Datasets:
anonymousatom
/
SFT15KQ_OCR

Dataset card Data Studio Files
xet
 Community
1
Paper_ID
stringlengths
10
10
Question
stringlengths
201
1.81k
ocr_output
stringlengths
252
54k
⌀
488A64eOf6
The method requires sampling in order to approximate the new parametric distribution (via the WIS algorithm). The runtime of this algorithm, and hence the additional computational complexity incurred by this decoding method, is not discussed. Further, there is an additional sampling + renormalization step that must hap...
LANGUAGE MODEL DECODING AS DIRECT METRICS OPTIMIZATION Haozhe Ji Pei Ke∗ Hongning Wang Minlie Huang∗ The CoAI Group, DCST, BNRist, Tsinghua University, Beijing 100084, China [email protected] [email protected] ABSTRACT Despite the remarkable advances in language modeling, current mainstream decodi...
EriR6Ec69a
Shouldn't a higher spectral norm contribute to higher dimensional dynamics? For a low spectral norm the network would quickly collapse onto a low-dimensional manifold. How do you explain then the RNNs have higher spectral norm and lower dimensionality in their dynamics?
LEVERAGING LOW-RANK AND SPARSE RECURRENT CONNECTIVITY FOR ROBUST CLOSED-LOOP CONTROL Neehal Tumma¹ Mathias Lechner² Noel Loo² Ramin Hasani² Daniela Rus² ¹Harvard University ²MIT CSAIL ABSTRACT Developing autonomous agents that can interact with changing environments is an open challenge in machine learning. Robustne...
79tJB1eTmb
Comparison with the gpt-3.5-turbo-based method (AutoCot) shows that the gap is 0.5 points. Comparing this with the data requirements that the proposed method imposes brings the approach's utility into question.
Meta-CoT: Generalizable Chain-of-Thought Prompting in Mixed-task Scenarios with Large Language Models Anonymous authors Paper under double-blind review Abstract Large language models (LLMs) have unveiled remarkable reasoning capabilities by exploiting chain-of-thought (CoT) prompting, which generates intermediate re...
ISq7Hnln0t
According to last paragraph in section 3, during evaluation, point prompts are used by default for quantitative evaluations. Is the UAP created under point prompts still effective under other types of SAM prompts? We can never assume the user to stick at one single prompt type.
SEGMENT ANYTHING MEETS UNIVERSAL ADVERSARIAL PERTURBATION Anonymous authors Paper under double-blind review ABSTRACT As Segment Anything Model (SAM) becomes a popular foundation model in computer vision, its adversarial robustness has become a concern that cannot be ignored. This work investigates whether it is poss...
v8eWha27jw
In Section 3.3 “distribution-aware unbiased quantization”, this work proposes two optimization problems to find the optimal quantization values to reduce NMSE. In the first optimization problem on page 4, the notations $S(z, x)$ and $R(x)$ are a bit confusing. Are $S$ and $R$ two functions to be optimized?
ABSTRACT Distributed Mean Estimation (DME), in which \( n \) clients communicate vectors to a parameter server that estimates their average, is a fundamental building block in communication-efficient federated learning. In this paper, we improve on previous DME techniques that achieve the optimal \( O(1/n) \) Normaliz...
5T46w5X3Go
The major concern I have is that the paper can separate the common features and task-specific features explicitly. In other words, in Option A and Option B, we can directly know that $w$ is a common feature, while $q$ is not. In practice, we do not have any idea about which part of the feature should be useful and we s...
THEORETICAL ANALYSIS ON THE GENERALIZATION POWER OF OVERFITTED TRANSFER LEARNING Anonymous authors Paper under double-blind review ABSTRACT Transfer learning is a useful technique for achieving improved performance and reducing training costs by leveraging the knowledge gained from source tasks and applying it to ta...
Fq8tKtjACC
In Table 2, it looks like for all prior models as well as phi-1-base (the model without finetuning), there is a significant gap between the new score and the HumanEval one. However, in both finetunened phi-1 models this gaps is removed. Is it not possible that this means that while the finetuning data may be unrelated ...
TEXTBOOKS ARE ALL YOU NEED Anonymous authors Paper under double-blind review ABSTRACT We introduce phi-1, a new large language model for code, with significantly smaller size than competing models: phi-1 is a Transformer-based model with 1.3B parameters, trained for 4 days on 8 A100s, using a selection of “textbook ...
PFdjJiZjPj
I do not understand what the following statement means: “the test cases generated by LLMs can show a descent pass rate, and this pass rate is even higher than the code pass rate on HumanEval+, which holds for both large and small LLMs.”
THE PROGRAM TESTING ABILITY OF LARGE LANGUAGE MODELS FOR CODE Anonymous authors Paper under double-blind review ABSTRACT Recent development of large language models (LLMs) for code like CodeX and CodeT5+ demonstrates tremendous promise in achieving code intelligence. Their ability of synthesizing code that completes...
x2rZGCbRRd
Another thought concerns use of the proposed DAG. Post-treatment variables are presumably omnipresent, and it can be difficult to know how to proceed when they are around (motivating this paper). For applicability in practice, I do not know when or whether investigators would be willing to assume the proposed graph in ...
EXTRACTING POST-TREATMENT COVARIATES FOR HETEROGENEOUS TREATMENT EFFECT ESTIMATION Anonymous authors Paper under double-blind review ABSTRACT The exploration of causal relationships between treatments and outcomes, and the estimating causal effects from observational data, have garnered considerable interest in the ...
EHKS0oXuku
In the experiments, are you evaluating deterministic predictions using the mean parameters of the variational distribution, or are you using BMA through sampling from the variational distribution? In the context of BNNs, performance evaluation with BMA is necessary.
JENSEN-SHANNON DIVERGENCE BASED NOVEL LOSS FUNCTIONS FOR BAYESIAN NEURAL NETWORKS Anonymous authors Paper under double-blind review ABSTRACT We aim to overcome the limitations of Kullback-Leibler (KL) divergence-based variational inference (VI) used in Bayesian Neural Networks (BNNs), which stem from the lack of bou...
4SrzKsJocx
And the measure RC_0 is introduced because the ideal uncorrelation is not achievable if the sample is few. But RC_0 is computed based on multiple random trials. That is to say, the evaluation metric is not deterministic.
SIMULTANEOUS DIMENSIONALITY REDUCTION: A DATA EFFICIENT APPROACH FOR MULTIMODAL REPRESENTATIONS LEARNING Anonymous authors Paper under double-blind review ABSTRACT Current experiments frequently produce high-dimensional, multimodal datasets—such as those combining neural activity and animal behavior or gene expressi...
bTMMNT7IdW
According to my understanding, Eq.(7) aims at learn a set of parameters for $f_k$ and $g_k$ so that the learned models can mimic the approximated trajectory of data. In my opinion, such operation relies heavily on the quality of data representations, especicially at the early phase of training, since the feature extrac...
Latent Trajectory Learning for Limited Timestamps under Distribution Shift over Time Qiuhao Zeng¹, Changjian Shui², Long-Kai Huang, Peng Liu³, Xi Chen⁴ Charles X. Ling¹, Boyu Wang¹∗ ¹University of Western Ontario ²Vector Institute ³University of Toronto ⁴Noah’s Ark Lab Abstract Distribution shifts over time are comm...
PdTe8S0Mkl
It was mentioned that the use of Roget’s thesaurus is to map words to related categories for a thematic-style analysis. But why are other, more common forms of thematic analysis not explored such as LDA, BERTopic, or Contextual Topic Models? These might even give better results on the differences between theme use as e...
Humans vs ChatGPT: Uncovering Non-trivial Distinctions by Evaluating Parallel Responses Anonymous authors Paper under double-blind review Abstract The advent of ChatGPT and similar Large Language Models has set the world in an uproar as it is able to generate human-like natural language. Due to the high similarity b...
Sy8upuD6Bw
- Page 5: - Would it be possible to give a sense of what the correlation between the presence of the noise token and the feedback request from the receiver is? It's not necessary for all of the experiments, maybe just the initial basic ones.
Emergent Communication with Conversational Repair Mitja Nikolaus * CerCo, CNRS [email protected] Abstract Research on conversation has put emphasis on the importance of a multi-level communication system, in which the interlocutors aim to establish and maintain common ground. In natural conversations, repair me...
KXOB15k1br
What is the significance of the improvements of TSAA when presenting the results with a critical diagram that includes a statistical test (Demšar, JMLR 2006, https://jmlr.org/papers/v7/demsar06a.html)?
TIME-SERIES AUTOAUGMENT: DATA AUGMENTATION POLICY SEARCH FOR LONG-TERM FORECASTING Anonymous authors Paper under double-blind review ABSTRACT Data augmentation is a popular regularization for addressing overfitting issues of neural networks. Recently, automatic augmentation showed strong results on image classificat...
l1U6sEgYkb
- It appears that the 3D DV deformable attention mechanism lifts PV features to 3D using known camera parameters. Also, when the authors concatenate DV features, they depend on these camera parameters. Have the authors attempted to test the tolerance of calibration parameters to noise?
DV-3DLane: End-to-end Multi-modal 3D Lane Detection with Dual-view Representation Yueru Luo\textsuperscript{1,2}, Shuguang Cui\textsuperscript{2,1}, Zhen Li\textsuperscript{2,1*} \textsuperscript{1} FNii, CUHK-Shenzhen \textsuperscript{2} School of Science and Engineering, CUHK-Shenzhen \{222010057@link., shuguangcui@...
oTRwljRgiv
What is the comparison in FLOPs between the ExeDec models and the baseline models (including the beam search, etc)? Even if the parameter count is the same, the loop in ExeDec could give these models significantly more power than the Transformer / Latent Programmer baselines.
ExeDec: Execution Decomposition for Compositional Generalization in Neural Program Synthesis Kensen Shi Google DeepMind [email protected] Joey Hong * UC Berkeley [email protected] Yinlin Deng * University of Illinois Urbana-Champaign [email protected] Pengcheng Yin Google DeepMind pcyin@googl...
fH2wf2w2Ss
Is it that unconditional sampling of CLIP image embeddings is somehow important or easier than sampling an image directly? Or is it the two-stage pipeline itself that is the important part? Could the condition generation and subsequent image generation be done in a single pipeline with end-to-end training? What exactly...
Two-Stage Diffusion Models: Better Image Synthesis by Explicitly Modeling Semantics Anonymous authors Paper under double-blind review Abstract Recent progress with conditional image diffusion models has been stunning, and this holds true whether we are speaking about models conditioned on a text description, a scene...
V3j5d0GQgH
Why the mean and variance of class means are evaluated in Figure 1? I cannot follow the logic in the discussion quoted below. > However, preliminary experiments benchmarking ETF with CIFAR-100 in Fed-LT suggest that only a few features have relatively large means, while most of the small-mean features are contaminated ...
FedLoGe: Joint Local and Generic Federated Learning under Long-tailed Data Zikai Xiao1∗, Zihan Chen2∗, Liyinglan Liu3, Yang Feng4, Jian Wu1, Wanlu Liu1, Joey Tianyi Zhou5,6, Howard Hao Yang1, Zuozhu Liu1† 1Zhejiang University, 2Singapore University of Technology and Design, 3University of Electronic Science and Techno...
btpgDo4u4j
The paper repeatedly uses terms like the latent action representation and latent action planning, without a carefully derived definition of it. For self consistency, it would be helpful to define these terms more concretely; otherwise, in its current format, the contributions of the paper can be hard to follow.
Efficient Planning with Latent Diffusion Wenhao Li School of Software Engineering, Tongji University Shanghai, 201804, China [email protected] Abstract Temporal abstraction and efficient planning pose significant challenges in offline reinforcement learning, mainly when dealing with domains that involve temporall...
QXCjvHnDmu
how is the success rate computed? the paper mentions that the attack succeeds if the target string is produced exactly. I imagine this has some limitations. Examples in fig. 4 do not show the output starting from
Open Sesame! Universal Black Box Jailbreaking of Large Language Models Anonymous authors Paper under double-blind review This paper contains unfiltered, possibly offensive content generated by LLMs Abstract Large language models (LLMs), designed to provide helpful and safe responses, often rely on alignment techniq...
gENfMmUIkT
Since the approach is worked on resource-constrained IoT devices, to compare, the authors did not bring any such metric (For instance, FLOPs, Parameters, Energy Consumptions) to compute the computational power needed by the models.
A PIPELINE-BASED APPROACH FOR OBJECT DETECTION ON RESOURCE CONSTRAINED INTERNET OF THINGS DEVICES Anonymous authors Paper under double-blind review ABSTRACT Object detection with computer vision and convolutional neural networks on resources constrained devices can be challenging. The limited power and processing ca...
cZo6pDtDZr
What specific properties of the hash family do you need? The reason I ask is that communicating a random hash function requires a large number of bits; however, this can be reduced drastically if one can settle for $\ell$-wise independence in the analysis using standard techniques from the pseudorandomness literature (...
NEAR-OPTIMAL ALGORITHMS FOR PRIVATE ESTIMATION AND SEQUENTIAL TESTING OF COLLISION PROBABILITY Anonymous authors Paper under double-blind review ABSTRACT We present new algorithms for estimating and testing collision probability, a fundamental measure of the spread of a discrete distribution that is widely used in m...
KbDzdqevfV
Bellman equations to characterise safety and equivalently reachability properties have been already developed in the literature, and also in the stochastic setting [1,2]. The one proposed by the authors is an extension of those ones, with the added constraint that only actions that have probability 1 of being safe are ...
CORRECT-BY-DESIGN SAFETY CRITICS USING NON-CONTRACTIVE BINARY BELLMAN OPERATORS Anonymous authors Paper under double-blind review ABSTRACT The inability to naturally enforce safety in Reinforcement Learning (RL), with limited failures, is a core challenge impeding its use in real-world applications. One notion of sa...
ORUiqcLpV6
The proposed idea which pays attention to not just the target object, has been proposed before in “PhraseRefer”. They use human annotators to obtain fine grained annotation. Which share the same underlying idea with this paper.
CoT3DRef: Chain-of-Thoughts Data-Efficient 3D Visual Grounding Eslam Mohamed Bakr, Mohamed Ayman, Mahmoud Ahmed, Habib Slim, Mohamed Elhoseiny King Abdullah University of Science and Technology (KAUST) {eslam.abdelrahman, mohamed.mohamed.2, mahmoud.ahmed, habib.slim, mohamed.elhoseiny}@kaust.edu.sa Abstract 3D visua...
0IaTFNJner
In section 3, the paper proposes Information Abundance to measure the degree of collapse of embedding matrices. As the paper focuses on the scaling law of embedding layers, the paper should discuss whether Information Abundance is a fair metric when comparing embedding matrices of different dimension sizes.
On the Embedding Collapse When Scaling up Recommendation Models Anonymous authors Paper under double-blind review Abstract Recent advances in deep foundation models have led to a promising trend of developing large recommendation models to leverage vast amounts of available data. However, we experiment to scale up e...
sxGugrYhP9
While BatteryML puts a lot of effort into standardizing publicly available datasets and making them accessible to work with, for the wider community, I'm curious how much of the detailed battery science gets communicated to the end-user, or whether the layer of abstraction necessary to standardize the framework obfusca...
BatteryML: An Open-source Platform for Machine Learning on Battery Degradation Han Zhang\textsuperscript{1,*}, Xiaofan Gui\textsuperscript{2}, Shun Zheng\textsuperscript{2}, Ziheng Lu\textsuperscript{2}, Yuqi Li\textsuperscript{3}\textsuperscript{*}, Jiang Bian\textsuperscript{2} \textsuperscript{1}Institute for Inter...
Q1vkAhdI6j
The mechanics of the “clicking annotation” remain somewhat elusive. How were the annotators guided in executing this task? Was there a specific strategy adopted for different instance types? Was this manual labeling extended to all point clouds during training? If not, how were the crucial clicking points discerned? Th...
MixSup: Mixed-Grained Supervision for Label-Efficient LiDAR-Based 3D Object Detection Yuxue Yang1,2,3,5 Lue Fan2,3,5† Zhaoxiang Zhang1,2,3,4,5† 1School of Artificial Intelligence, UCAS 2University of Chinese Academy of Sciences (UCAS) 3Institute of Automation, Chinese Academy of Sciences (CASIA) 4Centre for Artificial...
ikwEDva1JZ
**Why are the constructive proofs important for understanding in-context learning in transformers?** I am aware that there are prior works that design transformers that are capable of in-context learning. However, I am not convinced of the importance and significance of these results. Couldn't we also find weights for ...
How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations Tianyu Guo¹ Wei Hu² Song Mei¹ Huan Wang³ Caiming Xiong³ Silvio Savarese³ Yu Bai³ ¹UC Berkeley ²University of Michigan ³Salesforce AI Research [email protected] Abstract While large language models based ...
zhZXk5Ctz2
The widely used dual-domain loss in models, such as MIMOUNet (Cho et al, ICCV'21) and SFNet (Cui et al, ICLR'23), can introduce global information refinement. How does aRGB compare to this loss function? This function does not lead to much computation overhead.
Rethinking RGB Color Representation for Image Restoration Models Anonymous authors Paper under double-blind review Abstract The per-pixel distance loss defined in the RGB color domain has been almost a compulsory choice for training image restoration models, despite its well-known tendency to guide the model to prod...
DluJpvRF69
In equation(3), the activation of s^t_m includes phi^k_m and phi^t_m, it looks like a recursive definition, then the activation of s^k_m should include phi from its nearest model and the phi for itself, and so on following the recursive rules. Why all the historical nearest models are omitted in equation (3)?
StyleCL: Latent Dictionary Learning for StyleGAN without Forgetting Anonymous authors Paper under double-blind review Abstract StyleGAN is one of the most versatile generative models that have emerged in recent times. However, when it is trained continually on a stream of data (potentially previously unseen distribu...
zEkvV65Wi1
If setting 2 is adopted, the comparison between the proposed method and the baselines (e.g., KD with Mixup vs Mixup) is not fair, since the hyperparameters of baselines are not optimized in terms of the calibration error.
Understanding Calibration Transfer in Knowledge Distillation Anonymous authors Paper under double-blind review Abstract Modern deep neural networks are often miscalibrated, leading to overconfident mistakes that erode their reliability, and limit their use in critical applications. The existing confidence calibratio...
YLJs4mKJCF
One of the problem of introducing elastic penalty is how to choose properly the two penalty parameters $\lambda_{1}$ and $\lambda_2$. Though it can be chosen empirically, it can be dataset-dependent. Would it make significantly difference if we simply choose the L1 norm penalty instead?
Towards Poisoning Fair Representations Tianci Liu1, Haoyu Wang1, Feijie Wu1, Hengtong Zhang2, Pan Li3, Lu Su1, Jing Gao1 1Purdue University 2Tencent AI Lab 3Georgia Institute of Technology 1{liu3351,wang5346,wu1977,lusu,jinggao}@purdue.edu [email protected] [email protected] Abstract Fair machine learning seek...
sW95puhphh
While the paper addresses the challenges of policy discoordination and privacy concerns, how does the proposed method handle issues related to credit assignment, especially when agents have conflicting goals or when their contributions to the global reward are imbalanced?
DECENTRALIZED MULTI-AGENT REINFORCEMENT LEARNING VIA ANTICIPATION SHARING Anonymous authors Paper under double-blind review ABSTRACT Centralized multi-agent reinforcement learning requires global policy access and coordination, often infeasible in decentralized applications. A challenge in decentralized MARL with in...
hp4yOjhwTs
The authors implement ALP-GMM by fixing the color. I don't think this is a good way to construct baseline. One can run ALP-GMM as a task sampler on a fixed set of tasks. ALP-GMM essentially assigns a probability for different tasks at the each round. If one task has the underlying U that is not aligned with the target ...
Causally Aligned Curriculum Learning Mingxuan Li and Junzhe Zhang and Elias Barcinboim Causal Artificial Intelligence Lab, Columbia University, USA {ml,junzhez,eb}@cs.columbia.edu Abstract A pervasive challenge in Reinforcement Learning (RL) is the “curse of dimensionality” which is the exponential growth in the sta...
OkHHJcMroY
I am also a bit surprised by the fact that the mixing time of the original chain $(s_1,a_1,s_2,\ldots)$ does not pop up explicitly in the bounds. This would be a typical behavior for the optimization problems with dependent data. What is the explanation?
PILOT: AN $O(1/K)$-CONVERGENT APPROACH FOR POLICY EVALUATION WITH NONLINEAR FUNCTION APPROXIMATION Zhuqing Liu†, Xin Zhang‡, Jia Liu†, Zhengyuan Zhu‡, Songtao Lu∗ †Department of Electrical and Computer Engineering, The Ohio State University ‡Department of Statistics, Iowa State University ∗IBM Research, IBM Thomas J. ...
R1crLHQ4kf
Moreover, can the adversarial optimization problem be formulated to reduce divergence from the benign data distribution, while still fooling the ASR system? What are the challenges in constructing such
LEVERAGING CHARACTERISTICS OF THE OUTPUT DISTRIBUTION FOR IDENTIFYING ADVERSARIAL AUDIO EXAMPLES Anonymous authors Paper under double-blind review ABSTRACT Adversarial attacks can mislead automatic speech recognition (ASR) systems into producing an arbitrary desired output. This is easily achieved by adding impercep...
SQFDJLyJNB
Given that a Gaussian Mixture Model (GMM) is employed, does the assumption of similarity between learning stages in conventional continual learning scenarios, which may vary greatly, potentially limit the model's capabilities?
PROMPTCCD: LEARNING GAUSSIAN MIXTURE PROMPT POOL FOR CONTINUAL CATEGORY DISCOVERY Anonymous authors Paper under double-blind review ABSTRACT In this paper, we address the challenging open-world learning problem of continual category discovery (CCD). Initially, a labelled dataset consisting of known categories is pro...
wM01y5BPM9
Beyond knowing where the object is (for which one could also use object detection in general, for this simplified case with fixed background the computer vision literature is also providing various method for background subtraction), in which way are these representations interpretable? It seems to me that these repres...
IDENTIFIABLE REPRESENTATION LEARNING VIA ARCHITECTURE EQUIVARIANCES Anonymous authors Paper under double-blind review ABSTRACT Despite their immense success and usefulness, current deep learning systems are still lacking in interpretability, robustness, and out of distribution generalisation. In this work we propose...
z7K2faBrDG
Following the above comment, the proposed method would be unable to give a measure to predict the Mean Opinion Score (MOS) on distortion in databases such as TID [Ponomarenko et al. 13] or KADID [Lin et al. 19]. If this is not the case, the authors should mention how to infer this MOS.
Perceptual Scales Predicted by Fisher Information Metrics Jonathan Vacher∗ MAP5, Université Paris Cité, CNRS, F-75006, Paris, France [email protected] Pascal Mamassian LSP, Département d’études cognitives, École normale supérieure, PSL University, CNRS, 75005 Paris, France [email protected] Abstract ...
MhzKwuvpm6
`This reward structure allows us to utilize any single-agent reinforcement learning algorithm, instead of using supervised learning to optimize over loss functions defined in Equations 11 and 12.` What is the difference to GAIL? GAIL also allows any single-agent reinforcement learning algorithm. And why should we condu...
RILE: Reinforced Imitation Learning Anonymous authors Paper under double-blind review Abstract Learning to imitate behaviors from a limited set of expert trajectories is a promising way to acquire a policy. In imitation learning (IL), an expert policy is trained directly from data in a computationally efficient way,...
V8aD5pUcVX
The main concern is that the pretraining image-text pairs contain a significant amount of object-level annotations. This leads an unfair comparison when the downstream tasks are most object-centric questions.
What Makes for Good Visual Tokenizer Supervision for Large Language Models? Anonymous authors Paper under double-blind review Abstract We empirically investigate proper pre-training supervision to build good visual tokenizers, making Large Language Models (LLMs) powerful Multimodal Large Language Models (MLLMs). In ...
ekz1hN5QNh
In the re-scaling procedure, you assume that the variance direction is along the geodesic intersecting the origin, however, this may not be the case, therefore is not an accurate formulation. Can you elaborate if I have mis-understood.
Fully Hyperbolic Convolutional Neural Networks for Computer Vision Ahmad Bdeir\textsuperscript{1,*}, Kristian Schwethelm\textsuperscript{2,*,\dagger} & Niels Landwehr\textsuperscript{1} \textsuperscript{1} Data Science Department, University of Hildesheim \textsuperscript{2} Chair for Artificial Intelligence in Medici...
Nu7dDaVF5a
The model is not scale/rotation/translation invariant. I think the main reason is the use of point position as input to the decoder, which means if the coordinates system is changed, the output of the decoder is also changed. Similarly, the surface normal or view directions should also be in the local coordinates syste...
3D Reconstruction with Generalizable Neural Fields Using Scene Priors Yang Fu† Shalini De Mello Xueting Li Amey Kulkarni Jan Kautz Xiaolong Wang Sifei Liu 1University of California, San Diego 2NVIDIA Abstract High-fidelity 3D scene reconstruction has been substantially advanced by recent progress in neural fields. H...
WXXuORQwbQ
In Figure 6 and Table 4. It seems the model gets worse when using 50 or 10 masks which is strange.Why would using 3 masks be better than 10 or 50 masks? Why would it be even better than using the full dense tensor?
Sparse Mask Representation for Human-Scene Interaction Anonymous authors Paper under double-blind review Abstract Human-scene interaction is an active research topic with several applications in robotics, virtual experiences, gaming, surveillance, and healthcare. Despite efforts to improve the network architectures ...
6vtGG0WMne
At 500:1 imbalance ratios, multiple benchmark methods show numbers decimated to zero. A previous study, MMM, by Mirza et al. '21 appears to suggest that even at such a ratio, classical resampling and even baselines report an above zero performance. Could you reason about the disparity?
REGULATING IMBALANCED DEEP MODELS WITH USER-SPECIFIED METRICS Anonymous authors Paper under double-blind review ABSTRACT Deep learning models implemented in real-world applications still face challenges from imbalanced data. Existing methods address the imbalance problem by balancing the models between the minority ...
Va4t6R8cGG
How fast is the proposed method compared to other methods? I understand that there are some GFLOPs comparisons in the supplementary, but it is difficult to compare the methods due to the presence of other parts (such as LTC or person detector). Could we see a speed comparison instead?
END-TO-END SPATIO-TEMPORAL ACTION LOCALISATION WITH VIDEO TRANSFORMERS Anonymous authors Paper under double-blind review ABSTRACT The most performant spatio-temporal action localisation models use external person proposals and complex external memory banks. We propose a fully end-to-end, transformer based model that...
eO6lXIWyxn
There are so many metrics used for the evaluation (FID, CLIP-S, and OCR). Which one is the most appropriate to evaluate the overall performance? Or is there any way to combine all of them as a final metric?
SUBMISSION HAS BEEN WITHDRAWN Anonymous authors Paper under double-blind review ABSTRACT We thank the reviewers for their valuable comments. After careful consideration, we think our paper is inappropriate for ICLR and decided to withdraw our paper.
4pW8NL1UwH
Equation 3 seems to model generations as P_{\pi_{\theta}} (y^{i}_{j,k} | x^{i}) , but for autoregressive models that the authors study, the probability should be modelled as P_{\pi_{theta}}(y^{i}_{j,k} | x^{i}, y^{i}_{j, <k}), which in turn renders it intractable to compute
LIRE: LISTWISE REWARD ENHANCEMENT FOR PREFERENCE ALIGNMENT Anonymous authors Paper under double-blind review ABSTRACT Recently, tremendous strides have been made in the domain of Natural Language Generation (NLG) due to the vast advances in Large Language Models (LLMs). However, often trained on large-scale unsuperv...
Glcsog6zOe
The false negative is misleading. In the appendix, the error explanation says that the environment reports that there is no keyboard. Are keyboards part of the possible objects? The observations in the prompt mention a computer but no keyboard.
TREE-PLANNER: EFFICIENT CLOSE-LOOP TASK PLANNING WITH LARGE LANGUAGE MODELS Mengkang Hu ♦ Yao Mu ♦ Xinmiao Yu ♥ Mingyu Ding*♦ Shiguang Wu ◊ Wenqi Shao* Qiguang Chen♥ Bin Wang ♦ Yu Qiao* Ping Luo* ♦ ABSTRACT This paper studies close-loop task planning, which refers to the process of generating a sequence of skills (a...
wwotGBxtC3
I'm curious how making this approach multi-modal can be helpful. Could graph embeddings or vision embeddings of the molecules provide any benefit? I'm not a molecular properties expert, but I tried a couple of the figures (table 2 and figure 2) with GPT-4 vision, and it gave meaningful explanations. Have the authors in...
DATA-EFFICIENT MOLECULAR GENERATION WITH HIERARCHICAL TEXTUAL INVERSION Anonymous authors Paper under double-blind review ABSTRACT Developing an effective molecular generation framework even with a limited number of molecules is often important for its practical deployment, e.g., drug discovery, since acquiring task...
S3x7IcbwY8
In the method part, the authors put a lot of effort into introducing the overlap part with Pix2Seq V2, such as the tokenizer and masked modeling. However, the difference from Pix2Seq V2 is not well presented.
Masked AutoDecoder is Effective Multi-Task Vision Generalist Anonymous authors Paper under double-blind review Abstract Inspired by the success of general-purpose models in NLP, recent studies attempt to unify different vision tasks in the same sequence format and employ autoregressive Transformers for sequence pred...
AwyxtyMwaG
The text says 'We can then test the causal effect of an FV by adding it to hidden states at any layer ℓ as the model resolves a prompt and measuring its performance in executing the task.' I don't understand exactly how the vector is being 'added to hidden states at any layer'. Am I right that this vector is being adde...
Function Vectors in Large Language Models Eric Todd,* Millicent L. Li, Arnab Sen Sharma, Aaron Mueller, Byron C. Wallace, and David Bau Khoury College of Computer Sciences, Northeastern University Abstract We report the presence of a simple neural mechanism that represents an input-output function as a vector within...
ccxD4mtkTU
For instance, the paper by Veselovsky et al. [1] demonstrated that crowd workers are using LLMs to solve tasks, so I am wondering if the paper took any steps to ensure that the human evaluators solved the task on their own.
Can LLM-Generated Misinformation Be Detected? Canyu Chen Illinois Institute of Technology [email protected] Kai Shu Illinois Institute of Technology [email protected] Project website: https://llm-misinformation.github.io/ Abstract The advent of Large Language Models (LLMs) has made a transformative impact. ...
FoqZKsH9sE
In addition, it is unfair to compare the baseline in Table 1. The baseline in Table 1 is not training with the knowledge distillation, but the LSP models are fine-tuned with 300 epochs with knowledge distillation. And the accuracy of the distilled DeiT-Small and DeiT-Base should be 81.2% and 83.4%, respectively.
LSP: Low-Power Semi-Structured Pruning for Vision Transformers Anonymous authors Paper under double-blind review Abstract Vision transformers (ViTs) have emerged as a promising alternative to convolutional neural networks (CNNs) for various image analysis tasks, offering comparable or superior performance. However, ...
aZH1dM3GOX
How many experts are used for the Meta World evaluations? How does the number of experts impact the scalabiltiy of the proposed approach? Further elaborations on the computational overhead introduces, as well as possible limitations might be more insightful, than the presented remarks regarding interpretability.
Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts Ahmed Hendawy\textsuperscript{1,2}, Jan Peters\textsuperscript{1,2,3,4}, Carlo D’Eramo\textsuperscript{1,2,5} \textsuperscript{1}Department of Computer Science, TU Darmstadt, Germany \textsuperscript{2}Hessian Center for Artificial Intelligence (Hess...
UTLv72uDlS
A couple direct questions:- Is the matrix S actually being approximated at specific rows, or are entire rows being left out?- How is it more efficient to compute the gradient at a sampled point? Doesn’t this essentially require backpropagating through all time steps from the end of time to the beginning, regardless of ...
SCALING SAFE LEARNING-BASED CONTROL TO LONG-HORIZON TEMPORAL TASKS Anonymous authors Paper under double-blind review ABSTRACT This paper introduces a model-based approach for training parameterized policies for an autonomous agent operating in a highly nonlinear (albeit deterministic) environment. We desire the trai...
l3s3HwJYDm
In appendix A. 2, this paper uses a set of identical states to acquire the action vectors of the policy in the test set. What is the detailed process of obtaining the set of identical states? Are these states sampled by a certain policy?
OPPONENT MODELING BASED ON SUBGOAL INFERENCE Anonymous authors Paper under double-blind review ABSTRACT When an agent is in a multi-agent environment, it may face previously unseen opponents, and it is a challenge to cooperate with other agents to accomplish the task together or to maximize its own rewards. Most opp...
DjeQ39QoLQ
I am not convinced by the claim the S4-PTD model outperforms the S4D models on LRA. The LRU paper (https://arxiv.org/abs/2303.06349) reports results for S4D that are much better than the original reported S4D paper results. In addition, the appendix of the current paper under review states that mild hyperparameter tuni...
ROBUSTIFYING STATE-SPACE MODELS FOR LONG SEQUENCES VIA APPROXIMATE DIAGONALIZATION Annan Yu,1 Arnur Nigmatov,2 Dmitriy Morozov,2 Michael W. Mahoney,2,3,4 N. Benjamin Erichson2,3 1 Center for Applied Mathematics, Cornell University, Ithaca, NY 14853, USA 2 Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA ...
sCd7pHnXMG
I don't understand what this part is achieving. Why is a data poisoning algorithm optimizing weights in Equation (4)? Are you interfering with the training process? If that's the case, it is contradictory to your threat model in Section 2, where the attacker does not know the pre-training settings.
CORRUPTENCODER: DATA POISONING BASED BACKDOOR ATTACKS TO CONTRASTIVE LEARNING Anonymous authors Paper under double-blind review ABSTRACT Contrastive learning (CL) pre-trains general-purpose encoders using an unlabeled pre-training dataset, which consists of images or image-text pairs. CL is vulnerable to data poison...
wD8L86iCvD
In Table 2 and 3, some results are missing for the Video-LLAMA and are replaced with “-”. Why? As far as I understand, Video-LLAMA could be applied on all the tasks that the proposed method can be applied to by just removing the missing modality. One can just remove the audio or the visual branch in the Video-LLAMA and...
FINE-GRAINED AUDIO-VISUAL JOINT REPRESENTATIONS FOR MULTIMODAL LARGE LANGUAGE MODELS Anonymous authors Paper under double-blind review ABSTRACT Audio-visual large language models (LLM) have drawn significant attention, yet the fine-grained combination of both input streams is rather under-explored, which is challeng...
FMMF1a9ifL
In the Experiments section (Section 6), how much of the results are related to the selection of the molecules under study? Stated differently, how do the authors plan to address the generalizability of the approach?
GRADUAL OPTIMIZATION LEARNING FOR CONFORMATIONAL ENERGY MINIMIZATION Artem Tsypin\textsuperscript{1}\textsuperscript{\&}, Leonid Ugadiarov\textsuperscript{2,4}, Kuzma Khrabrov\textsuperscript{1}, Alexander Telepov\textsuperscript{1}, Egor Rumiantsev\textsuperscript{1}, Alexey Skrynnik\textsuperscript{1,2}, Aleksandr P...
MbfAK4s61A
”Impact of Fundamental Model”: GPT-4 has a higher unsafe rate than ChatGPT of smaller size. However, the trend does not work for Llama2 models (13B and 70B). How should we interpret the results? “GPT-4” was distinctively too smart to be safe? Can we generalize that the smarter llms is the unsafer?
GPT-4 IS TOO SMART TO BE SAFE: STEALTHY CHAT WITH LLMs VIA CIPHER WARNING: THIS PAPER CONTAINS UNSAFE MODEL RESPONSES. Youliang Yuan\textsuperscript{1,2,*} Wenxiang Jiao\textsuperscript{2} Wenxuan Wang\textsuperscript{2,3,*} Jen-tse Huang\textsuperscript{2,3,*} Pinjia He\textsuperscript{1†} Shuming Shi\textsuperscrip...
fcSDt7H8kI
What is the QRDQN algorithm baseline in Figure 5? It is not discussed in the paper. What is the difference between $\epsilon$-greedy in Figure 1 and Figure 5? While it is briefly mentioned in the footnote of the supplementary materials, detailed references are not presented.
Boosting Reinforcement Learning with Extremum Experiences Anonymous authors Paper under double-blind review Abstract Reinforcement learning research has achieved high acceleration in its progress starting from the initial installation of deep neural networks as function approximators to learn policies that make sequ...
lK0WxHeups
First math display in Section 1.3.2: this bound should mentioned somewhere that $b > n$ isn't possible and $b = n$ reduces to full-batch gradient descent. As a result, it is not always feasible to select the batch-size to minimize the oracle complexity.
Iteration and Stochastic First-order Oracle Complexities of Stochastic Gradient Descent using Constant and Decaying Learning Rates Anonymous authors Paper under double-blind review Abstract The performance of stochastic gradient descent (SGD), which is the simplest first-order optimizer for training deep neural netw...
w5oP27fmYW
Center of Mass: The authors opted to canonicalize the shapes using the point mean. In cases where the density is non-uniform, this approach could lead to issues. Specifically, very similar shapes that are sampled differently might end up having different
CCD-3DR: Consistent Conditioning in Diffusion for Single-Image 3D Reconstruction Anonymous authors Paper under double-blind review Abstract In this paper, we present a novel shape reconstruction method leveraging a diffusion model to generate a 3D sparse point cloud for the object captured in a single RGB image. Rec...
Sx7BIiPzys
Why was this specific subset of six UCI data sets chosen? The original work by Hernández-Lobato and Adams (2015), who introduced this set of experiments had ten, and even Watson et al. (2021) who the authors cite as relying on for their setup used ~~seven~~ different sets. _(PostRebuttal Edit: I misread the reference, ...
VARIATIONAL BAYESIAN LAST LAYERS James Harrison\textsuperscript{1}, John Willes\textsuperscript{2}, Jasper Snoek\textsuperscript{1} \textsuperscript{1}Google DeepMind, \textsuperscript{2}Vector Institute [email protected], [email protected], [email protected] ABSTRACT We introduce a deterministic...
IJBsKYXaH4
Section 4.5 is confusing to me. The first part regarding the marginal vs. joint seems to be saying the dependence between distances can be thrown out. This is done without explanation other than a hypothesis and throwing this out makes diffusion on distances no different than diffusion on particles to me. Even in the i...
MOLECULAR CONFORMATION GENERATION VIA SHIFTING SCORES Anonymous authors Paper under double-blind review ABSTRACT Molecular conformation generation, a critical aspect of computational chemistry, involves producing the three-dimensional conformer geometry for a given molecule. Generating molecular conformation via dif...
sGd02fkoAE
Table 1: from the writing, I assume the first row is performing 2D detection (comparing DETR, Swin, CameraViT). The second two rows are 3D detection with lidar and lidar-camera fusion, respectively. Why is the first group being compared to the second two?
FusionViT: Hierarchical 3D Object Detection via Lidar-Camera Vision Transformer Fusion Anonymous authors Paper under double-blind review Abstract For 3D object detection, both camera and lidar have been demonstrated to be useful sensory devices for providing complementary information about the same scenery with data...
RQk9srYfhj
Moreover, the authors claims that the SEELE addresses multiple generative sub-tasks in subject repositioning using a single diffusion model, on one hand, SEELE actually contains many components not only the diffusion model for generative sub-tasks, on the other hand, it is confusing what is the exact meaning of using a...
REPOSITIONING THE SUBJECT WITHIN IMAGE Anonymous authors Paper under double-blind review ABSTRACT Current image manipulation primarily centers on static manipulation, such as replacing specific regions within an image or altering its overall style. In this paper, we introduce an innovative dynamic manipulation task,...
dnaCBAP7X2
“Our identification result on the black-box attack is not as good as the white-box attack tested because of the noise gradient estimation used in the black-box attack.” Why is the identification accuracy for black-box attacks the best in Figure 2 (b) when the number of adversarial examples is greater than one, better t...
AN IMPLICIT WATERMARK FRAMEWORK FOR ADVERSARY IDENTIFICATION Anonymous authors Paper under double-blind review ABSTRACT Security of deep neural networks based machine learning systems has been an emerging research topic, especially after the discovery of adversarial attacks. In general, however, it is very difficult...
7Ttk3RzDeu
The sentence-level score may disproportionately favor summaries that contain a large number of (short) sentences. Furthermore, the evaluation of different models does not further investigate whether some of the score differences can be explained by the different length of the summaries (e.g., a shorter summary may be m...
BooookScore: A SYSTEMATIC EXPLORATION OF BOOK-LENGTH SUMMARIZATION IN THE ERA OF LLMs Yapei Chang University of Massachusetts Amherst [email protected] Kyle Lo Allen Institute for AI [email protected] Tanya Goyal Princeton University [email protected] Mohit Iyyer University of Massachusetts Amherst miyyer...
sOXKeeVxqW
It would be helpful if the authors clarify the specific ratios of atoms that are masked in both SMILES and graphs. Additionally, providing information about what replaces the masked atoms is crucial for a complete understanding of the masking strategy.
MOleSG: A Multi-Modality Molecular Pre-training Framework by Joint Non-overlapping Masked Reconstruction of SMILES and Graph Anonymous authors Paper under double-blind review Abstract Self-supervised pre-training plays an important role in molecular representation learning because labeled molecular data are usually ...
EWTFMkTdkT
In the performance comparison figure 4, is the X axis the time axis? The results surprised me because the prediction from the first 3 models is bad starting from the beginning of the prediction, which means the error mostly comes from decoder reconstruction rather than dynamics prediction. Especially for ODE2VAE, VAE s...
INVARIANCE-BASED LEARNING OF LATENT DYNAMICS Kai Lagemann∗ Statistics and Machine Learning, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany [email protected] Christian Lagemann∗ Department of Mechanical Engineering, University of Washington, Seattle, USA Sach Mukherjee Statistics and Machine Le...
vyGp9Mty2t
How does the method deal with registration problems in CT imaging? The validated datasets in the paper seem to be already registered. If we consider the real CT scanning in practice, for different patients, the patient’s positions will always be different. How can this method deal with the position shift when learning ...
Implicit Neural Representations for Joint Sparse-View CT Reconstruction Anonymous authors Paper under double-blind review Abstract Computed Tomography (CT) plays a crucial role in both medical diagnostics and industrial quality control. Sparse-view CT, in particular, has advantages over standard CT for its reduced i...
B1VWS7ZRm6
Also, the LGB, RTDL, and Resnet methods are listed alongside knowledge transfer baselines and I'm not sure how to interpret them. I assume they refer to scores of models without cross-modality knowledge transfer, but I also assume LGB and RTDL are supposed to target tabular data while Resnet image data. Since the targe...
ON TRANSFERRING EXPERT KNOWLEDGE FROM TABULAR DATA TO IMAGES Anonymous authors Paper under double-blind review ABSTRACT Transferring knowledge across modalities has gained considerable attention in machine learning. Expert knowledge in fields like medicine is often represented in tabular form, and transferring this ...
Zh047FhXqI
As mentioned in the paper, PCM adopt more advanced network architecture to enhance performance so that even the PAM achieves better performance. Does this mean that the advantage of PCM may not come from the algorithmic novelty but from the better network architecture instead?
EFFECTIVE OFFLINE ENVIRONMENT RECONSTRUCTION WHEN THE DATASET IS COLLECTED FROM DIVERSIFIED BEHAVIOR POLICIES Anonymous authors Paper under double-blind review ABSTRACT In reinforcement learning, it is crucial to have an accurate environment dynamics model to evaluate different policies’ value in tasks like offline ...
i92ssjkZCz
Table 8 (a): the detection performance wrt. masking ratio didn't change much when the masking ratio ranged from 0.1 to 0.7. A small masking ratio leads to little information loss, but this pre-training strategy still works well compared to other methods. Hence it is uncertain whether this performance improvement really...
UNiPAD: A UNIVERSAL PRE-TRAINING PARADIGM FOR AUTONOMOUS DRIVING Anonymous authors Paper under double-blind review ABSTRACT In the context of autonomous driving, the significance of effective feature learning is widely acknowledged. While conventional 3D self-supervised pre-training methods have shown widespread suc...
bYwEpQ96ng
Equation 1 seems to only describe how to obtain the prediction but does not involve how to reduce the difference between the prediction and the ground truth. This is not entirely consistent with what ERM describes. Therefore, another issue arises. It seems that no loss function was mentioned. It should be cross-entropy...
Hierarchical Long-tailed Classification with Visual Language Models Anonymous authors Paper under double-blind review Abstract Vision Language Models (VLMs) have shown promising capabilities in handling open vocabulary tasks but struggle with imbalanced data tuning, particularly when dealing with highly skewed label...
Nxn6vGgpI9
Compared to the vision-based action recognition datasets, the scale of WEAR is relatively small, with only 18 people operating 18 workout activities, and the total duration of each activity is 90 seconds, which does not show the minimum times the activity appears in this period.
WEAR: AN OUTDOOR SPORTS DATASET FOR WEARABLE AND EGOCENTRIC ACTIVITY RECOGNITION Anonymous authors Paper under double-blind review ABSTRACT Though research has shown the complementarity of camera- and inertial-based data, datasets which offer both egocentric video and inertial-based sensor data remain scarce. In thi...
0IaTFNJner
The idea of multi-facet embedding or polysemy embedding has been studied quite extensively in the past. From network embedding (Liu et al. Is a single vector enough? exploring node polysemy for network embedding) to recommender systems (Weston et al. Nonlinear latent factorization by embedding multiple user interests)....
On the Embedding Collapse When Scaling up Recommendation Models Anonymous authors Paper under double-blind review Abstract Recent advances in deep foundation models have led to a promising trend of developing large recommendation models to leverage vast amounts of available data. However, we experiment to scale up e...
D1w3huGGpu
For the `HARD HOLDOUTS` set-up, will the performance further improve if there are more easy combinations available for training? Are there any other possible solutions to address the easy-to-hard transfer problem?
COMPOSITIONAL INTERFACES FOR COMPOSITIONAL GENERALIZATION Anonymous authors Paper under double-blind review ABSTRACT With recent work such as GATO (Reed et al., 2022), we see the development of agents that can accomplish a variety of tasks, and are able to perceive the world and act in multiple observation and actio...
GOt2kP383R
x_i in Eq. (2) denotes the feature (activation). However, I wonder if it will lead to the homogenization of features since they are expected to have a low standard deviation. Did the authors try to minimize the difference in the mean of each channel?
ABSTRACT Quantization is a promising approach to reduce the high computational complexity of image super-resolution (SR) networks. However, compared to high-level tasks like image classification, low-bit quantization leads to severe accuracy loss in SR networks. This is because feature distributions of SR networks are...
LWDRiFzbHQ
Can you combine / learn different similarity functions and transformations? If the impact of the different image transformations is limited (as we observe in Table 2), can we (and should we) select them arbitrarily? Is there a reason some transformations might be more beneficial than others and in what settings would t...
VICINAL ASSESSMENT OF MODEL GENERALIZATION Anonymous authors Paper under double-blind review ABSTRACT This paper studies how to assess the generalization ability of classification models on out-of-distribution test sets without relying on test ground truths. Existing works usually compute an unsupervised indicator o...
WbR415lO2L
Wordnet is used for token substitutions in this work. My understanding is that Wordnet could be quite noisy, containing words and tokens that are very infrequent in human-written text. As a result, it is surprising to see that Wordnet substitutions do not result in decreases in readability. I wonder if the authors migh...
Large Language Models can be Guided to Evade AI-Generated Text Detection Anonymous authors Paper under double-blind review Abstract Large language models (LLMs) have shown remarkable performance in various tasks and have been extensively utilized by the public. However, the increasing concerns regarding the misuse o...
wk77w7DG1N
Given that the reference and candidate documents might have varying numbers of sentences, how do you handle sentence-level comparisons? Specifically, do you employ any sentence matching techniques, and if so, how are they implemented?
Evaluating and Improving Generation Consistency of Large Language Models via A Divide-Conquer-Reasoning Approach Anonymous authors Paper under double-blind review Abstract Evaluating the quality and variability of text generated by Large Language Models (LLMs) poses a significant, yet unresolved research challenge. ...
D9SA02esgh
The volume bound has to be selected a-priori for the dataset. This approach seems to not be extensible for non-local morphologies (e.g. considering long range axons would require looking at the entire brain volume)
MORPHOCC: AN IMPLICIT GENERATIVE MODEL OF NEURONAL MORPHOLOGIES Anonymous authors Paper under double-blind review ABSTRACT Understanding the diversity and complexity of the morphology of different types of neurons is important for understanding neural circuits. We need quantitative, unbiased methods to capture the s...
jJvXNpvOdM
The graph-based state representation requires shortest-path computation for all fully connected edges, but this seems quite computationally heavy, especially when we have a large number of nodes, leading to a drastically increasing number of edges.
Task Planning for Visual Room Rearrangement under Partial Observability Karan Mirakhor*, Sourav Ghosh*, Dipanjan Das & Brojeshwar Bhowmick Visual Computing and Embodied Intelligence Lab TCS Research, Kolkata, India {karan.mirakhor, g.sourav10, dipanjan.da, b.bhowmick}@tcs.com Abstract This paper presents a novel mod...
9ztL7Trdnx
If TAFS employs an MLP to approximate the optimal activation function, the performance of activation function would also depend on the number of layers in the MLP and the non-linear transformation functions. This has not been discussed in this paper.
TAFS: Task-aware Activation Function Search for Graph Neural Networks Anonymous authors Paper under double-blind review Abstract Since the inception of Graph Neural Networks (GNNs), extensive research efforts have concentrated on enhancing graph convolution, refining pooling operations, devising robust training stra...
oYjPk8mqAV
Would it be possible for the authors to expand on their improvement on the miniF2F benchmark? Specifically, I would be interested to know what are the specific unsolved theorems that Magnushammer was able to solve, and how complex are those proofs.
MAGNUSHAMMER: A TRANSFORMER-BASED APPROACH TO PREMISE SELECTION Maciej Mikula∗ Google DeepMind† Szymon Tworkowski∗ xAI† Szymon Antoniak∗ Mistral AI† Bartosz Piotrowski IDEAS NCBR Albert Qiaochu Jiang University of Cambridge Jin Peng Zhou Cornell University‡ Christian Szegedy xAI‡ Łukasz Kuciński IDEAS NCBR, IMPAN Piot...
MBIGXMT0qC
Despite the possible data and information leakage in the experimental framework, the proposed method fails to outperform standard protein language models like ESM-2 in tasks such as contact and secondary structure prediction. Given that the model incorporates a broader range of pre-training loss than ESM-2, these resul...
MULTI-SCALE PROTEIN LANGUAGE MODEL FOR UNIFIED MOLECULAR MODELING Anonymous authors Paper under double-blind review ABSTRACT Protein language models have shown great potential in protein engineering. However, the current protein language models mainly work in the residue scale, which cannot offer information in the ...
ruGY8v10mK
In scenarios where there's a very good classifier and a high-quality dataset (like MNIST or CIFAR10, with low aleatoric uncertainty), there might be a significant imbalance between positive and negative samples. This could lead to the situation that the trained uncertainty measure is unreliable. How should one address ...
A DATA-DRIVEN MEASURE OF RELATIVE UNCERTAINTY FOR MISCLASSIFICATION DETECTION Eduardo Dadalto∗ Laboratoire des signaux et systèmes (L2S) Université Paris-Saclay CNRS CentraleSupélec Gif-sur-Yvette, France [email protected] Marco Romanelli∗ New York University New York, NY, USA [email protected] Georg P...
i4kDKfllrz
OE achieves remarkable performance by solely requiring a single network to concurrently manage classification and the rejection of unknowns. It's worth noting that this paper doesn't make any references to OE, and there is a noticeable lack of in-depth discussion or comparison concerning methods and experimental result...
Synergistic Classification and Unknown Discrimination for Open Set Recognition Anonymous authors Paper under double-blind review Abstract Deep learners tend to perform well when trained under the closed set assumption but struggle when deployed under open set conditions. This motivates the field of Open Set Recognit...
4A5D1nsdtj
The design of heterophilic bases relies on the dataset's homophily rate, denoted as $h$ in Algorithm 1. I am concerned this approach is impractical due to obtaining the exact homophily rate $h$ from the training data is not feasible. It appears that the authors have directly utilized the entire dataset, including the l...
AN EFFECTIVE UNIVERSAL POLYNOMIAL BASIS FOR SPECTRAL GRAPH NEURAL NETWORKS Anonymous authors Paper under double-blind review ABSTRACT Spectral Graph Neural Networks (GNNs), also referred to as graph filters have gained increasing prevalence for heterophily graphs. Optimal graph filters rely on Laplacian eigendecompo...
2XkTz7gdpc
The equation after eq.1 combines two independent distributions together, which uses less information to predict v_l, and can lead to error. This is because that edges being refined can be very important for identifying which node is expected to expand.
EFFICIENT AND SCALABLE GRAPH GENERATION THROUGH ITERATIVE LOCAL EXPANSION Andreas Bergmeister* Karolis Martinkus† Nathanaël Perraudin† Roger Wattenhofer ETH Zürich Prescient Design SDSC, ETH Zürich DISCO, ETH Zürich ABSTRACT In the realm of generative models for graphs, extensive research has been conducted. However...
F76bwRSLeK
In section 6 about limitations, I would venture that two reasons for the poor reconstruction could be that (a) the proposed linear encoder does not have enough expressive power or (b) there exists no sparse linear basis to explain the LLM layers or (c) there is no compact natural language description for the sparse fea...
Sparse Autoencoders Find Highly Interpretable Features in Language Models Hoagy Cunningham∗12, Aidan Ewart∗13, Logan Riggs∗1, Robert Huben, Lee Sharkey4 1EleutherAI, 2MATS, 3University of Bristol, 4Apollo Research {hoagycunningham, aidanprattewart, logansmith5}@gmail.com Abstract One of the roadblocks to a better un...
9TG42oozQP
Since the authors have modeled the potential causal graph (Figure 1(c)), why not directly use the causal relationships on the graph to model variables C and M using VAE, instead of modeling variable Z first and then using causal discovery algorithms to distinguish them?
Causal Effect Estimation with Mixed Latent Confounders and Post-treatment Variables Anonymous authors Paper under double-blind review Abstract Causal inference from observational data has attracted considerable attention in recent years. One main obstacle is the handling of confounders. As the direct measure of conf...
EhmEwfavOW
I'm concerned about the datasets used in the node classification task. Results in Table 1 show that the performance of MagNet is far way from FaberNet. I guess that this is because MagNet operates as a low-pass filter and therefore cannot perform well in the heterophilic datasets.
HoloNets: Spectral Convolutions Do Extend to Directed Graphs Christian Koke & Daniel Cremers Technical University Munich and Munich Center for Machine Learning {christian.koke,cremers}@tum.de Abstract Within the graph learning community, conventional wisdom dictates that spectral convolutional networks may only be d...
XyB4VvF01X
Premise selection is one important step for theorem proving, which is not discussed in this work. How to guarantee that all relevant theorems/lemmas are properly included when constructing the graph representation?
GRAPH2TAC: LEARNING HIERARCHICAL REPRESENTATIONS OF MATH CONCEPTS IN THEOREM PROVING Anonymous authors Paper under double-blind review ABSTRACT Concepts abound in mathematics and its applications. They vary greatly between subject areas, and new ones are introduced in each mathematical paper or application. A formal...
9rzEPbs4Wg
According to the impact of masking probability alpha on generalization and safety metrics, it shows a direct correlation between alpha and corruption accuracy. However, why such a correlation is held? Will this correlation be held for more datasets or applications?
IMPROVING GENERALIZATION AND SAFETY OF DEEP NEURAL NETWORKS WITH MASKED ANCHORING Anonymous authors Paper under double-blind review ABSTRACT Anchoring is a recent architecture and task-agnostic technique that can produce state-of-the-art epistemic uncertainty estimates, and improve extrapolation capabilities. Howeve...