
ICLR.cc/2017/conference
2017
Combining policy gradient and Q-learning
["Brendan O'Donoghue", "Remi Munos", "Koray Kavukcuoglu", "Volodymyr Mnih"]
Policy gradient is an efficient technique for improving a policy in a reinforcement learning setting. However, vanilla online variants are on-policy only and not able to take advantage of off-policy data. In this paper we describe a new technique that combines policy gradient with off-policy Q-learning, drawing experie...
["Deep learning", "Reinforcement Learning"]
ABSTRACTPolicy gradient is an efficient technique for improving a policy in a reinforcementlearning setting. However, vanilla online variants are on-policy only and not ableto take advantage of off-policy data. In this paper we describe a new technique thatcombines policy gradient with off-policy Q-learning, drawing ex...
Hk2oimlSe
An excellent paper pointing out connections between existing algorithms, which also leads to new algorithms with excellent results
9: Top 15% of accepted papers, strong accept
This is a very nicely written paper which unifies some value-based and policy-based (regularized policy gradient) methods, by pointing out connections between the value function and policy which have not been established before. The theoretical results are new and insightful, and will likely be useful in the RL field m...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Combining policy gradient and Q-learning ### Paper Abstract Policy gradient is an efficient technique for improving a policy in a reinforcement learning setting. However...
Hklo5RNtwS
ICLR.cc/2020/Conference
2020
Behavior-Guided Reinforcement Learning
["Aldo Pacchiano", "Jack Parker-Holder", "Yunhao Tang", "Anna Choromanska", "Krzysztof Choromanski", "Michael I. Jordan"]
We introduce a new approach for comparing reinforcement learning policies, using Wasserstein distances (WDs) in a newly defined latent behavioral space. We show that by utilizing the dual formulation of the WD, we can learn score functions over trajectories that can be in turn used to lead policy optimization towards (...
["Reinforcement Learning", "Optimal Transport", "Evolution Strategies"]
ABSTRACTWe introduce a new approach for comparing reinforcement learning policies, usingWasserstein distances (WDs) in a newly defined latent behavioral space. We showthat by utilizing the dual formulation of the WD, we can learn score functionsover trajectories that can be in turn used to lead policy optimization towa...
r1l0QRy0FH
Official Blind Review #3
3: Weak Reject
This work explores two uses of Wasserstein distances (WD) within reinforcement learning: the first is a variant of policy gradient, where WD is used to guide the policy search (instead of alternative such as Trust-region used in TRPO); the second is a variant of evolutionary search where WD is used again to guide the p...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Behavior-Guided Reinforcement Learning ### Paper Abstract We introduce a new approach for comparing reinforcement learning policies, using Wasserstein distances (WDs) in...
rkxMKerYwr
ICLR.cc/2020/Conference
2020
Towards Interpreting Deep Neural Networks via Understanding Layer Behaviors
["Jiezhang Cao", "Jincheng Li", "Xiping Hu", "Peilin Zhao", "Mingkui Tan"]
Deep neural networks (DNNs) have achieved unprecedented practical success in many applications. However, how to interpret DNNs is still an open problem. In particular, what do hidden layers behave is not clearly understood. In this paper, relying on a teacher-student paradigm, we seek to understand the layer behaviors...
["Interpretability of DNNs", "Wasserstein distance", "Layer behavior"]
ABSTRACTDeep neural networks (DNNs) have achieved unprecedented practical success inmany applications. However, how to interpret DNNs is still an open problem. Inparticular, what do hidden layers behave is not clearly understood. In this paper,relying on a teacher-student paradigm, we seek to understand the layer behav...
HkxW9dLptr
Official Blind Review #1
3: Weak Reject
The authors intuitively, and then analytically, explain the behavior in the hidden layers of deep convolutional networks and show how the behavior can be used to improve performance by "early exiting." I give this paper a weak reject. I believe this paper does well by connecting the intuitive explanation with the proo...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Towards Interpreting Deep Neural Networks via Understanding Layer Behaviors ### Paper Abstract Deep neural networks (DNNs) have achieved unprecedented practical success ...
rkxNh1Stvr
ICLR.cc/2020/Conference
2020
Quantifying Point-Prediction Uncertainty in Neural Networks via Residual Estimation with an I/O Kernel
["Xin Qiu", "Elliot Meyerson", "Risto Miikkulainen"]
Neural Networks (NNs) have been extensively used for a wide spectrum of real-world regression tasks, where the goal is to predict a numerical outcome such as revenue, effectiveness, or a quantitative result. In many such tasks, the point prediction is not enough: the uncertainty (i.e. risk or confidence) of that predic...
["Uncertainty Estimation", "Neural Networks", "Gaussian Process"]
ABSTRACTNeural Networks (NNs) have been extensively used for a wide spectrum of real-world regression tasks, where the goal is to predict a numerical outcome suchas revenue, effectiveness, or a quantitative result. In many such tasks, the pointprediction is not enough: the uncertainty (i.e. risk or confidence) of that ...
HJx95fEwYH
Official Blind Review #2
6: Weak Accept
This paper proposes a new framework (RIO) to estimate uncertainty in pretrained neural networks. For this purpose, RIO employs Gaussian Processes whose kernels are calculated by kernel functions of input and output samples and the corresponding target values. - The proposed approach is interesting and the initial resu...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Quantifying Point-Prediction Uncertainty in Neural Networks via Residual Estimation with an I/O Kernel ### Paper Abstract Neural Networks (NNs) have been extensively use...
TYXs_y84xRj
ICLR.cc/2021/Conference
2021
PolarNet: Learning to Optimize Polar Keypoints for Keypoint Based Object Detection
["Wu Xiongwei", "Doyen Sahoo", "Steven HOI"]
A variety of anchor-free object detectors have been actively proposed as possible alternatives to the mainstream anchor-based detectors that often rely on complicated design of anchor boxes. Despite achieving promising performance on par with anchor-based detectors, the existing anchor-free detectors such as FCOS or Ce...
["Object Detection", "Deep Learning"]
ABSTRACTA variety of anchor-free object detectors have been actively proposed as possible al-ternatives to the mainstream anchor-based detectors that often rely on complicateddesign of anchor boxes. Despite achieving promising performance on par withanchor-based detectors, the existing anchor-free detectors such as FCO...
V4JzA1A6vB
Official review
3: Clear rejection
########################################################################## Summary: This paper proposes an anchor-free object detector that does bounding box regression in the polar coordinate instead of in the Cartesian coordinate. The motivation of doing this is because there are larger variance in offset vectors i...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title PolarNet: Learning to Optimize Polar Keypoints for Keypoint Based Object Detection ### Paper Abstract A variety of anchor-free object detectors have been actively propos...
diB-RsDQJeG
NeurIPS.cc/2022/Workshop/SyntheticData4ML
2022
Causal Structural Hypothesis Testing and Data Generation Models
["Sunay Gajanan Bhat", "Omead Pooladzandi", "Jeffrey Jiang", "Gregory Pottie"]
A vast amount of expert and domain knowledge is captured by causal structural priors, yet there has been little research on testing such priors for generalization and data synthesis purposes. We propose a novel model architecture, Causal Structural Hypothesis Testing, that can use nonparametric, structural causal knowl...
["causality", "generative modeling", "generalization", "hypothesis testing"]
Causal Structural Hypothesis Testing and DataGeneration ModelsJeffrey Jiang∗jimmery@ucla.eduOmead Pooladzandi∗opooladz@ucla.eduSunay Bhat∗sunaybhat1@ucla.eduGregory Pottie∗pottie@ee.ucla.eduAbstractA vast amount of expert and domain knowledge is captured by causal structuralpriors, yet there has been little research on...
zT-E-jl6AxL
Interesting idea but connections to related work should be clarified more
6: Marginally above acceptance threshold
# Summary The authors propose two generative models that assume access to a causal DAG to learn the structural assignment functions. They then apply this for hypothesis testing and out-of-distribution data generation. # Strengths * the causal hypothesis testing problem is well-motivated as causal DAG assumptions are ...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Causal Structural Hypothesis Testing and Data Generation Models ### Paper Abstract A vast amount of expert and domain knowledge is captured by causal structural priors, ...
SJMeTo09YQ
ICLR.cc/2019/Conference
2019
Guided Exploration in Deep Reinforcement Learning
["Sahisnu Mazumder", "Bing Liu", "Shuai Wang", "Yingxuan Zhu", "Xiaotian Yin", "Lifeng Liu", "Jian Li", "Yongbing Huang"]
This paper proposes a new method to drastically speed up deep reinforcement learning (deep RL) training for problems that have the property of \textit{state-action permissibility} (SAP). Two types of permissibility are defined under SAP. The first type says that after an action $a_t$ is performed in a state $s_t$ and t...
["deep reinforcement learning", "guided exploration", "RL training speed up"]
ABSTRACTThis paper proposes a new method to drastically speed up deep reinforcementlearning (deep RL) training for problems that have the property of state-actionpermissibility (SAP). Two types of permissibility are defined under SAP. The firsttype says that after an action atis performed in a state stand the agent rea...
HyeHZTbdn7
The application of SAP seems very narrow.
3: Clear rejection
This paper proposed the concept of state-action permissibility (SAP). Given a user-defined type 1 SAP function, the algorithm learns a classifier to predict whether an action at a given state is permissible or not. Based on this prediction, the reinforcement learning (RL) algorithms can limit the exploration only to th...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Guided Exploration in Deep Reinforcement Learning ### Paper Abstract This paper proposes a new method to drastically speed up deep reinforcement learning (deep RL) train...
cO1IH43yUF
ICLR.cc/2021/Conference
2021
Revisiting Few-sample BERT Fine-tuning
["Tianyi Zhang", "Felix Wu", "Arzoo Katiyar", "Kilian Q Weinberger", "Yoav Artzi"]
This paper is a study of fine-tuning of BERT contextual representations, with focus on commonly observed instabilities in few-sample scenarios. We identify several factors that cause this instability: the common use of a non-standard optimization method with biased gradient estimation; the limited applicability of sign...
["Fine-tuning", "Optimization", "BERT"]
ABSTRACTThis paper is a study of fine-tuning of BERT contextual representations, with focuson commonly observed instabilities in few-sample scenarios. We identify severalfactors that cause this instability: the common use of a non-standard optimizationmethod with biased gradient estimation; the limited applicability of...
DviyGBgd1TU
Thorough Analysis of Various Finetuning Factors
6: Marginally above acceptance threshold
The paper focuses on instability issues in BERT finetuning on small datasets. They list three factors which leads to instability, and provide simple fixes for each: 1. Lack of bias correction term in BertAdam -- Fix was to use standard Adam 2. Using all pretrained layers for finetuning -- Reinitializing the last few l...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Revisiting Few-sample BERT Fine-tuning ### Paper Abstract This paper is a study of fine-tuning of BERT contextual representations, with focus on commonly observed instab...
a0yodLze7gs
ICLR.cc/2021/Conference
2021
Disentangling Action Sequences: Discovering Correlated Samples
["Jiantao Wu", "Chunxiuzi Liu", "Lin Wang"]
Disentanglement is a highly desirable property of representation due to its similarity with human’s understanding and reasoning. This improves interpretability, enables the performance of down-stream tasks, and enables controllable generative models.However, this domain is challenged by the abstract notion and incomple...
["disentanglement", "representation learning", "unsupervised", "inductive bias"]
ABSTRACTDisentanglement is a highly desirable property of representation due to its similaritywith human’s understanding and reasoning. This improves interpretability, enablesthe performance of down-stream tasks, and enables controllable generative models.However, this domain is challenged by the abstract notion and in...
0Lnoczyks-X
Interesting model idea w.r.t. annealing capacity of latent representation, but relation to competing SOTA approaches requires more clarity
4: Ok but not good enough - rejection
### Summary: In this submission, a common modelling assumption for unsupervised disentanglement is challenged: that the disentangled representation follows the independence structure of the underlying (data generating) factors. Instead, the paper proposes to consider *action sequences* which describe how datapoints are...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Disentangling Action Sequences: Discovering Correlated Samples ### Paper Abstract Disentanglement is a highly desirable property of representation due to its similarity ...
w6PFZGCPm_T
ICML.cc/2020/Workshop/SAS
2020
Semi-supervised audio tagging with deep co-training and augmentations
["Leo Cances", "Thomas Pellegrini"]
In this work, we explore the task of audio tagging in a semi-supervised context. The recently proposed Deep Co-Training (DCT) algorithm has shown impressive results in visual object recognition and outperformed other semi-supervised state-of-the-art methods such as Mean Teacher and GANs. DCT uses two or more deep neura...
["Deep Co-Training", "semi supervised learning", "audio tagging"]
000001002003004005006007008009010011012013014015016017018019020021022023024025026027028029030031032033034035036037038039040041042043044045046047048049050051052053054Semi-supervised audio tagging with deep co-training and augmentationsAnonymous Authors1AbstractIn this work, we explored the task of audio tag-ging in a se...
tjtmVA5VACq
Interesting work but could be explained better
4: Ok but not good enough - rejection
Notes: -Paper considers semi-supervised audio tagging. -Paper uses "deep co-training" which had been successful for visual object recognition. -DCT uses two or more deep nets and adversarial examples? -Paper uses audio data augmentations (calling it duplication in the abstract is strange, since presu...
2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Semi-supervised audio tagging with deep co-training and augmentations ### Paper Abstract In this work, we explore the task of audio tagging in a semi-supervised context....
Byx93sC9tm
ICLR.cc/2019/Conference
2019
Deep Ensemble Bayesian Active Learning : Adressing the Mode Collapse issue in Monte Carlo dropout via Ensembles
["Remus Pop", "Patric Fulop"]
In image classification tasks, the ability of deep convolutional neural networks (CNNs) to deal with complex image data has proved to be unrivalled. Deep CNNs, however, require large amounts of labeled training data to reach their full potential. In specialised domains such as healthcare, labeled data can be difficult ...
["Active Learning", "Deep Learning", "Bayesian Neural Networks", "Bayesian Deep Learning", "Ensembles"]
ABSTRACTIn image classification tasks, the ability of deep convolutional neural networks(CNNs) to deal with complex image data has proved to be unrivalled. Deep CNNs,however, require large amounts of labeled training data to reach their full poten-tial. In specialized domains such as healthcare, labeled data can be dif...
BJgzh7Schm
Ensemble of MC-Dropout models is not an approximation of the posterior
4: Ok but not good enough - rejection
The authors propose to use the combination of model ensemble and MC dropout in Bayesian deep active learning. They empirically show that there exists the mode collapse problem due to the MC dropout which can be regarded as a variational approximation. The authors introduce an ensemble of MC-Dropout models with differen...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Deep Ensemble Bayesian Active Learning : Adressing the Mode Collapse issue in Monte Carlo dropout via Ensembles ### Paper Abstract In image classification tasks, the abi...
BJgctpEKwr
ICLR.cc/2020/Conference
2020
RPGAN: random paths as a latent space for GAN interpretability
["Andrey Voynov", "Artem Babenko"]
In this paper, we introduce Random Path Generative Adversarial Network (RPGAN) --- an alternative scheme of GANs that can serve as a tool for generative model analysis. While the latent space of a typical GAN consists of input vectors, randomly sampled from the standard Gaussian distribution, the latent space of RPGAN ...
["generative models", "GAN", "interpretability"]
ABSTRACTIn this paper, we introduce Random Path Generative Adversarial Network (RP-GAN) — an alternative design of GANs that can serve as a tool for generativemodel analysis. While the latent space of a typical GAN consists of input vec-tors, randomly sampled from the standard Gaussian distribution, the latent spaceof ...
HklH2Ycr9S
Official Blind Review #2
3: Weak Reject
This paper addresses the issue of interpretability of GAN generation through an alternative approach to the introduction of variability. To seed the generation, instead of providing a random input vector (typically sampled from a standard Gaussian distribution), the authors instead modify the generator architecture so ...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title RPGAN: random paths as a latent space for GAN interpretability ### Paper Abstract In this paper, we introduce Random Path Generative Adversarial Network (RPGAN) --- an a...
M9hdyCNlWaf
ICLR.cc/2021/Conference
2021
Sparse Uncertainty Representation in Deep Learning with Inducing Weights
["Hippolyt Ritter", "Martin Kukla", "Cheng Zhang", "Yingzhen Li"]
Bayesian neural networks and deep ensembles represent two modern paradigms of uncertainty quantification in deep learning. Yet these approaches struggle to scale mainly due to memory inefficiency issues, since they require parameter storage several times higher than their deterministic counterparts. To address this, we...
["Bayesian neural networks", "uncertainty estimation", "memory efficiency"]
ABSTRACTBayesian neural networks and deep ensembles represent two modern paradigmsof uncertainty quantification in deep learning. Yet these approaches struggle toscale mainly due to memory inefficiency issues, since they require parameter stor-age several times higher than their deterministic counterparts. To address t...
KUPFW1L8T_E
Interesting work, some key baselines are missing
6: Marginally above acceptance threshold
### Summary This work proposes a specific parametrisation for the Gaussian prior and approximate posterior distribution in variational Bayesian neural networks in terms of inducing weights. The general idea is an instance of the sparse variational inference scheme for GPs proposed by Titsias back in 2009; for a given m...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Sparse Uncertainty Representation in Deep Learning with Inducing Weights ### Paper Abstract Bayesian neural networks and deep ensembles represent two modern paradigms of...
lvXLfNeCQdK
ICLR.cc/2021/Conference
2021
Loss Landscape Matters: Training Certifiably Robust Models with Favorable Loss Landscape
["Sungyoon Lee", "Woojin Lee", "Jinseong Park", "Jaewook Lee"]
In this paper, we study the problem of training certifiably robust models. Certifiable training minimizes an upper bound on the worst-case loss over the allowed perturbation, and thus the tightness of the upper bound is an important factor in building certifiably robust models. However, many studies have shown that Int...
["Adversarial Examples", "Certifiable Robustness", "Certifiable Training", "Loss Landscape", "Deep Learning", "Security"]
ABSTRACTIn this paper, we study the problem of training certifiably robust models. Certifi-able training minimizes an upper bound on the worst-case loss over the allowedperturbation, and thus the tightness of the upper bound is an important factor inbuilding certifiably robust models. However, many studies have shown t...
2XK6OVV509P
Central claims unsupported by evaluation
3: Clear rejection
This paper claims to make three contributions: (i) a theoretical and experimental demonstration that the driving force behind performance in certifiable training methods is loss landscape smoothness (ii) a new training method and domain designed to have a smooth loss landscape (iii) an evaluation showing their new tra...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Loss Landscape Matters: Training Certifiably Robust Models with Favorable Loss Landscape ### Paper Abstract In this paper, we study the problem of training certifiably r...
HylINLUKuV
ICLR.cc/2019/Workshop/DeepGenStruct
2019
WiSE-ALE: Wide Sample Estimator for Aggregate Latent Embedding
["Shuyu Lin", "Ronald Clark", "Robert Birke", "Niki Trigoni", "Stephen Roberts"]
In this paper, we present a new generative model for learning latent embeddings. Compared to the classical generative process, where each observed data point is generated from an individual latent variable, our approach assumes a global latent variable to generate the whole set of observed data points. We then propose ...
["Generative models", "Latent variable models", "Variational inference", "Auto-encoder", "Representation learning"]
ABSTRACTIn this paper, we present a new generative model for learning latent embeddings.Compared to the classical generative process, where each observed data pointis generated from an individual latent variable, our approach assumes a globallatent variable to generate the whole set of observed data points. We then pro...
HklOGRbmc4
solid work
4: Top 50% of accepted papers, clear accept
This paper derives a variational lower bound to the data log likelihood, which allows us to impose a prior constraint on the bulk statistics of the aggregate posterior distribution for the entire dataset. The analysis shows that the proposed method achieves better reconstruction quality, as well as forming a smooth, co...
2: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title WiSE-ALE: Wide Sample Estimator for Aggregate Latent Embedding ### Paper Abstract In this paper, we present a new generative model for learning latent embeddings. Compar...
r1eVXa4KvH
ICLR.cc/2020/Conference
2020
Concise Multi-head Attention Models
["Srinadh Bhojanapalli", "Chulhee Yun", "Ankit Singh Rawat", "Sashank Reddi", "Sanjiv Kumar"]
Attention based Transformer architecture has enabled significant advances in the field of natural language processing. In addition to new pre-training techniques, recent improvements crucially rely on working with a relatively larger embedding dimension for tokens. This leads to models that are prohibitively large to b...
["Transformers", "Attention", "Multihead", "expressive power", "embedding size"]
ABSTRACTAttention based Transformer architecture has enabled significant advances in thefield of natural language processing. In addition to new pre-training techniques,recent improvements crucially rely on working with a relatively larger embeddingdimension for tokens. Unfortunately, this leads to models that are proh...
BJxY2KeTtH
Official Blind Review #2
3: Weak Reject
This work studies the head size <--> head number tradeoff in multihead attention. It argues and formally establishes that (1) the expressivity of an attention head is determined by its dimension and (b) fixing the head dimension, one gains additional expressive power by using more heads. In response to such observation...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Concise Multi-head Attention Models ### Paper Abstract Attention based Transformer architecture has enabled significant advances in the field of natural language process...
4PdgBOT21I
graphicsinterface.org/Graphics_Interface/2023/Conference
2023
Saliency Driven Gaze Control for Autonomous Pedestrians
["Anonymous"]
How and why an agent looks at its environment can inform its navigation, behaviour and interaction with the environment. A human agent's visual-motor system is complex and requires both an understanding of visual stimulus as well as adaptive methods to control and aim its gaze in accordance with goal-driven behaviour o...
["Gaze Control", "Saliency", "Virtual Agent Behaviour"]
ABSTRACTHow and why an agent looks at its environment can inform its navi-gation, behaviour and interaction with the environment. A humanagent’s visual-motor system is complex and requires both an under-standing of visual stimulus as well as adaptive methods to controland aim its gaze in accordance with goal-driven beh...
K4-733qv7C
Interesting topic, but limited technical contribution and inadequate evaluation.
4: Ok but not good enough - rejection
This paper advocates generating gaze movement for virtual characters using only agent-perceivable information. To achieve that, it presents two procedural techniques to control the gaze direction based on saliency maps extracted from visual input: the first technique treats the saliency map as a potential field and si...
2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Saliency Driven Gaze Control for Autonomous Pedestrians ### Paper Abstract How and why an agent looks at its environment can inform its navigation, behaviour and interac...
36G2rwDbk1k
ICLR.cc/2021/Conference
2021
On the Discovery of Feature Importance Distribution: An Overlooked Area
["Yuxiao Huang"]
Detecting feature's predictive power is a key problem in Machine Learning. Previous methods have been focusing on providing a single value, usually named feature importance, as a point estimate of the power. However, it is difficult to interpret the predictive power using feature importance. Moreover, in reality featur...
["feature importance", "predictive power", "discovery", "feature importance distribution", "overlooked area", "feature", "point estimate", "feature values", "key problem", "machine learning"]
ABSTRACTDetecting feature’s predictive power is a key problem in Machine Learning. Pre-vious methods have been focusing on providing a single value, usually named fea-ture importance, as a point estimate of the power. However, it is difficult to inter-pret the predictive power using feature importance. Moreover, in rea...
5bYrvlVfwv
Review of "On the Discovery of Feature Importance Distribution: An Overlooked Area"
4: Ok but not good enough - rejection
Summary: The authors propose estimating a feature/variable importance measure that varies as the value of the variable varies. In this way, one can discover significant variation in feature importance that cannot be summarized using a single point estimate. Pros: + In the variable importance literature, much of the fo...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title On the Discovery of Feature Importance Distribution: An Overlooked Area ### Paper Abstract Detecting feature's predictive power is a key problem in Machine Learning. Pre...
Aj4_e50nB8
ICLR.cc/2021/Conference
2021
Contextual Knowledge Distillation for Transformer Compression
["Geondo Park", "Gyeongman Kim", "Eunho Yang"]
A computationally expensive and memory intensive neural network lies behind the recent success of language representation learning. Knowledge distillation, a major technique for deploying such a vast language model in resource-scarce environments, transfers the knowledge on individual word representations learned witho...
["Knowledge Distillation", "Transformer Compression", "BERT"]
ABSTRACTA computationally expensive and memory intensive neural network lies behindthe recent success of language representation learning. Knowledge distillation, amajor technique for deploying such a vast language model in resource-scarce en-vironments, transfers the knowledge on individual word representations learne...
x5HeTbUJw6j
A paper that proposes two new distillation objective based on word relations and layer transforming relations that outperform previous distillation methods.
6: Marginally above acceptance threshold
This paper proposes two new distillation objective, word relations and layer transforming relations. Word relations constrain the pairwise/triplet relations of embeddings at each layer to be closer to the teacher network. The layer transforming relations constrain that the pairwise/triplet relations of embeddings betwe...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Contextual Knowledge Distillation for Transformer Compression ### Paper Abstract A computationally expensive and memory intensive neural network lies behind the recent s...
n7wIfYPdVet
ICLR.cc/2021/Conference
2021
Auxiliary Learning by Implicit Differentiation
["Aviv Navon", "Idan Achituve", "Haggai Maron", "Gal Chechik", "Ethan Fetaya"]
Training neural networks with auxiliary tasks is a common practice for improving the performance on a main task of interest. Two main challenges arise in this multi-task learning setting: (i) designing useful auxiliary tasks; and (ii) combining auxiliary tasks into a single coherent loss. Here, we propose a novel frame...
["Auxiliary Learning", "Multi-task Learning"]
ABSTRACTTraining neural networks with auxiliary tasks is a common practice for improvingthe performance on a main task of interest. Two main challenges arise in thismulti-task learning setting: (i) designing useful auxiliary tasks; and (ii) combiningauxiliary tasks into a single coherent loss. Here, we propose a novel ...
4hTy4SxESlT
The intuition makes sense. Comprehensive experiments are done.
6: Marginally above acceptance threshold
This paper studies a variant of multi-task learning, auxiliary learning, where one main task dominates, and other tasks are used to learn a good representation. To achieve this goal, the authors propose a learning-to-learn algorithm. In particular, the auxiliary losses are represented by a vector and then transformed t...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Auxiliary Learning by Implicit Differentiation ### Paper Abstract Training neural networks with auxiliary tasks is a common practice for improving the performance on a m...
ByecAoAqK7
ICLR.cc/2019/Conference
2019
Zero-shot Dual Machine Translation
["Lierni Sestorain", "Massimiliano Ciaramita", "Christian Buck", "Thomas Hofmann"]
Neural Machine Translation (NMT) systems rely on large amounts of parallel data.This is a major challenge for low-resource languages. Building on recent work onunsupervised and semi-supervised methods, we present an approach that combineszero-shot and dual learning. The latter relies on reinforcement learning, to explo...
["unsupervised", "machine translation", "dual learning", "zero-shot"]
ABSTRACTNeural Machine Translation (NMT) systems rely on large amounts of parallel data.This is a major challenge for low-resource languages. Building on recent work onunsupervised and semi-supervised methods, we present an approach that combineszero-shot and dual learning. The latter relies on reinforcement learning, ...
ryek_zaK2Q
Limited novelty; Experiments are not enough
5: Marginally below acceptance threshold
[Summary] This paper proposed an algorithm for zero-shot translation by using both dual learning (He et al, 2016) and multi-lingual neural machine translation (Johnson et al 2016). Specially, a multilingual model is first trained following (Johnson et al 2016) and then the dual learning (He et al 2016) is applied to t...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Zero-shot Dual Machine Translation ### Paper Abstract Neural Machine Translation (NMT) systems rely on large amounts of parallel data.This is a major challenge for low-r...
hZlwUFmka-U
graphicsinterface.org/Graphics_Interface/2023/Conference_SD
2023
Generating Packed Rectilinear Display Text Layouts with Weighted Word Emphasis
["Cheryl Lao", "Craig S. Kaplan", "Daniel Vogel", "Jose Echevarria", "Paul Asente"]
A common text layout style is a "packed rectilinear layout", in which non-overlapping word bounding boxes are packed so that their union forms a rectangle with no holes other than word and line spacing. Designing variations of these layouts while preserving word emphasis is a difficult and time-consuming process. We pr...
["Typography", "Packing", "Display Text", "Emphasis"]
ABSTRACTA common text layout style is a “packed rectilinear layout”, in whichnon-overlapping word bounding boxes are packed so that their unionforms a rectangle with no holes other than word and line spacing.Designing variations of these layouts while preserving word empha-sis is a difficult and time-consuming process....
CO4bpCMtyI
review for rectilinear display text layouts
5: Marginally below acceptance threshold
Paper Summary. The paper introduces a method for generating rectilinear text layouts. In order to enable various layout designs, the proposed method takes use of a phrase tree structure inspired from an existing image layout algorithm. During this layout generation process, reading order and spacing are considered. The...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Generating Packed Rectilinear Display Text Layouts with Weighted Word Emphasis ### Paper Abstract A common text layout style is a "packed rectilinear layout", in which n...
DBBM-7ObAem
MICCAI.org/2021/Workshop/COMPAY
2021
Synthesis of Annotated Colorectal Cancer Tissue Images from Gland Layout
["Anonymous"]
Generating annotated pairs of realistic tissue images along with their annotations is a challenging task in computational histopathology. Such synthetic images and their annotations can be useful in training and evaluation of algorithms in the domain of computational pathology. To address this, we present an interactiv...
["Computational Pathology", "Generative Adversarial Networks", "Tissue Image Synthesis"]
Synthesis of Annotated Colorectal Cancer Tissue Imagesfrom Gland LayoutEditor: UnnamedAbstractGenerating annotated pairs of realistic tissue images along with their annotations is a chal-lenging task in computational histopathology. Such synthetic images and their annotationscan be useful in training and evaluation of ...
_ebTQF9AE_o
The authors of this paper propose a method to generate synthetic colorectal tissue images with glands.
5: Marginally below acceptance threshold
The authors of this paper propose a method to generate synthetic colorectal tissue images with glands. I like the main idea of generating synthetic images interactively to train and evaluate algorithms. But the paper could be strengthened by describing more details. Please see my comments below. 1. The authors introduc...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Synthesis of Annotated Colorectal Cancer Tissue Images from Gland Layout ### Paper Abstract Generating annotated pairs of realistic tissue images along with their annota...
0h9cYBqucS6
ICLR.cc/2021/Conference
2021
Communication-Computation Efficient Secure Aggregation for Federated Learning
["Beongjun Choi", "Jy-yong Sohn", "Dong-Jun Han", "Jaekyun Moon"]
Federated learning has been spotlighted as a way to train neural network models using data distributed over multiple clients without a need to share private data. Unfortunately, however, it has been shown that data privacy could not be fully guaranteed as adversaries may be able to extract certain information on local ...
["Federated Learning", "Privacy", "Graphs", "Secure Aggregation", "Communication-Efficient", "Computation-Efficient"]
ABSTRACTFederated learning has been spotlighted as a way to train neural network modelsusing data distributed over multiple clients without a need to share private data.Unfortunately, however, it has been shown that data privacy could not be fullyguaranteed as adversaries may be able to extract certain information on l...
EAiwZ9i1fjx
An Ok paper. Unrealistic assumptions.
4: Ok but not good enough - rejection
This paper proposes a secure (or private) algorithm that is communication and computationally efficient for the purpose of federated learning. The main idea is to allow each client to share its public keys and shares them secretly with a subset of its clients. This is a very natural idea. The main theoretical contribut...
3: The reviewer is fairly confident that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Communication-Computation Efficient Secure Aggregation for Federated Learning ### Paper Abstract Federated learning has been spotlighted as a way to train neural network...
ryT9R3Yxe
ICLR.cc/2017/conference
2017
Generative Paragraph Vector
["Ruqing Zhang", "Jiafeng Guo", "Yanyan Lan", "Jun Xu", "Xueqi Cheng"]
The recently introduced Paragraph Vector is an efficient method for learning high-quality distributed representations for pieces of texts. However, an inherent limitation of Paragraph Vector is lack of ability to infer distributed representations for texts outside of the training set. To tackle this problem, we introdu...
["Natural language processing", "Deep learning", "Unsupervised Learning", "Supervised Learning"]
ABSTRACTThe recently introduced Paragraph Vector is an efficient method for learning high-quality distributed representations for pieces of texts. However, an inherent lim-itation of Paragraph Vector is lack of ability to infer distributed representationsfor texts outside of the training set. To tackle this problem, we...
B1_E5k84e
Not convincing
3: Clear rejection
It feels that this paper is structured around a shortcoming of the original paragraph vectors paper, namely an alleged inability to infer representation for text outside of the training data. I am reasonably sure that this is not the case. Unfortunately on that basis, the premise for the work presented here no longer h...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Generative Paragraph Vector ### Paper Abstract The recently introduced Paragraph Vector is an efficient method for learning high-quality distributed representations for ...
BJgvg30ctX
ICLR.cc/2019/Conference
2019
Information Regularized Neural Networks
["Tianchen Zhao", "Dejiao Zhang", "Zeyu Sun", "Honglak Lee"]
We formulate an information-based optimization problem for supervised classification. For invertible neural networks, the control of these information terms is passed down to the latent features and parameter matrix in the last fully connected layer, given that mutual information is invariant under invertible map. We ...
["supervised classification", "information theory", "deep learning", "regularization"]
ABSTRACTWe formulate an information-based optimization problem for supervised classifi-cation. For invertible neural networks, the control of these information terms ispassed down to the latent features and parameter matrix in the last fully connectedlayer, given that mutual information is invariant under invertible ma...
B1l4KIw9nX
Theoretically grounded regularizer that penalizes confident predictions, experimental section needs to be improved
6: Marginally above acceptance threshold
The authors propose a regularizer placed on the final linear layer of invertible networks that penalizes confident predictions, leading to better generalization. The algorithm is theoretically grounded and even though SOTA networks do not meet some theoretical requirements in practice, it seems to be effective. The i...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Information Regularized Neural Networks ### Paper Abstract We formulate an information-based optimization problem for supervised classification. For invertible neural ne...
HDYMjiukjn
robot-learning.org/CoRL/2023/Conference
2023
RoboPianist: Dexterous Piano Playing with Deep Reinforcement Learning
["Kevin Zakka", "Philipp Wu", "Laura Smith", "Nimrod Gileadi", "Taylor Howell", "Xue Bin Peng", "Sumeet Singh", "Yuval Tassa", "Pete Florence", "Andy Zeng", "Pieter Abbeel"]
Replicating human-like dexterity in robot hands represents one of the largest open problems in robotics. Reinforcement learning is a promising approach that has achieved impressive progress in the last few years; however, the class of problems it has typically addressed corresponds to a rather narrow definition of dext...
["high-dimensional control", "bi-manual dexterity"]
ROBOPIANIST : Dexterous Piano Playing with DeepReinforcement LearningKevin Zakkaβ, δ, Philipp Wuβ, Laura Smithβ,Nimrod Gileadiδ, Taylor Howellσ, Xue Bin Pengψ, Sumeet Singhδ, Yuval Tassaδ,Pete Florenceδ, Andy Zengδ, Pieter AbbeelββUC Berkeley,δGoogle DeepMind,σStanford University,ψSimon Fraser UniversityCorrespondence ...
MSB7b5NuCG
### Strengths - The paper is looking at a potentially profoundly interesting problem (i.e., robot pianist), which can go beyond just the topic of dexterous control but more deeply into the space of AI creativity and music performing arts. For example, can a robot play classical pieces in idiosyncratic but acoustically ...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title RoboPianist: Dexterous Piano Playing with Deep Reinforcement Learning ### Paper Abstract Replicating human-like dexterity in robot hands represents one of the largest op...
ByxduJBtPB
ICLR.cc/2020/Conference
2020
When Covariate-shifted Data Augmentation Increases Test Error And How to Fix It
["Sang Michael Xie*", "Aditi Raghunathan*", "Fanny Yang", "John C. Duchi", "Percy Liang"]
Empirically, data augmentation sometimes improves and sometimes hurts test error, even when only adding points with labels from the true conditional distribution that the hypothesis class is expressive enough to fit. In this paper, we provide precise conditions under which data augmentation hurts test accuracy for min...
["data augmentation", "adversarial training", "interpolation", "overparameterized"]
ABSTRACTWe study covariate-shifted data augmentation where the augmented targets aredrawn from the true predictive distribution but the inputs are shifted. Empirically,some forms of data augmentation such as adversarial training improve robustness,but increase test error. We provide precise conditions under which data ...
r1eQCCNUYH
Official Blind Review #3
3: Weak Reject
This paper studies data augmentation in the regime where labels for the augmented datapoints are known. Special emphasis is put on the study of overparametrised linear models with minimum Euclidean norm of the regression weights (a.k.a. ridgeless regression). The results of this study are then used to motivate their “X...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title When Covariate-shifted Data Augmentation Increases Test Error And How to Fix It ### Paper Abstract Empirically, data augmentation sometimes improves and sometimes hurts ...
SJxE3jlA-
ICLR.cc/2018/Conference
2018
Now I Remember! Episodic Memory For Reinforcement Learning
["Ricky Loynd", "Matthew Hausknecht", "Lihong Li", "Li Deng"]
Humans rely on episodic memory constantly, in remembering the name of someone they met 10 minutes ago, the plot of a movie as it unfolds, or where they parked the car. Endowing reinforcement learning agents with episodic memory is a key step on the path toward replicating human-like general intelligence. We analyze why...
["Reinforcement learning", "Deep learning", "Episodic memory"]
ABSTRACTHumans rely on episodic memory constantly, in remembering the name of some-one they met 10 minutes ago, the plot of a movie as it unfolds, or where theyparked the car. Endowing reinforcement learning agents with episodic memory isa key step on the path toward replicating human-like general intelligence. We an-a...
r1gpVv_gM
Interesting similarity function, but insufficient evidence of generality
4: Ok but not good enough - rejection
There are a number of attempts to add episodic memory to RL agents. A common approach is to use some sort of recurrent model with a model-free agent. This work follows this approach using what could be considered a memory network with a identity embedding function and tests on 'Concentration', a game which requires mat...
5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Now I Remember! Episodic Memory For Reinforcement Learning ### Paper Abstract Humans rely on episodic memory constantly, in remembering the name of someone they met 10 m...
Bklfsi0cKm
ICLR.cc/2019/Conference
2019
Deep Convolutional Networks as shallow Gaussian Processes
["Adri\u00e0 Garriga-Alonso", "Carl Edward Rasmussen", "Laurence Aitchison"]
We show that the output of a (residual) CNN with an appropriate prior over the weights and biases is a GP in the limit of infinitely many convolutional filters, extending similar results for dense networks. For a CNN, the equivalent kernel can be computed exactly and, unlike "deep kernels", has very few parameters: onl...
["Gaussian process", "CNN", "ResNet", "Bayesian"]
ABSTRACTWe show that the output of a (residual) convolutional neural network (CNN) withan appropriate prior over the weights and biases is a Gaussian process (GP) in thelimit of infinitely many convolutional filters, extending similar results for densenetworks. For a CNN, the equivalent kernel can be computed exactly a...
HJxvPo8t3m
An efficient convolutional kernel for GPs without proper evaluation of probabilistic predictions
5: Marginally below acceptance threshold
This paper shows that deep convolutional networks (CNNs, without pooling) with a suitable prior over weights can be seen as shallow Gaussian processes (GPs) with a specific covariance function. It shows that this covariance function can be computed efficiently (when compared to previous attempts at resembling convoluti...
4: The reviewer is confident but not absolutely certain that the evaluation is correct
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Deep Convolutional Networks as shallow Gaussian Processes ### Paper Abstract We show that the output of a (residual) CNN with an appropriate prior over the weights and b...
HygQ7TNtPr
ICLR.cc/2020/Conference
2020
Rethinking Neural Network Quantization
["Qing Jin", "Linjie Yang", "Zhenyu Liao"]
Quantization reduces computation costs of neural networks but suffers from performance degeneration. Is this accuracy drop due to the reduced capacity, or inefficient training during the quantization procedure? After looking into the gradient propagation process of neural networks by viewing the weights and intermediat...
["Deep Learning", "Convolutional Network", "Network Quantization", "Efficient Learning"]
ABSTRACTQuantization reduces computation costs of neural networks but suffers from per-formance degeneration. Is this accuracy drop due to the reduced capacity, or in-efficient training during the quantization procedure? After looking into the gra-dient propagation process of neural networks by viewing the weights and ...
HkgNx-OiqB
Official Blind Review #1
1: Reject
Quick Summary: Based on semi-quantitative analysis the paper first proposes two rules for quantization of DNNs, then extends previous methods based on these rules to propose specific technique for quantizing activations and weights. Experimental results are provided on MobileNet V1/VB2 and ResNet50 and compare favourab...
<|im_start|>system You are a helpful assistant that reviews papers and provides feedback on their quality.<|im_end|> <|im_start|>user ### Paper Title Rethinking Neural Network Quantization ### Paper Abstract Quantization reduces computation costs of neural networks but suffers from performance degeneration. Is this ac...