Paper_ID
stringlengths
10
10
Question
stringlengths
201
1.81k
ocr_output
stringlengths
252
54k
OZitfSXpdT
In addition, the teacher model indeed may underperform on the outliers. According to previous discussion, directly decreasing the $\alpha$ also can solve this issue. Thus, why to introduce the inter-sample relations? This motivation needs to be further claified.
Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation Chengming Hu\textsuperscript{1,2}\footnote{Equal contribution with random order.}, Haolun Wu\textsuperscript{1,2}\footnote{To whom the correspondence should be addressed.}, Xuan Li\textsuperscript{1,2}, Chen Ma\textsuperscript{3}, Xi ...
TFR0GrzERG
In the introduction, the authors claim that task descriptions with minimal information can impair in-context learning performance because they hinder the model’s ability to learn from in-context examples. I don’t think that this is correct given the presented results (Figure 1).
On Task Description of In-context Learning: A Study from Information Perspective Anonymous authors Paper under double-blind review Abstract Transformers have demonstrated remarkable performance in a wide range of applications, making in-context learning an essential technique. Although the in-context learning has be...
esqRHCwTJ2
In other words, as the marginal feature distribution $P_X$ changes under strategic feedback, it would be expected that the qualification rate $Q(\mathcal{S}_{o,t-1})$ would also change in response, no?
Long-Term Impacts of Model Retraining with Strategic Feedback Anonymous authors Paper under double-blind review Abstract When machine learning (ML) models need to be frequently retrained, it is often too expensive to obtain human-annotated samples, so recent ML models have started to label samples by themselves. Thi...
NddKiWtdUm
Additionally, given the use of simulation with LMs is touted as a benefit of this approach, what are the drawbacks of using this instead of collecting representative feedback (and training a reward model), whose values are the models implicitly aligning to with such simulation? These should be explicitly called out and...
Training Socially Aligned Language Models in Simulated Human Society Ruibo Liu Dartmouth College Ruixin Yang University of British Columbia Chenyan Jia Stanford University, Northeastern University Ge Zhang University of Michigan, Ann Arbor Diyi Yang Stanford University Soroush Vosoughi Dartmouth Colle...
qTlcbLSm4p
Besides, from the equation and algorithm, I can’t find an explicit connection between low-resolution diffusion and high-resolution diffusion; such an explicit statement may be necessary for reader to understand.
Relay Diffusion: Unifying Diffusion Process Across Resolutions for Image Synthesis Jiayan Teng∗1, Wendi Zheng∗1, Ming Ding∗12†, Wenyi Hong1, Jianqiao Wangni2, Zhuoyi Yang1, Jie Tang1† ∗equal contribution 1Tsinghua University 2Zhipu AI †corresponding authors {tengjy20@mails,zhengwd23@mails,jietang@mail}.tsinghua.edu.cn...
RAA0vCLMhp
In Table 4, the results of existing state-of-the-art solutions reported are zero-shot or fine-tuned, and SeMDiff is the zero-shot or fine-tuned? If the results of SOTA are zero-shot, how about the fine-tuned performance on these datasets?
Semantic Memory Guided Diffusion Networks for Image-to-Long Text Generation Anonymous authors Paper under double-blind review Abstract Automatic describing image with comprehensive textual content is often demanded by different real-world applications, which motivates image-to-text generation tasks such as image cap...
LTHWoQ9ac1
Sequential recourse: the problem in Eq (7) doesn’t seem like the real problem that needs to be solved. The set U_P updates after every step, and this changes the argmax A and therefore w_ij. Am I missing anything?
Cost-Adaptive Recourse Recommendation by Adaptive Preference Elicitation Anonymous authors Paper under double-blind review Abstract Algorithmic recourse recommends a cost-efficient action to a subject to reverse an unfavorable machine learning classification decision. Most existing methods in the literature generate...
oJ1tx3fXDA
In Eq.(5), what is the MVR term? Could the author explain this in detail? Its first part is a sum of $J_i$-step updates, while its second part is a sum of $j$-step updates. I know these two parts come from the additional term from $\theta_{i,j}^t-\theta_i^{t-\tau_i}$ to $\theta^{t-1}-\theta^{t-\tau_i-1}$, but how does ...
COMMUNICATION-EFFICIENT HETEROGENEOUS FEDERATED LEARNING WITH GENERALIZED HEAVY-BALL MOMENTUM Anonymous authors Paper under double-blind review ABSTRACT Federated Learning (FL) is the state-of-the-art approach for learning from decentralized data in privacy-constrained scenarios. As the current literature reports, t...
gBLEHzKOfF
Proposition 3.1 assumes $Y \sim \pi_{\epsilon}^{\star} (\cdot | x)$ in Eq 7. In practice, Algorithm 1 adopts the mini-batch estimate for $\hat{\pi}_{ \epsilon}$. Can GENOT still recovers Optimal Conditional Generators by taking expectations over mini-batch estimate, e.g. [4]? In this respect, I am curious about the qua...
Generative Entropic Neural Optimal Transport To Map Within and Across Spaces Anonymous authors Paper under double-blind review Abstract Learning measure-to-measure mappings is a crucial task in machine learning, featured prominently in generative modeling. Recent years have witnessed a surge of techniques that draw ...
WqsYs05Ri7
Is there a way to abstract away the specific form of the concept activation of Oikarinen et al. (2023) in this formulation? Stated differently, what is $\vec{m}(x)$, and $\vec{s}(x)$ for a standard CBM?
ESTIMATION OF CONCEPT EXPLANATIONS SHOULD BE UNCERTAINTY AWARE Anonymous authors Paper under double-blind review ABSTRACT Model explanations are very valuable for interpreting and debugging prediction models. We study a specific kind of global explanations called Concept Explanations, where the goal is to interpret ...
a745RnSFLT
Besides, using the pre-trained LLaMA can improve the PAC-Bayes bound, is it some form of transferring the ''generalization problem'' of the prompt engineering to the generalization problem of the pre-trained language model?
UNDERSTANDING PROMPT ENGINEERING MAY NOT REQUIRE RETHINKING GENERALIZATION Victor Akinwande¹, Yiding Jiang¹, Dylan Sam¹ & J. Zico Kolter¹,² ¹Carnegie Mellon University, ²Bosch Center for AI ABSTRACT Zero-shot learning in prompted vision-language models, the practice of crafting prompts to build classifiers without a...
SkETBJRKH7
I am a bit confused about the results shown in the middle panel of Figures 4 and 5: The PFC-LLM architecture produced zero invalid action proposals in both tasks. Does this imply that the Monitor module is unnecessary, given that its role is to identify invalid action proposals? However, this contradicts the ablation s...
A Prefrontal Cortex-inspired Architecture for Planning in Large Language Models Anonymous authors Paper under double-blind review Abstract Large language models (LLMs) demonstrate impressive performance on a wide variety of tasks, but they often struggle with tasks that require multi-step reasoning or goal-directed ...
y01KGvd9Bw
The major concern I have is the necessity to utilize the token from LLM for image decoding. What is is going to be if you let the LLM to first output the image description, then extract it and feed it directly to the diffusion model?
DREAMLLM: Synergistic Multimodal Comprehension and Creation Runpei Dong \(^{1,2}\) Chunrui Han \(^3\) Yuang Peng \(^4\) Zekun Qi \(^{1,2}\) Zheng Ge \(^3\) Jinrong Yang \(^5\) Liang Zhao \(^3\) Jianjian Sun \(^3\) Hongyu Zhou \(^3\) Haoran Wei \(^3\) Xiangwen Kong \(^3\) Xiangyu Zhang \(^3\) Kaisheng Ma \(^4\) Li Yi \...
kxpswbhr1r
The paper could benefit from a discussion on the practical implications of the findings. For instance, how can the insights on convergence be used to improve training methodologies or model selection in real-world applications?
In-Context Convergence of Transformers Anonymous authors Paper under double-blind review Abstract Transformers have recently revolutionized many domains in modern machine learning and one salient discovery is their remarkable in-context learning capability, where models can solve an unseen task by utilizing task-spe...
dEz3ge8QSo
In section 2, the authors introduce convex risk measures, where there is only one input in the measure $\sigma$ in (5) and (6), while there are two in examples 1 and 2. Could the authors explain what is the difference between these two $\sigma$'s?
SOFT ROBUST MDPs AND RISK-SENSITIVE MDPs: EQUIVALENCE, POLICY GRADIENT, AND SAMPLE COMPLEXITY Runyu (Cathy) Zhang Harvard University runyuzhang@fas.harvard.edu Yang Hu Harvard University yanghu@g.harvard.edu Na Li Harvard University nali@seas.harvard.edu ABSTRACT Robust Markov Decision Processes (MDPs)...
mOTiVzTgF2
The discussion of Adam condition number assumes very very small values of gradient, which is not observed realistically. This discussion should be contextualized given realistic values of the gradient.
ResiDual: Transformer with Dual Residual Connections Anonymous authors Paper under double-blind review Abstract Transformer networks have become the preferred architecture for many tasks due to their state-of-the-art performance. However, the optimal way to implement residual connections in Transformer, which are es...
1vDArHJ68h
In Appendix G (BSuite environment), is there any hypothesis on why sometimes harder environments (longer memory steps) present better performance than easier ones? For instance, R2I’s performance on 31 memory steps is better than 15 memory steps. Similarly, the performance in 81 memory steps seems better (or more stabl...
MASTERING MEMORY TASKS WITH WORLD MODELS Mohammad Reza Samsami∗1,2 Artem Zholus∗1,3 Janarthanan Rajendran1,2 Sarath Chandar1,3,4 1Mila – Quebec AI Institute 2Université de Montréal 3Polytechnique Montréal 4CIFAR AI Chair ABSTRACT Current model-based reinforcement learning (MBRL) agents struggle with long-term depen...
cYksYKbf6K
Maybe I’m missing something, but I do not see why “Our method is computationally simpler compared to previous methods”. Despite being conceptually straightforward, the method requires running a complete rollout up to the maximum length before any truncation occurs. Could you elaborate on the computational saving?
Imagine Within Practice: Conservative Rollout Length Adaptation for Model-Based Reinforcement Learning Anonymous authors Paper under double-blind review Abstract Model-based reinforcement learning (MBRL) algorithms achieve high sample efficiency by leveraging imagined rollouts from a world model for policy optimizat...
5KUiMKRebi
- I wonder about the integer inputs (i.e. the tensor coordinates) to the INR network. Do you normalise these coordinates somehow or directly input the integers into the INR network? Did such an integer input space cause any problems during training?
Implicit Neural Representation Inference for Low-Dimensional Bayesian Deep Learning Panagiotis Dimitrakopoulos¹, Giorgos Sfikas² & Christophoros Nikou¹ ¹Dept. of Computer Science & Engineering, University of Ioannina, Ioannina Greece ²Dept. of Surveying & Geoinformatics, University of West Attica, Athens Greece ¹{p.di...
OinvjdvPjp
Could you please demonstrate how well the proposed method can handle situations where it needs to refer to previously mentioned real numbers in the context, ensuring these numbers remain unaltered? How does this embedding method impact a Language Model's capability to preserve real numbers in the given input?
xVal: A Continuous Number Encoding for Large Language Models Anonymous authors Paper under double-blind review Abstract Large Language Models (LLMs) have not yet been broadly adapted for the analysis of scientific datasets due in part to the unique difficulties of tokenizing numbers. We propose xVal, a numerical enc...
cZo6pDtDZr
As best I can tell, the lower bound in Theorem 3 is not really comparable to the upper bound in Theorem 2 since one imposes $\beta = 0$ in the former while the upper bound requires $\beta>0$. For instance, when $\beta>0$, it does not seem to be necessary to have any dependence on $\alpha$, since it seems to me that sim...
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...
4Qz9BT4mpM
Could you elaborate on the theoretical underpinnings of the 'agreement-on-the-line' method? How does it theoretically and practically differ from existing methods for assessing model performance on out-of-distribution data?
Predicting the Performance of Foundation Models via Agreement-on-the-Line Anonymous authors Paper under double-blind review Abstract Estimating out-of-distribution performance is critical to safely deploy machine learning models. Recently, Baek et al. showed that the phenomenon “agreement-on-the-line” can be a relia...
77N93tc3o5
In the manuscript, it sounds like identifiability should follow trivially from previous works, such as the iVAE framework (Khemakhem et al., 2020). Even if this were the case, it would be helpful to restate and discuss the required assumptions.
DEEP INDEPENDENT VECTOR ANALYSIS Anonymous authors Paper under double-blind review ABSTRACT We introduce a deep multivariate latent variable model, Deep Independent Vector Analysis (DeepIVA), for learning linked and identifiable latent sources across multiple data modalities by unifying multidataset independent subs...
IcR1OOFzxm
One thing I'm not particularly sure is how is the answer selected in RAISE. When you generate the answer, how do you pick the candidate from the given set? PrAE and ALANS actually only generate the hidden latents and compare in the latent space. Do you compare in the pixel space? Do you think comparing in the hidden sp...
Towards Generative Abstract Reasoning: Completing Raven’s Progressive Matrix via Rule Abstraction and Selection Fan Shi Bin Li∗ Xiangyang Xue Shanghai Key Laboratory of Intelligent Information Processing School of Computer Science, Fudan University fshi22@m.fudan.edu.cn {libin,xyxue}@fudan.edu.cn Abstract Endo...
8sKcAWOf2D
The section detailing the use of CMAP to patch activations from Goat-7B to Llama-7B lacks clarity, particularly when validating the impact of QKV on the model's performance. The rationale behind why patching the QK-circuit of the Value Fetcher and the value vector of the Position Transmitter heads results in the most s...
Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking Nikhil Prakash¹* Tamar Rott Shaham² Tal Haklay³ Yonatan Belinkov³ David Bau¹ ¹Northeastern University ²MIT CSAIL ³Technion – IIT Abstract Fine-tuning on generalized tasks such as instruction following, code generation, and mathematics has been...
qODvxQ8TXW
While the authors argue LRR finds a better mask than WR in Figure 3, I wonder if a longer training epochs within each IMP cycle would help WR to find a superior mask. In other words, are both WR and LRR fully converged?
Masks, Signs, And Learning Rate Rewinding Advait Gadhikar & Rebekka Burkholz CISPA Helmholtz Center for Information Security Saarbrücken, Germany {advait.gadhikar, burkholz}@cispa.de Abstract Learning Rate Rewinding (LRR) has been established as a strong variant of Iterative Magnitude Pruning (IMP) to find lottery t...
VAvSUG3hwI
If the agent has to be trained step by step, could this affect how useful the solution is in situations where the agent must work with many teammates simultaneously? For instance, think of an online chatbot on an e-commerce platform where customers show up all together.
One by One, Continual Coordinating with Humans via Hyper-Teammate Identification Anonymous authors Paper under double-blind review Abstract One of the primary objectives in modern artificial intelligence researches is to empower agents to effectively coordinate with diverse teammates, particularly human teammates. P...
G0vdDSt9XM
The paper could provide a more in-depth analysis of the tool creation and retrieval components of CRAFT. Understanding how different types of tools contribute to performance improvements and how the retrieval mechanism interacts with various tasks would offer valuable insights.
CRAFT: Customizing LLMs by Creating and Retrieving from Specialized Toolsets Lifan Yuan∗, Yangyi Chen∗, Xingyao Wang, Yi R. Fung, Hao Peng, Heng Ji University of Illinois Urbana-Champaign {lievanyuan173}@gmail.com {yangyic3,xingyao6,yifung2,haopeng,hengji}@illinois.edu Abstract Large language models (LLMs) are often...
tAcEidZ1Y2
Although reconstruction is an important topic in MRI, are there any other reasons that make the proposed method tie to MRI reconstruction? In other words, whether the proposed method is suitable for potential reconstruction tasks in natural image domains?
Self-supervision Meets Bootstrap Estimation: New Paradigm for Unsupervised Reconstruction with Uncertainty Quantification Anonymous authors Paper under double-blind review Abstract Deep learning-based self-supervised reconstruction (SSR) plays a vital role in diverse domains, including unsupervisedly reconstructing ...
lOsF9k1sxW
In the relevant work, it has been written that the previous defense methods have high calculation costs, which limits their practicability in the actual environment. But won't the calculation of FIM in SFT increase the complexity and cost of calculation?
FISHER INFORMATION GUIDED BACKDOOR PURIFICATION VIA NAÏVE EXPLOITATION OF SMOOTHNESS Anonymous authors Paper under double-blind review ABSTRACT Backdoor attacks during deep neural network (DNN) training have gained popularity in recent times since they can easily compromise the safety of a model of high importance, ...
Y8OaqdX5Xt
How does PToM perform compared with an agent that performs neurally-guided MCTS trained via self-play like AlphaZero, and how much of the fast convergence benefit comes from model-based planning vs. ToM?
PLANNING WITH THEORY OF MIND FOR FEW-SHOT ADAPTATION IN SEQUENTIAL SOCIAL DILEMMAS Anonymous authors Paper under double-blind review ABSTRACT Despite the recent successes of multi-agent reinforcement learning (MARL) algorithms, efficiently adapting to other agents in mixed-motive environments remains a significant c...
WSzRdcOkEx
The definition of robustness in Eq. (1) and Eq.(4) seems to be confusing and possibly wrong. In Eq. (1), $\Delta_{\min}$ is defined as the minimal perturbation of a sample-label pair causing the change of the top-1 class prediction. Then, I understand $g(x)$ is a lower bound of $\Delta_{\min}$. However, in Eq. (3), $g(...
GREAT Score: Global Robustness Evaluation of Adversarial Perturbation Using Generative Models Anonymous authors Paper under double-blind review Abstract Current studies on adversarial robustness mainly focus on aggregating local robustness results from a set of data samples to evaluate and rank different models. How...
LOyYjE0blM
The top and bottom molecule in Fig 1c are identical yet lead to different phenotypes. The title of the figure says that different molecules yield different phenotypes in cells. If there was an intention for some of these aspects, could you possibly clarify it? If not, perhaps drop the second copy of the molecule.
Neural scaling laws for phenotypic drug discovery Anonymous authors Paper under double-blind review Abstract Recent breakthroughs by deep neural networks (DNNs) in natural language processing (NLP) and computer vision have been driven by a scale-up of models and data rather than the discovery of novel computing para...
lygdvIKDxi
How does the similarity of the public dataset to the target model’s dataset affect results? The main results in Table 1 show significant overlap: CIFAR-10 is used as the public dataset when CIFAR-100 is the hidden dataset, and vice versa, and even Tiny-ImageNet has strong similarities compared to CIFAR 10/100.
SEEKER: SEMI-SUPERVISED KNOWLEDGE TRANSFER FOR QUERY-EFFICIENT MODEL EXTRACTION Anonymous authors Paper under double-blind review ABSTRACT Model extraction attacks against neural networks aim at extracting models without white-box access to model internals and training datasets. Unfortunately, most existing methods ...
RIEW6M9YoV
In Table 2, we can observe that the novelty of the generated molecules is low compared to those of the baselines (mainly on QM9). I would expect the authors to provide some explanation or intuitions about why the proposed model fails to produce novel graphs.
GRAPH GENERATION WITH $K^2$–TREES Yunhui Jang, Dongwoo Kim, Sungsoo Ahn Pohang University of Science and Technology {uni5510, dongwookim, sungsoo.ahn}@postech.ac.kr ABSTRACT Generating graphs from a target distribution is a significant challenge across many domains, including drug discovery and social network analys...
EODzbQ2Gy4
If the method is not intended to transfer to different behaviours (only different object displacements), why not just perform planning on the target task directly? TransferStep itself appears to be a form of trajectory optimisation at each step, and at least for the tasks presented in the experiments (which are all rel...
DIFF-TRANSFER: MODEL-BASED ROBOTIC MANIPULATION SKILL TRANSFER VIA DIFFERENTIABLE PHYSICS SIMULATION Anonymous authors Paper under double-blind review ABSTRACT The capability to transfer mastered skills to accomplish a range of similar yet novel tasks is crucial for intelligent robots. In this work, we introduce Dif...
2lDQLiH1W4
Let's say we input a set of views V = [v1, v2, v3, v4] and generate a shape A. Then we multiply all the views with a transformation matrix M and generate a shape B. Will shape A and shape B under the same canonicalized coordinate frame?
Instant3D: Fast Text-to-3D with Sparse-View Generation and Large Reconstruction Model Jiahao Li1,2∗ Hao Tan1 Kai Zhang1 Zexiang Xu1 Fujun Luan1 Yinghao Xu1,3 Yicong Hong1,4 Kalyan Sunkavalli1 Greg Shakhnarovich2 Sai Bi1 1Adobe Research 2TTIC 3Stanford University 4Australian National University {jiahao,greg}@ttic.edu y...
dIjwC8A0N6
The Lion optimizer demonstrates greater efficiency when employed with larger batch sizes and lower learning rates. It would be great if the authors can address how these specific characteristics of the Lion optimizer influence the training process within their experimental setup.
QFT: Quantized Full-parameter Tuning of LLMs with Affordable Resources Anonymous authors Paper under double-blind review Abstract Large Language Models (LLMs) have showcased remarkable impacts across a wide spectrum of natural language processing tasks. Fine-tuning these pre-trained models on downstream datasets pro...
52fz5sUAy2
why $r_{u,i}$ is affected by $o_{u,i}$? In my opinion, $o_{u,i}$ is just a treatment to observe $r_{u,i}$, and does not affect the 'value' of $r_{u,i}$. (i.e., the value of $r$ is affected only by $x$ and observed only when $o=1$).
Be Aware of the Neighborhood Effect: Modeling Selection Bias under Interference Haoxuan Li1 Chunyuan Zheng1 Sihao Ding2 Peng Wu3,* Zhi Geng3 Fuli Feng2 Xiangnan He2 1Peking University 2University of Science and Technology of China 3Beijing Technology and Business University hxli@stu.pku.edu.cn dsihao@mail.ustc.edu.cn ...
Spp2i1hKwV
ideally, rather than (arbitrarily?) choosing k = 18 & k = 100, you should provide an automated approach to AUTOMATICALLY assess the smallest amount of annotations that leads to the best possible performance; you only introduce Auto-IDEAL in
IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models Shaokun Zhang\(^1*\) Xiaobo Xia\(^2*\)\(^†\) Zhaoqing Wang\(^2\) Ling-Hao Chen\(^3\) Jiale Liu\(^4\) Qingyun Wu\(^1†\) Tongliang Liu\(^2\) \(^1\)Pennsylvania State University \(^2\)The University of Sydney \(^3\)Tsingh...
xRiZddh5Pb
The authors explain the idea behind designing the input features in Proposition 1, wherein a ranking metric using the designed distance input features following a specific format that meets the RankPres property. It is not straightforward to me why this format would satisfy the property. Detailed proof would be helpful...
Learning from A Single Graph is All You Need for Near-Shortest Path Routing Anonymous authors Paper under double-blind review Abstract We propose a simple algorithm that needs only a few data samples from a single graph for learning local routing policies that generalize across classes of geometric random graphs in ...
xAqcJ9XoTf
In Figure 2, how did you compute the Lipschitz constant of MLPs? We can compute the Lipschitz constant for models that consist of a single layer, but exact computation of the Lipschitz constant of MLPs is NP-hard [1].
ON THE STABILITY OF EXPRESSIVE POSITIONAL ENCODINGS FOR GRAPHS Yinan Huang*1, William Lu*2, Joshua Robinson3, Yu Yang4, Muhan Zhang5, Stefanie Jegelka6, Pan Li1 1Georgia Institute of Technology, 2Purdue University, 3Stanford University, 4Tongji University, 5Peking University, 6MIT CSAIL {yhuang903, panli}@gatech.edu, ...
oGNdBvymod
This could be due to my limited understanding to Bayesian NN: For the image classification experiement, how do you compare the results obtained from sampling-based and optimization-based algorithm? Specifically, do you obtain certain point estimates of NN's weights from MCMC samples, and compute metrics on test dataset...
Entropy-MCMC: Sampling from Flat Basins with Ease Bolian Li, Ruqi Zhang Department of Computer Science, Purdue University, USA {li4468,ruqiz}@purdue.edu Abstract Bayesian deep learning counts on the quality of posterior distribution estimation. However, the posterior of deep neural networks is highly multi-modal in ...
qxLVaYbsSI
In the scenario of both unlabeled and labeled data, some combinations of federated algorithms and semi-supervised algorithms does not perform as well as federated learning using only labeled data, which violates normal cognition. The authors can try to explain the reasons for this phenomenon.
Robust Training of Federated Models with Extremely Label Deficiency Yonggang Zhang\textsuperscript{1}\thanks{Equal contributions.} \quad Zhiqin Yang\textsuperscript{1}\thanks{Equal contributions.} \quad Xinmei Tian\textsuperscript{2} \quad Nannan Wang\textsuperscript{3} \quad Tongliang Liu\textsuperscript{4} \quad Bo ...
TiY8Cvc2SR
In clinical practice the smaller the areas of tumor the more challenging the case is so degradation in performance in the datasets with smaller number of positive tiles may make it so this method is not the best choice for more challenging datasets.
PROGRESSIVE PSEUDO BAG AUGMENTATION WITH INSTANCE IMPORTANCE ESTIMATION FOR WHOLE SLIDE IMAGE CLASSIFICATION Anonymous authors Paper under double-blind review ABSTRACT In the field of computational pathology, the classification of whole-slide images (WSI) remains a challenging task due to the vast amount of gigapixe...
s9bCeJGUJi
D3GAT is proposed as an adaptation of DDGAT to overcome the issue that DDCAT “jointly calculates the invariant pattern and the variant pattern of the node feature, which limits pattern learning, because they are decided jointly and cannot adjust with regard to the other.” This claim seems to be unjustified. Through rea...
CURRICULUM DYNAMIC GRAPH INVARIANT LEARNING UNDER DISTRIBUTION SHIFT Anonymous authors Paper under double-blind review ABSTRACT Dynamic graph neural networks have attracted intensive research interests recently but generally suffer from handling distribution shifts that widely exist in dynamic graphs. Although the e...
3VD4PNEt5q
The methodology, while promising, seems to be narrowly tailored for a specific set of fusion models. This raises concerns about its universality. An in-depth exploration into its effectiveness against a broader spectrum of fusion models would have provided a more comprehensive perspective, allowing for a holistic under...
FUSION IS NOT ENOUGH: SINGLE MODAL ATTACKS ON FUSION MODELS FOR 3D OBJECT DETECTION Zhiyuan Cheng¹ Hongjun Choi² Shiwei Feng¹ James Liang³ Guanhong Tao¹ Dongfang Liu³ Michael Zuzak³ Xiangyu Zhang¹ ¹Purdue University {cheng443, feng292, taog, xyzhang}@purdue.edu ²DGIST hongjun@dgist.ac.kr ³Rochester Institute of Techno...
9XdLlbxZCC
Missing analysis. From Table 4, we can see that the model's performance is quite sensitive to the used backbone (more than 10% between Rsenet50 and ConvNext. However, the authors didn't give explanation. Besides, the proposed model uses six loss terms in total for both flow estimation and content learning. I am wonderi...
MC-JEPA: A JOINT-EMBEDDING PREDICTIVE ARCHITECTURE FOR SELF-SUPERVISED LEARNING OF MOTION AND CONTENT FEATURES Anonymous authors Paper under double-blind review ABSTRACT Self-supervised learning of visual representations has been focusing on learning content features, which do not capture object motion or location, ...
rpwES4pe9W
Including coordinate inputs inevitably increase the computational cost, as each coordinate has to go through a (usually larger) MLP network instead of grid sampling and small MLP feed-forward as multi-plane feature input. In fact, one key motivation of multi-plane methods is decreasing the training and rendering time r...
REFINED TENSORIAL RADIANCE FIELD: HARNESSED COORDINATE-BASED NETWORKS FOR NOVEL VIEW SYNTHESIS FROM SPARSE INPUTS Anonymous authors Paper under double-blind review ABSTRACT The multi-plane encoding approach has been highlighted for its ability to serve as static and dynamic neural radiance fields without sacrificing...
1zhM0XkQh0
The PGD-20 metric of the proposed method is pretty worse than the other SOTA in most cases but it is not adequately discussed or mentioned. Could the authors provide some intuitions why does such degradation on PGD-20 happen, any investigation to address that drawback?
PROFeAT: PROJECTED FEATURE ADVERSARIAL TRAINING FOR SELF-SUPERVISED LEARNING OF ROBUST REPRESENTATIONS Anonymous authors Paper under double-blind review ABSTRACT Supervised adversarial training has been the most successful approach for improving the robustness of Deep Neural Networks against adversarial attacks. Whi...
IOrnCVIKIZ
In table 2 and table 5, are the pretrained model performances on HumanEval measured with or without postprocessing? I believe a more apples-to-apples comparison would be between Pretrained model w/ postprocessing vs LETI w/o postprocessing. Given that LETI seems to learn how to clean up syntax errors, it seems a more f...
LETI: Learning to Generate from Textual Interactions Anonymous authors Paper under double-blind review Abstract Finetuning pre-trained language models (LMs) is essential for enhancing their capabilities and is a crucial phase in their lifecycles. Existing techniques commonly fine-tune on input-output pairs (e.g., in...
K804zYw6Wc
For GMM and LMM, NIR and RGB are concatenated along channel dimensions to input the same subsequent modules, and then split to obtain the two estimated NIR weights and RGB weights to apply to the corresponding branches, respectively. It seems to me that two estimated weights are the same, how to achieve selective fusio...
NIR-Assisted Image Denoising: A Selective Fusion Approach and a Real-World Benchmark Dataset Anonymous authors Paper under double-blind review Abstract Despite the significant progress in image denoising, it is still challenging to restore fine-scale details while removing noise, especially in extremely low-light en...
yqIJoALgdD
However, one of the appealing aspects of SNNs is their utilization of binary events for information processing. In essence, a spiking neuron is active only when it encounters spikes, enabling an event-driven regime and an energy-efficient system. However, the proposed neuron nodes deviate from this advantageous feature...
Towards Zero Memory Footprint Spiking Neural Network Training Anonymous authors Paper under double-blind review Abstract Spiking Neural Networks (SNNs), as representative brain-inspired neural networks, emulate the intrinsic characteristics and functional principles of the biological brain. With their unique structu...
fjwZHuQ3cm
In Table 1, what is meant by Detectors $d$ having different attack names as subscripts to them such as $d_{\text{FGSM}}$? Is this detector tuned to defend against FGSM? If so, why is there only one detector for $LIMANS_{10}$ and not one for the claimed strongest attack of $\text{LIMANS}_{4000}$? What is SA (standard ac...
LIMANS: Linear Model of the Adversarial Noise Space Anonymous authors Paper under double-blind review Abstract Recent works have revealed the vulnerability of deep neural network (DNN) classifiers to adversarial attacks. Among such attacks, it is common to distinguish specific attacks adapted to each example from un...
fvse7bMkAs
Although there are CLT results for the estimators, the validity of the statistical test is lacking --- in particular, the validity of the Bootstrap approximation. The CLT result is not sufficient for the validity of the test since the limiting distribution contains the unknown variance term; does Bootstrap approximatio...
RISK ASSESSMENT AND STATISTICAL SIGNIFICANCE IN THE AGE OF FOUNDATION MODELS Anonymous authors Paper under double-blind review ABSTRACT We propose a distributional framework for assessing socio-technical risks of foundation models with quantified statistical significance. Our approach hinges on a new statistical rel...
3ROGsTX3IR
Am I right to understand the theory quantitatively predicts the experiments only in the GFL phase, and only holds qualitatively in terms of phenomenology to explain the transition to GMFL-I and II phases?
GROKKING AS A FIRST ORDER PHASE TRANSITION IN TWO LAYER NETWORKS Noa Rubin∗ Inbar Seroussi† Zohar Ringel ‡ ABSTRACT A key property of deep neural networks (DNNs) is their ability to learn new features during training. This intriguing aspect of deep learning stands out most clearly in recently reported Grokking pheno...
KqbCvIFBY7
The illustrated plots in Figure 1 are a bit misleading to me. Why are the initial points concentrated in one mode? In a high-dimensional setting, the chance of sampling close-by Gaussian points is low.
Particle Guidance: non-I.I.D. Diverse Sampling with Diffusion Models Gabriele Corso∗1, Yilun Xu1, Valentin de Bortoli2, Regina Barzilay1, Tommi Jaakkola1 1CSAIL, Massachusetts Institute of Technology, 2ENS, PSL University Abstract In light of the widespread success of generative models, a significant amount of resea...
XUCAA0XnPC
Additionally, the client is supposed to be computationally restricted, which is why a server is involved in the first place. How, then, is the client able to train the entire model locally (Section 4.2)?
ENSEMBLER: COMBATING MODEL INVERSION ATTACKS USING MODEL ENSEMBLE DURING COLLABORATIVE INFERENCE Anonymous authors Paper under double-blind review ABSTRACT Deep learning models have exhibited remarkable performance across various domains. Nevertheless, the burgeoning model sizes compel edge devices to offload a sign...
ASppt1L3hx
The parameter (\kappa) for batch dependency is set as 256, however, Figure 4 indicates minimal difference between 64, 256, and even infinity. Furthermore, Figure 3 shows that the GNN model validation F1-score drops when \kappa is 256 (or larger). Given these observations, how do you justify the choice of \kappa as 256 ...
COOPERATIVE MINIBATCHING IN GRAPH NEURAL NETWORKS Anonymous authors Paper under double-blind review ABSTRACT Significant computational resources are required to train Graph Neural Networks (GNNs) at a large scale, and the process is highly data-intensive. One of the most effective ways to reduce resource requirement...
PCm1oT8pZI
According to the experimental results in Table 4, we still have a high OoDWSR after model extraction. Could you explain why the model extraction methods using i.i.d data can also extract the watermarks that are generated using the OoD data?
SAFE AND ROBUST WATERMARK INJECTION WITH A SINGLE OoD IMAGE Shuyang Yu1, Junyuan Hong1,2, Haobo Zhang1, Haotao Wang2, Zhaoyang Wang2 and Jiayu Zhou1 1Department of Computer Science and Engineering, Michigan State University 2Department of Electrical and Computer Engineering, University of Texas at Austin {yushuyan,hon...
86zAUE80pP
The hard cutoff based on hyperparameter k seems a bit weird to me. Towards what metric would the hyperparameter k be optimized for if it can't be equation 4 itself? This gets even weirder for me when the hard cutoff is then relaxed in equation 6. I don't get why it was ever introduced to begin with.
CPPO: Continual Learning for Reinforcement Learning with Human Feedback Han Zhang\textsuperscript{1,2}, Yu Lei\textsuperscript{2,*}, Lin Gui\textsuperscript{3}, Min Yang\textsuperscript{4}, Yulan He\textsuperscript{4}, Hui Wang\textsuperscript{2}, Ruifeng Xu\textsuperscript{1,2,5,*} \textsuperscript{1} Harbin Institu...
7bIpWYhCdu
To justify that FILI outperforms large language models (LLMs), the edit accuracy subject to some delta-distance is used, with δ denoting the number of changes the fixer makes to the incorrect program. It turns out this edit accuracy is inadequate and potentially biased, as it overlooks the semantic correctness of the p...
FILI: Syntax Repair By Learning From Own Mistakes Anonymous authors Paper under double-blind review Abstract Automatically fixing syntax errors in programs is a key challenge in Software Engineering community. Although, there are millions of programs on the web, both syntactically correct and incorrect, finding a la...
T8Rf1CRbHQ
All the convergence rates in this paper are in terms of the weighted sum of squared norms, while for unbiased SA or TTSA, the last-iterate convergence is achievable. Is this an inevitable issue when there exist structured errors? Or how does the last iterate behave in this case?
ABSTRACT Two-time-scale stochastic approximation is a recursive algorithm for solving a system of two equations. The method has found broad applications in many areas including machine learning and reinforcement learning. Recent works have revealed that single-time-scale stochastic approximation (especially its varian...
FlEUIydMMh
Theorem 3.6 states that the DAG G is identifiable with the conditions that $M_{i}\perp M_{j}\iff i-j\notin E^{\mathcal{U}}$, which is problematic. Since a DAG identifiable means that every direction for every edge in the DAG will be identified, which is impossible without any assumptions under the existence of a latent...
Neuro-Causal Factor Analysis Anonymous authors Paper under double-blind review Abstract Factor analysis (FA) is a statistical tool for studying how observed variables with some mutual dependences can be expressed as functions of mutually independent unobserved factors, and it is widely applied throughout the psychol...
xHmCdSArUC
Also, $\nu$ is currently set to some small value in the experiments, so $(1-\nu)^t$ decays significantly slower than $\binom{1/2}{t}$. Thus, I wonder if $\nu$ can be dropped to save some tuning effort.
CORRELATED NOISE PROVABLY BEATS INDEPENDENT NOISE FOR DIFFERENTIALLY PRIVATE LEARNING Christopher A. Choquette-Choo* Krishnamurthy (Dj) Dvijotham* Krishna Pillutla* Arun Ganesh Thomas Steinke Abhradeep Guha Thakurta Google ABSTRACT Differentially private (DP) learning algorithms inject noise into the learning proces...
KrtGfTGaGe
Perhaps related to this additional assumption, the future work section mentioned adaptations under relaxed assumptions, with a citation to (Lambrechts et al. 2022). It would help to elaborate on some of those potential relaxations.
THE WASSERSTEIN BELIEVER LEARNING BELIEF UPDATES FOR PARTIALLY OBSERVABLE ENVIRONMENTS THROUGH RELIABLE LATENT SPACE MODELS Raphael Avalos\textsuperscript{1}*\hspace{1em} Florent Delgrange\textsuperscript{1,2}*\hspace{1em} Ann Nowé\textsuperscript{1}†\hspace{1em} Guillermo A. Pérez\textsuperscript{2,3}†\hspace{1em} Di...
RJDjSXNuAZ
The method is evaluated on only virus detection in electron microscopy images, where viruses do not overlap. Thus, the method may not generalize to object detection (e.g., cell or nuclei detection) in other microscopy imaging modalities, such as hematoxylin and eosin (H&E) or immunohistochemistry (IHC) stained brightfi...
Weakly Supervised Virus Capsid Detection with Image-Level Annotations in Electron Microscopy Images Hannah Kniesel, Leon Sick, Tristan Payer, Tim Bergner, Kavitha Shaga Devan, Clarissa Read, Paul Walther, Timo Ropinski Ulm University Pedro Hermosilla TU Vienna Abstract Current state-of-the-art methods for object de...
hdCDVSPQ7v
What is the update frequency used in Shampoo when comparing wall-clock time? Is possible to use a second-order update interval for Shampoo such that it runs at a similar speed to Jorge to achieve the target accuracy?
Jorge: Approximate Preconditioning for GPU-Efficient Second-Order Optimization Anonymous authors Paper under double-blind review Abstract Despite their better convergence properties compared to first-order optimizers, second-order optimizers for deep learning have been less popular due to their significant computati...
r2ve0q6cIO
I am not sure I understand why one would want only the inference to be distributed if the training is not. More precisely, if the graphs are so large that inference needs to be distributed, how was the model even trained?
Graph Neural Networks Gone Hogwild Anonymous authors Paper under double-blind review Abstract Graph neural networks (GNNs) constitute a dominant class of architectures for modeling graph-structured data. Message-passing GNNs in particular appear to be ideal for applications where distributed inference is desired, si...
kE9bsfMgin
In the light that there is no or little correlation between intrinsic bias measures and bias observed in a downstream task, how do you think the analysis of bias in attention heads is useful for downstream tasks?
ABSTRACT Transformer-based pretrained large language models (PLM) such as BERT and GPT have achieved remarkable success in NLP tasks. However, PLMs are prone to encoding stereotypical biases. Although a burgeoning literature has emerged on stereotypical bias mitigation in PLMs, such as work on debiasing gender and rac...
2XwBIcywWM
Algorithm 1 indicates the use of $n$ samples for each domain. Could the authors provide guidance on how to effectively balance these samples across various domains to ensure a harmonized and representative dataset for each domain involved?
Learning Variational Neighbor Labels for Test-Time Domain Generalization Anonymous authors Paper under double-blind review Abstract This paper strives for domain generalization, where models are trained exclusively on source domains before being deployed on unseen target domains. We follow the strict separation of s...
JiTVtCUOpS
As a continuation of my previous question, the use of cross-correlation for identifying lead-lag relationships suggests a focus on linear associations. May I inquire if this suggests that the algorithm's applicability is confined to variables that share a linear relationship (e.g., $X_1 = X_2^2$, the cross-correlation ...
Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators Lifan Zhao Shanghai Jiao Tong University mogician233@sjtu.edu.cn Yanyan Shen* Shanghai Jiao Tong University shenyy@sjtu.edu.cn Abstract Recently, channel-independent methods have achieved state-of-the-ar...
lgmCGI2IpI
In experiments, baseline “FULL” is using the entire labeled set for training? And the proposed AL algorithm querying much less labels could often surpass “FULL” from Figure 3? If so, seems interesting to understand the reasons behind. E.g., Is it because AQOT is better at filtering out noise?
AN EFFICIENT QUERY STRATEGY FOR ACTIVE LEARNING VIA OPTIMAL TRANSPORT Anonymous authors Paper under double-blind review ABSTRACT Active Learning (AL) aims to reduce labeling costs by iteratively querying instances. Existing AL methods typically query instances based on either informativeness or representativeness. O...
5Lp6qU9hzV
In Table 4, it's intriguing to observe that BERT-MPU performs better than Roberta-MPU on full text but worse on short text. The authors could provide further insights or hypotheses as to why this performance gap exists. This could help readers better understand the nuances of the approach and its applicability in diffe...
MULTISCALE POSITIVE-UNLABELED DETECTION OF AI-GENERATED TEXTS Yuchuan Tian¹, Hanting Chen², Xutao Wang², Zheyuan Bai², Qinghua Zhang², Ruifeng Li⁴, Chao Xu¹, Yunhe Wang²∗ ¹ National Key Lab of General AI, School of Intelligence Science and Technology, Peking University ² Huawei Noah’s Ark Lab ³ Huawei Group Finance ⁴...
SZn1Ex72Lv
What specific issues or challenges in the field of Feed Forward Neural Networks (FFNNs) are you addressing with your proposed concept of block-operations and the Multiplexer? How does this concept enhance FFNNs, and what practical applications or benefits can be derived from it?
Block-operations: Creating an Inductive Bias to Route Data and Reuse Subnetworks Anonymous authors Paper under double-blind review Abstract Feed Forward Neural Networks (FNNs) often suffer from poor generalization due to their inability to effectively develop and reuse subnetworks for related tasks. Csordás et al. (...
pbLjYjjWqd
Sec 4.4 observed that most clients suffer from the dominant class issues. However, what is the loss for local clients when they are not well fitting across different classes? Also, is dominate class are “easy samples”?
FedBPT: Efficient Federated Black-box Prompt Tuning for Large Language Models Anonymous authors Paper under double-blind review Abstract Pre-trained language models (PLM) have revolutionized the NLP landscape, achieving stellar performances across diverse tasks. These models, while benefiting from vast training data...
H8CtXin7mZ
In the context of traditional approaches of multigrid, either solver or preconditioner, either geometric or algebraic, imposing (nonhomogeneous) Neumann BC can be challenging and is usually an ad-hoc business.
A NEURAL-PRECONDITIONED POISSON SOLVER FOR MIXED DIRICHLET AND NEUMANN BOUNDARY CONDITIONS Anonymous authors Paper under double-blind review ABSTRACT We introduce a neural-preconditioned iterative solver for Poisson equations with mixed boundary conditions. The Poisson equation is ubiquitous in scientific computing:...
FhbZ1PQCaG
Is the memory not initialized throughout the entire training process? Clarifying this point could help readers better understand the novelty of this approach, as memory initialization is typically performed per episode (e.g., NTM).
THINK BEFORE YOU ACT: DECISION TRANSFORMERS WITH INTERNAL MEMORY Anonymous authors Paper under double-blind review ABSTRACT Decision transformer model-based decision-making agents have shown the ability to generalize across multiple tasks. However, their performance relies on massive data and computation. We argue t...
rKMQhP6iAv
On the probing experiment. Technically speaking, if your data split is 50/50 yet the F1 is only 65%, isn’t it unconvincing that we could decode persona before the answers being generated? Could you provide other metrics, like accuracy, which is more widely adopted in probing literature?
PERSONAS AS A WAY TO MODEL TRUTHFULNESS IN LANGUAGE MODELS Anonymous authors Paper under double-blind review ABSTRACT Large Language Models (LLMs) are trained on vast amounts of text from the internet, which contains both factual and misleading information about the world. Can language models discern truth from fals...
ESq3U7z6FD
In Sec 4.3, why is using beam-size = branching factored referred to as exact search? Are all documents visited in this setting? Also, how does beam-size = 0.1*branching_factor ensure that we search up to 10% documents? My understanding is that beam-size=b means that we end up at b leaf nodes and then exhaustively rank ...
EHI: End-to-end Learning of Hierarchical Index for Efficient Dense Retrieval Anonymous authors Paper under double-blind review Abstract Dense embedding-based retrieval is now the industry standard for semantic search and ranking problems, like obtaining relevant web documents for a given query. Such techniques use a...
bUv5gJAAxH
The authors have not accounted for the potential for bias in their Z-score test. The variance of the ID (and therefore the significance of the Z-score) itself strongly depends on the dimensionality within which it is assessed. Setting a threshold for hypothesis testing that is uniform across all dimensions may not be a...
Relating Implicit Bias and Adversarial Attacks through Intrinsic Dimension Anonymous authors Paper under double-blind review Abstract Despite their impressive performance in classification, neural networks are known to be vulnerable to adversarial attacks. These attacks are small perturbations of the input data desi...
sFJr7okOBi
I don't understand the overall evaluation setup. What does 'We randomly generate 1000 protein sequences from these models'. What metadata did you condition on? Was it 1000 different sets of metadata? How do you make this comparison fair when using models like ESM that don't have the ability to condition on metadata?
NL2ProGPT: TAMING LARGE LANGUAGE MODEL FOR CONVERSATIONAL PROTEIN DESIGN Anonymous authors Paper under double-blind review ABSTRACT Large Language Models (LLMs), like ChatGPT, excel in cross-modal tasks thanks to their powerful abilities in natural language comprehension, generalization, and reasoning. Meanwhile, th...
DOerIFfUbs
In Tab. 3, could you explain why you useLLaVA-Bench which is often used to test the multimodal instruction ability? And why the UTA with G/14 model did not beat that of L/14 model but exceeded in conversation and reasoning by a large margin?
Enhancing Vision-Language Model with Unmasked Token Alignment at Scale Anonymous authors Paper under double-blind review Abstract Contrastive pre-training on image-text pairs, exemplified by CLIP, becomes a standard technique for learning multi-modal visual-language representations. Although CLIP has demonstrated re...
cmcD05NPKa
In claim U2, one can see that 21 is wrongly classified into $C_4$. As a result I wonder, for other bases that are product of larger primes, e.g., again, $2021=43\times 47$, can this phenomenon still be observe?
LEARNING THE GREATEST COMMON DIVISOR: EXPLAINING TRANSFORMER PREDICTIONS François Charton Meta AI fcharton@meta.com ABSTRACT The predictions of small transformers, trained to calculate the greatest common divisor (GCD) of two positive integers, can be fully characterized by looking at model inputs and outputs. As tr...
AOpJ3vPNu8
Descriptions in page 1 “However, optimally minimizing the individual task losses in the inner loop may not essentially lead to minimizing the average loss in the outer-loop…”. Based on my understanding, the goal of MAML is to find the meta initialization that minimizes the average inner task losses. So will the failure...
A Game Theoretic Approach to Meta-Learning Nash Model-Agnostic Meta-Learning Anonymous authors Paper under double-blind review Abstract Meta-learning, or learning to learn, aims to develop algorithms that can quickly adapt to new tasks and environments. Model-agnostic meta-learning (MAML), proposed as a bi-level opt...
b3LNKq6tfA
While the introduction of the RUID dataset (and its creation) are very interesting and useful, I argue if the randomness of the approach could generate many samples that are very hard to transform to code, thus impeding the improvement of performance at training time.
Learning UI-to-Code Reverse Generator using Visual Critic without Rendering Anonymous authors Paper under double-blind review Abstract Automated reverse engineering of HTML/CSS code from UI screenshots is an important yet challenging problem with broad applications in website development and design. In this paper, w...
9k0krNzvlV
The motivation of this paper is there is a limitation of decoding-based watermarking, namely, replacing it with a normal decoder. Is the assumption practical? In practice, for an LLM API, how do we conduct such an operation?
ON THE LEARNABILITY OF WATERMARKS FOR LANGUAGE MODELS Chenchen Gu, Xiang Lisa Li, Percy Liang, Tatsunori Hashimoto Stanford University {cygu, xlisali, thashim}@stanford.edu, pliang@cs.stanford.edu ABSTRACT Watermarking of language model outputs enables statistical detection of model-generated text, which can mitigat...
SFCHv2G33F
Could you please provide more details on the secondary structure prediction (SSP) task? How many classes is this task composed of? Is there a separate prediction head for structure prediction on top of the cryptic pocket prediction head? Why and how is its weight in the objective function chosen to be 1?
Protein Language Models Enable Accurate Cryptic Ligand Binding Pocket Prediction Anonymous authors Paper under double-blind review Abstract Accurate prediction of protein-ligand binding pockets is a critical task in protein functional analysis and small molecule pharmaceutical design. However, the flexible and dynam...
QuIiLSktO4
For the synthetic noise a log-normal distribution is used. I believe you meant the parameters for this distribution are $\mu = 0$ and $\sigma \in [0,50]$. The mean of a log-normal distribution cannot be 0...
ALGORITHMS FOR CACHING AND MTS WITH REDUCED NUMBER OF PREDICTIONS∗ Karim Abdel Sadek University of Amsterdam† karim.abdel.sadek@student.uva.nl Marek Eliáš Department of Computing Sciences Bocconi University marek.elias@unibocconi.it ABSTRACT ML-augmented algorithms utilize predictions to achieve performance beyond ...
qg5JENs0N4
Just curious: regarding Lemma 4.1, do you have any comments or implications on the (sort of) bias term $\E_{p(h)} \left[ p^{\beta_h}_+ (s_{t+} \mid s) p^{\beta_h}_+ (s) \right] - p^{\beta}_+ (s_{t+} \mid s) p^{\beta}_+ (s)$ ?
CLOSING THE GAP BETWEEN TD LEARNING AND SUPERVISED LEARNING – A GENERALISATION POINT OF VIEW Raj Ghugare\textsuperscript{1} Matthieu Geist\textsuperscript{2} Glen Berseth\textsuperscript{1,*} Benjamin Eysenbach\textsuperscript{3,*} \textsuperscript{1}Mila, Université de Montréal \textsuperscript{2}Google DeepMind \te...
xwZhyKynCB
Definition 12: For an abstract query graph G, a grounding is a function I that maps G into a query graph. Do you impose any restrictions on this mapping? For example, could two distinct nodes with the type
EFO$_k$-CQA: Towards Knowledge Graph Complex Query Answering beyond Set Operation Anonymous authors Paper under double-blind review Abstract To answer complex queries on knowledge graphs, logical reasoning over incomplete knowledge is required due to the open-world assumption. Learning-based methods are essential be...
9OevMUdods
Data leaky may happen. Note that the questions and answers are annotated based on the text of previous public datasets. The pretraining data of evaluated LLMs is likely to contain the text. This may cause data leaky, and LLMs might have memorized/learned shortcuts to answer the question. It makes the evaluation results...
Towards Understanding Factual Knowledge of Large Language Models Xuming Hu\textsuperscript{1,2*}, Junzhe Chen\textsuperscript{1*}, Xiaochuan Li\textsuperscript{1*}, Yufei Guo\textsuperscript{1}, Lijie Wen\textsuperscript{1†}, Philip S. Yu\textsuperscript{3}, Zhijiang Guo\textsuperscript{4†} \textsuperscript{1} Tsingh...
wlqkRFRkYc
The primary motivation of the paper is to address the limitations of two-dimensional image features in effectively representing global features within autonomous driving scenarios. The authors propose the use of BEV representations. The reviewer wonders why the authors did not explore 3D occupancy as a potential soluti...
BEV-CLIP: Multi-modal BEV Retrieval Methodology for Complex Scene in Autonomous Driving Anonymous authors Paper under double-blind review Abstract The demand for the retrieval of complex scene data in autonomous driving is increasing, especially as passenger vehicles have been equipped with the ability to navigate u...
Ax9cPWDKkR
While using a post-hoc metric like Shapley or the proposed AI is beneficial in some circumstances, it should be mentioned that doing so requires extra (test time) compute. Algorithms like Q-MIX or VDN do not suffer from this issue, in the sense that they provide an estimation of contribution at any point in training.
EFFICIENTLY QUANTIFYING INDIVIDUAL AGENT IMPORTANCE IN COOPERATIVE MARL Anonymous authors Paper under double-blind review ABSTRACT Measuring the contribution of individual agents is challenging in cooperative multi-agent reinforcement learning (MARL). In cooperative MARL, team performance is typically inferred from ...
22to0JZ4zh
Table 2 suggests that SSBM is more effective at preserving the marginal than minimizing the KL divergences. Do the authors have any insights or intuitions about why this is the case? It may be helpful to explain how this observation arises, as it is not explicitly implied by the algorithm itself.
SYMMETRIZED SCHRÖDINGER BRIDGE MATCHING Anonymous authors Paper under double-blind review ABSTRACT Schrödinger bridge (SB) has demonstrated numerous applications in probabilistic generative modeling. Finding the solution of probability paths aligns with entropy-regularized optimal transport that employs the Sinkhorn...
ykW3hvy6DL
In theorem 6.4, the results show the existence of a subsequence of weights produced by NGD and GD that have a slow margin maximization rate. Is there a dataset where if one instead considers the sequence $w'_t$ where $$ w'_t \in \arg\min_{s \leq t} \| \frac{ w*(s) }{ \| w*{s}\| } - w^*} \| $$ has a much better margin m...
ACHIEVING Margin Maximization EXPONENTIALLY FAST via PROGRESSIVE NORM RESCALING Anonymous authors Paper under double-blind review ABSTRACT In this work, we investigate the margin-maximization bias exhibited by gradient-based algorithms in classifying linearly separable data. We present an in-depth analysis of the sp...
IOrnCVIKIZ
Since the performance of LETI relies on the solution evaluator's implementation, could you elaborate on the potential biases that may arise from different evaluator designs? How might these biases impact the performance of LETI in optimizing towards certain metrics?
LETI: Learning to Generate from Textual Interactions Anonymous authors Paper under double-blind review Abstract Finetuning pre-trained language models (LMs) is essential for enhancing their capabilities and is a crucial phase in their lifecycles. Existing techniques commonly fine-tune on input-output pairs (e.g., in...
sBSC0OXEQG
Eq. (2) presents the linear combination between auto- and hetero-associative memory using the hyperparameter $b$ and $(1-b). But this notation hints at convex combination $b \in [0, 1]$, where $b=0$ implies only hetero-associative memory and $b=1$ implies only auto-associative memory. **Why do we need a full linear com...
CORRELATED DENSE ASSOCIATIVE MEMORIES Anonymous authors Paper under double-blind review ABSTRACT We introduce a novel associative memory model named Correlated Dense Associative Memory (CDAM), which integrates both auto- and hetero-association in a unified framework for continuous-valued memory patterns. Employing a...
EBUoTvVtMM
Framing existing attack terminology (“user inference”) as something new is confusing as there are already multiple works proposing the stronger threat model where either X texts of a user’s data or none were used to train the model (e.g., Song and Shmatikov). To the best of my understanding, it seems that user-level + ...
USER INFERENCE ATTACKS ON LARGE LANGUAGE MODELS Anonymous authors Paper under double-blind review ABSTRACT Fine-tuning is a common and effective method for tailoring large language models (LLMs) to specialized tasks and applications. In this paper, we study the privacy implications of fine-tuning LLMs on user data. ...
JuyFppXzh2
The authors focus on short text. While it is widely used across the industry, it will be good to demonstrate why their approach is better for short text when compared to other approaches. In a sense what makes the approach more suited for short-text?
GANDALF: LEARNING LABEL CORRELATIONS IN EXTREME MULTI-LABEL CLASSIFICATION VIA LABEL FEATURES Anonymous authors Paper under double-blind review ABSTRACT Extreme Multi-label Text Classification (XMC) involves learning a classifier that can assign an input with a subset of most relevant labels from millions of label c...