title
stringlengths
17
149
paper_url
stringlengths
42
42
authors
listlengths
1
34
type
stringclasses
3 values
primary_area
stringclasses
31 values
abstract
large_stringlengths
468
2.59k
keywords
listlengths
1
17
TL;DR
large_stringlengths
4
250
submission_number
int64
3
21.8k
arxiv_id
stringlengths
10
10
arxiv_id_source
stringclasses
2 values
embedding
listlengths
768
768
No Regrets: Investigating and Improving Regret Approximations for Curriculum Discovery
https://openreview.net/forum?id=iEeiZlTbts
[ "Alexander Rutherford", "Michael Beukman", "Timon Willi", "Bruno Lacerda", "Nick Hawes", "Jakob Nicolaus Foerster" ]
Poster
reinforcement_learning
What data or environments to use for training to improve downstream performance is a longstanding and very topical question in reinforcement learning. In particular, Unsupervised Environment Design (UED) methods have gained recent attention as their adaptive curricula promise to enable agents to be robust to in- and o...
[ "MARL", "UED", "Robotics" ]
An improved score function for unsupervised environment design in binary outcome settings, which we use to train agents for real-world tasks, and an improved adversarial evaluation protocol that assesses policy robustness.
21,766
2408.15099
title_snapshot
[ -0.011210364289581776, -0.045384109020233154, -0.00723948422819376, 0.0522133931517601, 0.05524815246462822, 0.028261512517929077, 0.018843874335289, -0.0023650138173252344, -0.00020760353072546422, -0.042588092386722565, -0.050922758877277374, 0.04427992179989815, -0.04910656809806824, -0...
Autoregressive Image Diffusion: Generation of Image Sequence and Application in MRI
https://openreview.net/forum?id=jIh4W7r0rn
[ "Guanxiong Luo", "Shoujin Huang", "Martin Uecker" ]
Poster
machine_learning_for_healthcare
Magnetic resonance imaging (MRI) is a widely used non-invasive imaging modality. However, a persistent challenge lies in balancing image quality with imaging speed. This trade-off is primarily constrained by k-space measurements, which traverse specific trajectories in the spatial Fourier domain (k-space). These measu...
[ "Autoregressive models", "Diffusion models", "Inverse problems", "Medical imaging", "MRI" ]
null
21,747
2405.14327
title_snapshot
[ -0.004525674972683191, -0.004820038564503193, -0.009375137276947498, 0.04703538119792938, 0.04252992197871208, 0.03508898243308067, 0.049210645258426666, 0.006557699758559465, -0.010362469591200352, -0.07843168079853058, 0.020184066146612167, -0.038461748510599136, -0.033746685832738876, 0...
The Implicit Bias of Gradient Descent on Separable Multiclass Data
https://openreview.net/forum?id=JlWn80mTJi
[ "Hrithik Ravi", "Clayton Scott", "Daniel Soudry", "Yutong Wang" ]
Poster
learning_theory
Implicit bias describes the phenomenon where optimization-based training algorithms, without explicit regularization, show a preference for simple estimators even when more complex estimators have equal objective values. Multiple works have developed the theory of implicit bias for binary classification under the assum...
[ "gradient descent", "multiclass classification", "hard-margin SVM", "implicit bias" ]
We prove implicit bias of gradient descent for linearly separable multiclass problems.
21,665
2411.01350
title_snapshot
[ -0.01362928468734026, -0.006378329358994961, -0.0040938518941402435, 0.031356338411569595, 0.0021120691671967506, 0.01816076971590519, 0.03348998352885246, -0.007958490401506424, -0.0124782994389534, -0.0433724969625473, -0.010930771939456463, 0.020448431372642517, -0.08164489269256592, -0...
How many classifiers do we need?
https://openreview.net/forum?id=m5dyKArVn8
[ "Hyunsuk Kim", "Liam Hodgkinson", "Ryan Theisen", "Michael W. Mahoney" ]
Poster
learning_theory
As performance gains through scaling data and/or model size experience diminishing returns, it is becoming increasingly popular to turn to ensembling, where the predictions of multiple models are combined to improve accuracy. In this paper, we provide a detailed analysis of how the disagreement and the polarization (a...
[ "ensemble", "model aggregation", "machine learning", "computer vision" ]
We develop bounds on the majority vote error that are tight enough to predict ensemble performance.
21,653
2411.00328
title_snapshot
[ -0.01323518343269825, -0.03446738421916962, -0.008921273052692413, 0.030814774334430695, -0.0011137420078739524, 0.016624102368950844, 0.02634459361433983, -0.02316446043550968, -0.053663723170757294, -0.02813912183046341, -0.014626038260757923, 0.00039368460420519114, -0.08206860721111298, ...
Learning to Reason via Program Generation, Emulation, and Search
https://openreview.net/forum?id=te6VagJf6G
[ "Nathaniel Weir", "Muhammad Khalifa", "Linlu Qiu", "Orion Weller", "Peter Clark" ]
Poster
natural_language_processing
Program synthesis with language models (LMs) has unlocked a large set of reasoning abilities; code-tuned LMs have proven adept at generating programs that solve a wide variety of algorithmic symbolic manipulation tasks (e.g. word concatenation). However, not all reasoning tasks are easily expressible as code, e.g. task...
[ "language models", "instruction tuning", "code generation", "reasoning", "program search", "program emulation" ]
We show that fine-tuning LMs to generate and then emulate the execution of programs creates models can learn new tasks via program search.
21,651
2405.16337
title_snapshot
[ -0.016291432082653046, -0.014608461409807205, -0.046342119574546814, 0.03480838984251022, 0.06110350415110588, 0.029587650671601295, 0.02509952522814274, 0.025929495692253113, -0.013543663546442986, -0.015995582565665245, -0.029490532353520393, 0.056295763701200485, -0.06960318982601166, -...
Beyond the Doors of Perception: Vision Transformers Represent Relations Between Objects
https://openreview.net/forum?id=8puv3c9CPg
[ "Michael A. Lepori", "Alexa R. Tartaglini", "Wai Keen Vong", "Thomas Serre", "Brenden Lake", "Ellie Pavlick" ]
Poster
interpretability_and_explainability
Though vision transformers (ViTs) have achieved state-of-the-art performance in a variety of settings, they exhibit surprising failures when performing tasks involving visual relations. This begs the question: how do ViTs attempt to perform tasks that require computing visual relations between objects? Prior efforts to...
[ "visual reasoning", "mechanistic interpretability", "transformers", "cognitive science" ]
We use methods from mechanistic interpretability to investigate how Vision Transformers perform an abstract visual reasoning task.
21,637
2406.15955
title_snapshot
[ -0.0029021932277828455, 0.019692862406373024, 0.019202103838324547, 0.0384954959154129, 0.013660081662237644, 0.004873357247561216, 0.029448023065924644, 0.01677437499165535, -0.028400275856256485, -0.02822250872850418, -0.05094936862587929, 0.05759909376502037, -0.06957262754440308, 0.012...
FactorizePhys: Matrix Factorization for Multidimensional Attention in Remote Physiological Sensing
https://openreview.net/forum?id=qrfp4eeZ47
[ "Jitesh Joshi", "Sos Agaian", "Youngjun Cho" ]
Poster
deep_learning_architectures
Remote photoplethysmography (rPPG) enables non-invasive extraction of blood volume pulse signals through imaging, transforming spatial-temporal data into time series signals. Advances in end-to-end rPPG approaches have focused on this transformation where attention mechanisms are crucial for feature extraction. However...
[ "Time-series estimation", "remote photo-plethysmography", "spatial-temporal attention", "non-negative matrix factorization" ]
This work introduces the Factorized Self-Attention Module, which computes multidimensional attention through nonnegative matrix factorization, and integrate it into FactorizePhys, a proposed 3D-CNN model for robust rPPG signal extraction.
21,624
2411.01542
title_snapshot
[ 0.02600422129034996, -0.008307782001793385, 0.026005923748016357, 0.001797666773200035, 0.040036726742982864, 0.04191362112760544, 0.03656867891550064, -0.009018603712320328, -0.02885274961590767, -0.045059699565172195, 0.0274586733430624, -0.028253847733139992, -0.07283543050289154, -0.00...
Multi-Group Proportional Representation in Retrieval
https://openreview.net/forum?id=BRZYhVHvSg
[ "Alex Oesterling", "Claudio Mayrink Verdun", "Alexander Glynn", "Carol Xuan Long", "Lucas Monteiro Paes", "Sajani Vithana", "Martina Cardone", "Flavio Calmon" ]
Poster
fairness
Image search and retrieval tasks can perpetuate harmful stereotypes, erase cultural identities, and amplify social disparities. Current approaches to mitigate these representational harms balance the number of retrieved items across population groups defined by a small number of (often binary) attributes. However, most...
[ "Fairness", "Proportional Representation", "Multi-Group Fairness" ]
We introduce a novel metric for ensuring multi-group proportional representation over sets of images. We apply this metric to retrieval and propose an algorithm that maximizes similarity under a multi-group proportional representation constraint.
21,620
2407.08571
title_snapshot
[ -0.0088723823428154, -0.014379017055034637, -0.020659156143665314, 0.042600829154253006, 0.009697751142084599, 0.04507756233215332, 0.004649006295949221, -0.011775585822761059, -0.05953187495470047, -0.02392813004553318, -0.0409080795943737, -0.021893104538321495, -0.05658499151468277, -0....
NeuralSteiner: Learning Steiner Tree for Overflow-avoiding Global Routing in Chip Design
https://openreview.net/forum?id=oEKFPSOWpp
[ "Ruizhi Liu", "ZhishengZeng", "Shizhe Ding", "Jingyan Sui", "Xingquan Li", "Dongbo Bu" ]
Poster
machine_learning_for_other_sciences_and_fields
Global routing plays a critical role in modern chip design. The routing paths generated by global routers often form a rectilinear Steiner tree (RST). Recent advances from the machine learning community have shown the power of learning-based route generation; however, the yielded routing paths by the existing approache...
[ "Global routing", "chip design", "neural network", "Steiner tree", "deep learning", "congestion" ]
We propose NeuralSteiner, a learning-based method to optimize overflow and wirelength simultaneously for global routing problem in chip design.
21,619
null
null
[ -0.0015334623167291284, -0.031163565814495087, -0.0003052824758924544, 0.017912186682224274, 0.03959954157471657, 0.026634497568011284, 0.0006694078911095858, -0.004412021022289991, -0.0044405171647667885, -0.05853015556931496, 0.04206811636686325, -0.05093015357851982, -0.03194849193096161,...
The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations
https://openreview.net/forum?id=eHzIwAhj06
[ "Tyler LaBonte", "John Collins Hill", "Xinchen zhang", "Vidya Muthukumar", "Abhishek Kumar" ]
Poster
fairness
Modern machine learning models are prone to over-reliance on spurious correlations, which can often lead to poor performance on minority groups. In this paper, we identify surprising and nuanced behavior of finetuned models on worst-group accuracy via comprehensive experiments on four well-established benchmarks across...
[ "spurious correlations", "group robustness", "distribution shift", "class balancing" ]
We identify surprising and nuanced behavior of finetuned models on worst-group accuracy in settings including class-balancing, model scaling, and spectral analysis.
21,606
2407.13957
title_snapshot
[ 0.00782328937202692, -0.02408536709845066, 0.029717210680246353, 0.025876978412270546, 0.028000444173812866, 0.015011069364845753, 0.05348983407020569, 0.011095574125647545, -0.029330380260944366, -0.04782083258032799, -0.01822277344763279, 0.002152385888621211, -0.07154297083616257, 0.010...
Limits of Transformer Language Models on Learning to Compose Algorithms
https://openreview.net/forum?id=x7AD0343Jz
[ "Jonathan Thomm", "Giacomo Camposampiero", "Aleksandar Terzic", "Michael Hersche", "Bernhard Schölkopf", "Abbas Rahimi" ]
Poster
natural_language_processing
We analyze the capabilities of Transformer language models in learning compositional discrete tasks. To this end, we evaluate training LLaMA models and prompting GPT-4 and Gemini on four tasks demanding to learn a composition of several discrete sub-tasks. In particular, we measure how well these models can reuse primi...
[ "Few-shot Compositional Learning", "Compositionality", "Sample Efficiency", "Algorithmic Learning", "Large Language Models", "Transformers" ]
We analyze the capabilities of Transformer language models in learning compositional discrete tasks and observe that compositional learning is very sample inefficient.
21,583
2402.05785
title_snapshot
[ 0.007754070218652487, -0.023912420496344566, -0.023575836792588234, 0.0368199422955513, 0.03268527612090111, 0.019214270636439323, 0.03543631359934807, 0.049040716141462326, -0.01727338507771492, -0.003833409631624818, -0.03125632926821709, 0.04417296126484871, -0.061405617743730545, 0.007...
Constrained Diffusion Models via Dual Training
https://openreview.net/forum?id=Es2Ey2tGmM
[ "Shervin Khalafi", "Dongsheng Ding", "Alejandro Ribeiro" ]
Poster
diffusion_based_models
Diffusion models have attained prominence for their ability to synthesize a probability distribution for a given dataset via a diffusion process, enabling the generation of new data points with high fidelity. However, diffusion processes are prone to generating samples that reflect biases in a training dataset. To add...
[ "Constrained diffusion model", "constrained optimization", "Lagrangian method", "dual algorithm" ]
null
21,568
2408.15094
title_snapshot
[ -0.02824985980987549, -0.021693652495741844, -0.035359643399715424, 0.05371713265776634, 0.06755712628364563, 0.04291357845067978, 0.009353694505989552, -0.01678231544792652, -0.0006766789592802525, -0.047797899693250656, -0.008756420575082302, 0.0015774135245010257, -0.039919376373291016, ...
Universality in Transfer Learning for Linear Models
https://openreview.net/forum?id=MhWaMOkoN3
[ "Reza Ghane", "Danil Akhtiamov", "Babak Hassibi" ]
Poster
learning_theory
We study the problem of transfer learning and fine-tuning in linear models for both regression and binary classification. In particular, we consider the use of stochastic gradient descent (SGD) on a linear model initialized with pretrained weights and using a small training data set from the target distribution. In the...
[ "Gaussian Universality", "Transfer Learning", "Linear Regression", "Binary Classification" ]
null
21,567
2410.02164
title_snapshot
[ -0.02818167954683304, -0.01791813224554062, 0.0023098888341337442, 0.01662719063460827, 0.05985357612371445, 0.03994900733232498, 0.04336069896817207, 0.013445659540593624, -0.005046280100941658, -0.027841191738843918, -0.016358738765120506, 0.017461560666561127, -0.08133388310670853, -0.0...
Gorilla: Large Language Model Connected with Massive APIs
https://openreview.net/forum?id=tBRNC6YemY
[ "Shishir G Patil", "Tianjun Zhang", "Xin Wang", "Joseph E. Gonzalez" ]
Poster
generative_models
Large Language Models (LLMs) have seen an impressive wave of advances, with models now excelling in a variety of tasks, such as mathematical reasoning and program synthesis. However, their potential to effectively use tools via API calls remains unfulfilled. This is a challenging task even for today’s state-of-the-art ...
[ "LLM", "Tool Use", "APIs", "Function Calling" ]
Teaching LLMs to use tools at scale with innvoations in finetuning (RAT) and a novel way to mesasure hallucination using AST.
21,559
2305.15334
title_snapshot
[ -0.0253200251609087, -0.0384320467710495, -0.044258058071136475, 0.035679515451192856, 0.021434778347611427, 0.016727855429053307, 0.027639251202344894, 0.027539635077118874, -0.047321055084466934, -0.002605514135211706, -0.006430189125239849, 0.02773858793079853, -0.0738355740904808, 0.00...
Stepping Forward on the Last Mile
https://openreview.net/forum?id=yCh1z6Dcto
[ "Chen Feng", "Shaojie Zhuo", "Xiaopeng Zhang", "Ramchalam Kinattinkara Ramakrishnan", "Zhaocong Yuan", "Andrew Zou Li" ]
Poster
optimization_for_deep_networks
Continuously adapting pre-trained models to local data on resource constrained edge devices is the \emph{last mile} for model deployment. However, as models increase in size and depth, backpropagation requires a large amount of memory, which becomes prohibitive for edge devices. In addition, most existing low power neu...
[ "On-device model adaptation", "Fixed-point forward gradient learning", "Low memory", "Edge devices" ]
On-device model adaptation with fixed-point forward gradient learning
21,556
2411.04036
title_snapshot
[ -0.039624687284231186, -0.025258561596274376, 0.0037947827950119972, 0.019142644479870796, 0.054488442838191986, 0.02209143526852131, 0.045207779854536057, 0.02740795537829399, -0.006787567865103483, -0.0395335778594017, 0.01563606970012188, -0.014296269975602627, -0.05287621542811394, 0.0...
Semi-Random Matrix Completion via Flow-Based Adaptive Reweighting
https://openreview.net/forum?id=XZp1uP0hh2
[ "Jonathan Kelner", "Jerry Li", "Allen Liu", "Aaron Sidford", "Kevin Tian" ]
Poster
learning_theory
We consider the well-studied problem of completing a rank-$r$, $\mu$-incoherent matrix $\mathbf{M} \in \mathbb{R}^{d \times d}$ from incomplete observations. We focus on this problem in the semi-random setting where each entry is independently revealed with probability at least $p = \frac{\textup{poly}(r, \mu, \log d)}...
[ "matrix completion", "semi-random model", "flow solver", "short-flat decomposition", "adaptive reweighting" ]
We give the first nearly-linear time algorithm for solving semi-random matrix completion to high accuracy and with noisy observations.
21,541
null
null
[ 0.000060439957451308146, -0.04892776533961296, 0.020577019080519676, 0.045746173709630966, 0.02278425171971321, 0.033588018268346786, 0.01194010116159916, 0.0011338480981066823, -0.030338408425450325, -0.07337553054094315, -0.033348266035318375, -0.04281046614050865, -0.062284015119075775, ...
Discovering plasticity rules that organize and maintain neural circuits
https://openreview.net/forum?id=nw4TWuEPGx
[ "David G Bell", "Alison Duffy", "Adrienne Fairhall" ]
Poster
neuroscience_and_cognitive_science
Intrinsic dynamics within the brain can accelerate learning by providing a prior scaffolding for dynamics aligned with task objectives. Such intrinsic dynamics would ideally self-organize and self-sustain in the face of biological noise including synaptic turnover and cell death. An example of such dynamics is the form...
[ "biologically plausible learning rules plasticity self-organization RNNs homeostasis meta-learning" ]
A supervised approach yields biologically plausible learning rules that self-organize and maintain robust representations of time within RNNs.
21,525
null
null
[ -0.01446001511067152, 0.017086226493120193, -0.014440581202507019, 0.021417371928691864, 0.023452650755643845, -0.007440534420311451, 0.04963209852576256, -0.002074189716950059, -0.0747113972902298, -0.030341902747750282, 0.013255912810564041, -0.008752530440688133, -0.05618041753768921, 0...
Stochastic Amortization: A Unified Approach to Accelerate Feature and Data Attribution
https://openreview.net/forum?id=ZdWTN2HOie
[ "Ian Connick Covert", "Chanwoo Kim", "Su-In Lee", "James Zou", "Tatsunori Hashimoto" ]
Poster
interpretability_and_explainability
Many tasks in explainable machine learning, such as data valuation and feature attribution, perform expensive computation for each data point and are intractable for large datasets. These methods require efficient approximations, and although amortizing the process by learning a network to directly predict the desired ...
[ "Amortization", "feature attribution", "data valuation", "stochastic optimization" ]
We accelerate feature and data attribution by training amortized models with noisy supervision
21,516
2401.15866
title_snapshot
[ 0.00790130253881216, -0.0030805696733295918, -0.027298251166939735, 0.07943371683359146, 0.02065533958375454, 0.04959690570831299, 0.0044367085210978985, 0.0034638666547834873, -0.00006549079989781603, -0.03804653882980347, -0.013623589649796486, 0.018118444830179214, -0.0741986557841301, ...
Unelicitable Backdoors via Cryptographic Transformer Circuits
https://openreview.net/forum?id=a560KLF3v5
[ "Andis Draguns", "Andrew Gritsevskiy", "Sumeet Ramesh Motwani", "Christian Schroeder de Witt" ]
Poster
safety_in_machine_learning
The rapid proliferation of open-source language models significantly increases the risks of downstream backdoor attacks. These backdoors can introduce dangerous behaviours during model deployment and can evade detection by conventional cybersecurity monitoring systems. In this paper, we introduce a novel class of backd...
[ "Backdoor attacks", "Transformers", "handcrafting model parameters", "cryptographic circuits" ]
We demonstrate how cryptographic backdoors can be embedded in transformer model weights and present a hardness scale for a class of backdoor detection methods.
21,509
2406.02619
title_judge
[ -0.014482774771749973, -0.022272858768701553, -0.02728956937789917, 0.05068276450037956, 0.053392957895994186, 0.0051377033814787865, 0.04007744789123535, -0.005270452704280615, -0.02183550037443638, -0.0054237074218690395, -0.015266269445419312, 0.012872117571532726, -0.023648224771022797, ...
Synatra: Turning Indirect Knowledge into Direct Demonstrations for Digital Agents at Scale
https://openreview.net/forum?id=KjNEzWRIqn
[ "Tianyue Ou", "Frank F. Xu", "Aman Madaan", "Jiarui Liu", "Robert Lo", "Abishek Sridhar", "Sudipta Sengupta", "Dan Roth", "Graham Neubig", "Shuyan Zhou" ]
Poster
natural_language_processing
LLMs can now act as autonomous agents that interact with digital environments and complete specific objectives (e.g., arranging an online meeting). However, accuracy is still far from satisfactory, partly due to a lack of large-scale, direct demonstrations for digital tasks. Obtaining supervised data from humans is cos...
[ "AI agents", "sythetic data", "web navigation" ]
we introduce a data synthesize approach for computer agents and achieve strong performance
21,508
2409.15637
title_snapshot
[ 0.01364938449114561, -0.033411066979169846, -0.011707022786140442, 0.053677625954151154, 0.05107151344418526, -0.017970014363527298, 0.03324473276734352, 0.006954388227313757, -0.028002789244055748, -0.00948344636708498, -0.05225181579589844, 0.020059971138834953, -0.06840913742780685, -0....
Learning Group Actions on Latent Representations
https://openreview.net/forum?id=HGNTcy4eEp
[ "Yinzhu Jin", "Aman Shrivastava", "Tom Fletcher" ]
Poster
deep_learning_architectures
In this work, we introduce a new approach to model group actions in autoencoders. Diverging from prior research in this domain, we propose to learn the group actions on the latent space rather than strictly on the data space. This adaptation enhances the versatility of our model, enabling it to learn a broader range of...
[ "group action", "representation learning", "image rendering" ]
A model to learn group action on latent factors, e.g. manipulating objects on 2D images.
21,500
null
null
[ 0.037136420607566833, -0.034206900745630264, -0.008172840811312199, 0.03954685479402542, 0.003505518427118659, 0.025582948699593544, 0.039496589452028275, -0.0005892976769246161, -0.006080439314246178, -0.02434413880109787, -0.028083568438887596, -0.02531077153980732, -0.0589095763862133, ...
Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks
https://openreview.net/forum?id=aVh9KRZdRk
[ "Tianyu He", "Darshil Doshi", "Aritra Das", "Andrey Gromov" ]
Oral
interpretability_and_explainability
Large language models can solve tasks that were not present in the training set. This capability is believed to be due to in-context learning and skill composition. In this work, we study the emergence of in-context learning and skill composition in a collection of modular arithmetic tasks. Specifically, we consider a ...
[ "In-Context Learning", "Grokking", "Modular Arithmetic", "Interpretability" ]
null
21,497
2406.02550
title_snapshot
[ -0.011526093818247318, 0.007442369591444731, -0.003003051970154047, 0.024932894855737686, 0.03267828747630119, 0.022479550912976265, 0.023601723834872246, 0.016523106023669243, -0.04490986093878746, 0.0009102426120080054, -0.021088464185595512, 0.02612363174557686, -0.0730055570602417, -0....
Neural Embeddings Rank: Aligning 3D latent dynamics with movements
https://openreview.net/forum?id=Hlcek7AYgP
[ "Chenggang Chen", "Zhiyu Yang", "Xiaoqin Wang" ]
Poster
neuroscience_and_cognitive_science
Aligning neural dynamics with movements is a fundamental goal in neuroscience and brain-machine interfaces. However, there is still a lack of dimensionality reduction methods that can effectively align low-dimensional latent dynamics with movements. To address this gap, we propose Neural Embeddings Rank (NER), a techni...
[ "Dimensionality reduction", "Latent dynamics", "Brain-machine interfaces", "Neural decoding", "Contrastive learning" ]
We reduce high-dimensional neural dynamics to only three dimensions and decode movements using just linear and logistic regression
21,487
null
null
[ -0.01820632629096508, -0.007219267077744007, 0.004041579551994801, 0.015023666433990002, 0.019744684919714928, 0.040815360844135284, 0.05067557469010353, -0.01790553331375122, -0.06656388938426971, -0.04908798262476921, -0.006497285328805447, -0.03443149849772453, -0.03487559035420418, 0.0...
Understanding the Limits of Vision Language Models Through the Lens of the Binding Problem
https://openreview.net/forum?id=Q5RYn6jagC
[ "Declan Iain Campbell", "Sunayana Rane", "Tyler Giallanza", "C. Nicolò De Sabbata", "Kia Ghods", "Amogh Joshi", "Alexander Ku", "Steven M Frankland", "Thomas L. Griffiths", "Jonathan D. Cohen", "Taylor Whittington Webb" ]
Poster
neuroscience_and_cognitive_science
Recent work has documented striking heterogeneity in the performance of state-of-the-art vision language models (VLMs), including both multimodal language models and text-to-image models. These models are able to describe and generate a diverse array of complex, naturalistic images, yet they exhibit surprising failures...
[ "visual reasoning", "foundation models", "multi-object reasoning", "cognitive science" ]
We show that vision language models exhibit human-like capacity constraints in multi-object visual reasoning and image generation.
21,466
2411.00238
title_snapshot
[ -0.019422948360443115, 0.03299581632018089, -0.013393738307058811, 0.023310288786888123, 0.038391366600990295, 0.0007768875220790505, 0.03715388476848602, 0.029411163181066513, -0.04604313522577286, -0.0234816987067461, -0.05362171679735184, 0.04845653474330902, -0.08713870495557785, -0.00...
Reward Machines for Deep RL in Noisy and Uncertain Environments
https://openreview.net/forum?id=Cc0ckJlJF2
[ "Andrew C Li", "Zizhao Chen", "Toryn Q. Klassen", "Pashootan Vaezipoor", "Rodrigo Toro Icarte", "Sheila A. McIlraith" ]
Poster
reinforcement_learning
Reward Machines provide an automaton-inspired structure for specifying instructions, safety constraints, and other temporally extended reward-worthy behaviour. By exposing the underlying structure of a reward function, they enable the decomposition of an RL task, leading to impressive gains in sample efficiency. Althou...
[ "Reward Machines", "LTL", "Linear Temporal Logic", "Automata", "RL", "Reinforcement Learning", "Formal Language" ]
We investigate the use of Reward Machines in deep RL under an uncertain interpretation of the domain-specific vocabulary.
21,459
2406.00120
title_snapshot
[ -0.029962969943881035, -0.011618191376328468, -0.028024839237332344, 0.06559209525585175, 0.03801950812339783, 0.05385924503207207, 0.02506513148546219, -0.003961456008255482, -0.031515803188085556, -0.015350336208939552, -0.03861911594867706, 0.0410497784614563, -0.06990589201450348, -0.0...
Nonlocal Attention Operator: Materializing Hidden Knowledge Towards Interpretable Physics Discovery
https://openreview.net/forum?id=uSKzEaj9zJ
[ "Yue Yu", "Ning Liu", "Fei Lu", "Tian Gao", "Siavash Jafarzadeh", "Stewart A Silling" ]
Spotlight
machine_learning_for_physical_sciences
Despite recent popularity of attention-based neural architectures in core AI fields like natural language processing (NLP) and computer vision (CV), their potential in modeling complex physical systems remains under-explored. Learning problems in physical systems are often characterized as discovering operators that ma...
[ "Foundation Model", "Neural Operators", "Inverse PDE Problems", "Physical Modeling" ]
null
21,440
2408.07307
title_snapshot
[ -0.020186183974146843, 0.002281036926433444, 0.006838693283498287, 0.031964246183633804, 0.023203056305646896, 0.012274732813239098, 0.00823651161044836, -0.013577158562839031, -0.03918982297182083, -0.023003926500678062, -0.016017930582165718, 0.004280895460397005, -0.04381154105067253, 0...
Sample Efficient Bayesian Learning of Causal Graphs from Interventions
https://openreview.net/forum?id=RfSvAom7sS
[ "Zihan Zhou", "Muhammad Qasim Elahi", "Murat Kocaoglu" ]
Poster
causal_inference
Causal discovery is a fundamental problem with applications spanning various areas in science and engineering. It is well understood that solely using observational data, one can only orient the causal graph up to its Markov equivalence class, necessitating interventional data to learn the complete causal graph. Most w...
[ "Causal Discovery", "Bayesian Learning", "Sample Efficiency" ]
We propose a sample efficient causal discovery algorithm that learns the causal graph in a Bayesian approach.
21,436
2410.20089
title_snapshot
[ -0.005730366799980402, -0.012829253450036049, -0.014393570832908154, 0.03817952051758766, 0.04371662065386772, 0.013476807624101639, 0.053981296718120575, -0.003995615988969803, 0.006955004297196865, -0.05517984926700592, 0.008682695217430592, 0.018003998324275017, -0.05826661363244057, -0...
Emergence of Hidden Capabilities: Exploring Learning Dynamics in Concept Space
https://openreview.net/forum?id=owuEcT6BTl
[ "Core Francisco Park", "Maya Okawa", "Andrew Lee", "Ekdeep Singh Lubana", "Hidenori Tanaka" ]
Spotlight
interpretability_and_explainability
Modern generative models demonstrate impressive capabilities, likely stemming from an ability to identify and manipulate abstract concepts underlying their training data. However, fundamental questions remain: what determines the concepts a model learns, the order in which it learns them, and its ability to manipulate ...
[ "Learning Dynamics", "Compositional Generalization", "Emergent Abilities", "Diffusion Models", "Mechanistic Interpretability" ]
We find that compositional generalization abilities of diffusion models emerge suddenly and robustly, while models might not actively exhibit this ability.
21,435
2406.19370
title_snapshot
[ -0.033589232712984085, -0.03219333663582802, -0.035326845943927765, 0.051058053970336914, 0.05458931252360344, -0.002513977000489831, 0.026660919189453125, 0.012710127979516983, -0.029711304232478142, -0.020825209096074104, -0.045341700315475464, 0.017044903710484505, -0.03487768769264221, ...
Non-asymptotic Convergence of Training Transformers for Next-token Prediction
https://openreview.net/forum?id=NfOFbPpYII
[ "Ruiquan Huang", "Yingbin Liang", "Jing Yang" ]
Poster
learning_theory
Transformers have achieved extraordinary success in modern machine learning due to their excellent ability to handle sequential data, especially in next-token prediction (NTP) tasks. However, the theoretical understanding of their performance in NTP is limited, with existing studies focusing mainly on asymptotic perfor...
[ "Transformer", "Convergence Rate", "Next-token Prediction", "Self-attention", "Implicit Bias" ]
We show fast convergence rate of training transformers for next-token prediction.
21,376
2409.17335
title_snapshot
[ -0.008698023855686188, -0.05495746061205864, 0.005134070757776499, 0.02070547640323639, 0.016080589964985847, 0.03287136182188988, 0.004526350647211075, 0.01887139491736889, -0.02673277072608471, -0.01313638873398304, -0.024123335257172585, 0.018417470157146454, -0.06629128754138947, 0.017...
Active, anytime-valid risk controlling prediction sets
https://openreview.net/forum?id=4ZH48aGD60
[ "Ziyu Xu", "Nikos Karampatziakis", "Paul Mineiro" ]
Poster
active_learning
Rigorously establishing the safety of black-box machine learning models with respect to critical risk measures is important for providing guarantees about the behavior of the model. Recently, a notion of a risk controlling prediction set (RCPS) has been introduced by Bates et. al. (JACM '24) for producing prediction se...
[ "distribution free", "conformal prediction", "e-process", "confidence sequence" ]
An extension of risk controlling prediction sets to anytime-valid and active labeling regime.
21,369
2406.10490
title_snapshot
[ -0.014761066995561123, -0.018005143851041794, -0.03376433253288269, 0.03882567584514618, 0.06241806969046593, 0.012625379487872124, 0.006815268192440271, -0.015250403434038162, -0.02631310559809208, -0.03770263493061066, -0.022809786722064018, 0.03291139006614685, -0.05485600605607033, -0....
Continual learning with the neural tangent ensemble
https://openreview.net/forum?id=qOSFiJdVkZ
[ "Ari S Benjamin", "Christian-Gernot Pehle", "Kyle Daruwalla" ]
Spotlight
online_learning
A natural strategy for continual learning is to weigh a Bayesian ensemble of fixed functions. This suggests that if a (single) neural network could be interpreted as an ensemble, one could design effective algorithms that learn without forgetting. To realize this possibility, we observe that a neural network classifier...
[ "continual learning", "catastrophic forgetting", "Bayesian ensembles", "Boosting and Ensemble Methods", "mixture of experts" ]
All network classifiers are ensembles; each edge provides a classifier. If you weigh them by their posterior probability you (almost) get SGD.
21,363
2408.17394
title_snapshot
[ -0.03550240397453308, -0.01940329186618328, -0.018487416207790375, 0.02950834296643734, 0.03185482695698738, 0.045067403465509415, 0.018703682348132133, 0.019930854439735413, -0.04923626407980919, -0.04665335640311241, -0.02924422174692154, 0.04084605351090431, -0.06411832571029663, -0.014...
What Makes and Breaks Safety Fine-tuning? A Mechanistic Study
https://openreview.net/forum?id=JEflV4nRlH
[ "Samyak Jain", "Ekdeep Singh Lubana", "Kemal Oksuz", "Tom Joy", "Philip Torr", "Amartya Sanyal", "Puneet K. Dokania" ]
Poster
interpretability_and_explainability
Safety fine-tuning helps align Large Language Models (LLMs) with human preferences for their safe deployment. To better understand the underlying factors that make models safe via safety fine-tuning, we design a synthetic data generation framework that captures salient aspects of an unsafe input by modeling the interac...
[ "Mechanistic Interpretability", "AI Safety", "Safety fine tuning", "Large Language Models" ]
null
21,355
2407.10264
title_snapshot
[ -0.001511211390607059, -0.008547951467335224, -0.021809455007314682, 0.023685919120907784, 0.05256623402237892, 0.010442705824971199, 0.04033999890089035, -0.019111962988972664, -0.018397213891148567, -0.032959070056676865, -0.03357076272368431, 0.03530624136328697, -0.06593257188796997, -...
Neglected Hessian component explains mysteries in sharpness regularization
https://openreview.net/forum?id=m6pVpdIN0y
[ "Yann Dauphin", "Atish Agarwala", "Hossein Mobahi" ]
Spotlight
other
Recent work has shown that methods that regularize second order information like SAM can improve generalization in deep learning. Seemingly similar methods like weight noise and gradient penalties often fail to provide such benefits. We investigate this inconsistency and reveal its connection to the the structure of th...
[ "sharpness", "flatness", "regularization" ]
Understanding the neglected indefinite part of the Hessian explains important phenomena in sharpness regularization
21,353
2401.10809
title_snapshot
[ -0.05661933869123459, -0.02479204349219799, 0.0051838550716638565, 0.043199025094509125, 0.03234434872865677, 0.04115241393446922, 0.036769721657037735, -0.004606206435710192, -0.05897412449121475, -0.05510982498526573, -0.003958855755627155, 0.04013023152947426, -0.03171106055378914, 0.00...
Dynamics of Supervised and Reinforcement Learning in the Non-Linear Perceptron
https://openreview.net/forum?id=doaJTihgIZ
[ "Christian Schmid", "James M Murray" ]
Poster
learning_theory
The ability of a brain or a neural network to efficiently learn depends crucially on both the task structure and the learning rule. Previous works have analyzed the dynamical equations describing learning in the relatively simplified context of the perceptron under assumptions of a student-teacher framework or a linear...
[ "Learning Dynamics", "non-linear perceptron", "supervised learning", "reinforcement learning" ]
We derive learning dynamics for a non-linear perceptron performing a binary Gaussian classification task.
21,331
2409.03749
title_snapshot
[ -0.03433443605899811, -0.004226998426020145, -0.023136412724852562, 0.025038864463567734, 0.03703612834215164, 0.022195765748620033, 0.023813584819436073, 0.030627354979515076, -0.05549445003271103, -0.02925332821905613, 0.015164069831371307, 0.020305320620536804, -0.06219180300831795, 0.0...
Semantics and Spatiality of Emergent Communication
https://openreview.net/forum?id=me1MpmENpw
[ "Rotem Ben Zion", "Boaz Carmeli", "Orr Paradise", "Yonatan Belinkov" ]
Poster
natural_language_processing
When artificial agents are jointly trained to perform collaborative tasks using a communication channel, they develop opaque goal-oriented communication protocols. Good task performance is often considered sufficient evidence that meaningful communication is taking place, but existing empirical results show that commun...
[ "emergent communication", "Lewis' games" ]
null
21,316
2411.10173
title_snapshot
[ -0.009445694275200367, -0.011077451519668102, -0.015684135258197784, 0.034333083778619766, 0.03423960879445076, -0.0005131539655849338, 0.046324651688337326, 0.0243380069732666, -0.011400888673961163, -0.053778037428855896, -0.03395715728402138, 0.008621278218925, -0.08136788755655289, 0.0...
Sample-Efficient Geometry Reconstruction from Euclidean Distances using Non-Convex Optimization
https://openreview.net/forum?id=Yu7H8ZOuI2
[ "Ipsita Ghosh", "Abiy Tasissa", "Christian Kümmerle" ]
Poster
optimization
The problem of finding suitable point embedding or geometric configurations given only Euclidean distance information of point pairs arises both as a core task and as a sub-problem in a variety of machine learning applications. In this paper, we aim to solve this problem given a minimal number of distance samples. To ...
[ "Euclidean distance geometry", "non-convex optimization", "iteratively reweighted least squares", "low-rank", "data efficiency", "convergence guarnatees", "restricted isometry property", "dual basis" ]
null
21,313
2410.16982
title_snapshot
[ -0.01835077814757824, -0.030058009549975395, 0.004330806899815798, 0.05143396183848381, 0.02641235664486885, 0.04893757775425911, 0.01429121196269989, 0.0007443542708642781, -0.03418352082371712, -0.07065653055906296, -0.023917993530631065, -0.015107777900993824, -0.04820127785205841, -0.0...
Normalization Layer Per-Example Gradients are Sufficient to Predict Gradient Noise Scale in Transformers
https://openreview.net/forum?id=S7THlpvH8i
[ "Gavia Gray", "Aman Tiwari", "Shane Bergsma", "Joel Hestness" ]
Poster
optimization_for_deep_networks
Per-example gradient norms are a vital ingredient for estimating gradient noise scale (GNS) with minimal variance. Observing the tensor contractions required to compute them, we propose a method with minimal FLOPs in 3D or greater tensor regimes by simultaneously computing the norms while computing the parameter gradie...
[ "Efficient deep learning", "gradient noise scale", "critical batch size", "language models" ]
While using a trick to compute per-example gradients efficiently we discover that normalization layers statistics predict GNS accurately.
21,305
2411.00999
title_snapshot
[ -0.027567625045776367, -0.02979525551199913, 0.01452943217009306, 0.027140432968735695, 0.034185416996479034, 0.0414084866642952, 0.03621161729097366, 0.01099227275699377, -0.021906815469264984, -0.01965252310037613, -0.008837511762976646, 0.00864908006042242, -0.04584597423672676, 0.00751...
An Analysis of Tokenization: Transformers under Markov Data
https://openreview.net/forum?id=wm9JZq7RCe
[ "Nived Rajaraman", "Jiantao Jiao", "Kannan Ramchandran" ]
Spotlight
interpretability_and_explainability
While there has been a large body of research attempting to circumvent tokenization for language modeling (Clark et al. 2022, Xue et al. 2022), the current consensus is that it is a necessary initial step for designing state-of-the-art performant language models. In this paper, we investigate tokenization from a theore...
[ "Tokenization", "LLMs", "interpretability" ]
Transformers without tokenization are very slow to learn Markov sources; we present a theoretical model and empirical observations showing that the addition of tokenization enables them to learn such processes much more efficiently
21,294
null
null
[ -0.03656494617462158, -0.022393936291337013, -0.01885519176721573, 0.034966107457876205, 0.035047851502895355, 0.04301028326153755, 0.01127239502966404, 0.023713277652859688, -0.01802304945886135, -0.0019394827540963888, -0.026349656283855438, 0.017249878495931625, -0.053452882915735245, -...
Reducing Transformer Key-Value Cache Size with Cross-Layer Attention
https://openreview.net/forum?id=M2UzLRoqic
[ "William Brandon", "Mayank Mishra", "Aniruddha Nrusimha", "Rameswar Panda", "Jonathan Ragan-Kelley" ]
Poster
deep_learning_architectures
Key-value (KV) caching plays an essential role in accelerating decoding for transformer-based autoregressive large language models (LLMs). However, the amount of memory required to store the KV cache can become prohibitive at long sequence lengths and large batch sizes. Since the invention of the transformer, two of th...
[ "transformers", "attention", "KV cache", "LLMs" ]
By sharing key and value activations between adjacent layers, we can reduce the key-value cache memory footprint of multi-query attention transformers by 2x with negligible impact on accuracy.
21,293
2405.12981
title_snapshot
[ -0.03403643146157265, -0.025083258748054504, 0.007641157601028681, 0.02708255685865879, 0.016557328402996063, 0.031407199800014496, 0.043749041855335236, 0.01969418115913868, -0.015707092359662056, -0.008511013351380825, -0.027319634333252907, 0.01959020271897316, -0.058927059173583984, 0....
Symmetry Discovery Beyond Affine Transformations
https://openreview.net/forum?id=qo7NtGMr2u
[ "Ben Shaw", "Abram Magner", "Kevin R. Moon" ]
Poster
interpretability_and_explainability
Symmetry detection has been shown to improve various machine learning tasks. In the context of continuous symmetry detection, current state of the art experiments are limited to the detection of affine transformations. Under the manifold assumption, we outline a framework for discovering continuous symmetry in data bey...
[ "symmetry detection", "isometries", "infinitesimal generators", "Killing vectors", "Riemannian metric", "transformation groups", "manifold learning" ]
null
21,287
2406.03619
title_snapshot
[ -0.021224314346909523, -0.008216550573706627, 0.021650316193699837, 0.008949026465415955, 0.01654684543609619, 0.006646016612648964, 0.038059916347265244, -0.010436013340950012, -0.02517453208565712, -0.05290912836790085, -0.012226012535393238, -0.03143417835235596, -0.06018893048167229, 0...
On the Inductive Bias of Stacking Towards Improving Reasoning
https://openreview.net/forum?id=3ZAfFoAcUI
[ "Nikunj Saunshi", "Stefani Karp", "Shankar Krishnan", "Sobhan Miryoosefi", "Sashank J. Reddi", "Sanjiv Kumar" ]
Poster
optimization_for_deep_networks
Given the increasing scale of model sizes, efficient training strategies like gradual stacking have garnered interest. Stacking enables efficient training by gradually growing the depth of a model in stages and using layers from a smaller model in an earlier stage to initialize the next stage. Although efficient for tr...
[ "stacking", "language model", "reasoning", "inductive bias", "efficient training" ]
An intriguing inductive bias of efficient training methods like stacking towards particularly improving tasks that require reasoning, despite having the same pretraining perplexity
21,267
2409.19044
title_snapshot
[ -0.03420599550008774, -0.01056729257106781, -0.02162158116698265, 0.03551609069108963, 0.02959003485739231, 0.002494462998583913, 0.051310744136571884, 0.00776038458570838, -0.04945952072739601, -0.013788106851279736, 0.0020101964473724365, 0.02248300425708294, -0.05595375597476959, 0.0058...
Model Reconstruction Using Counterfactual Explanations: A Perspective From Polytope Theory
https://openreview.net/forum?id=9uolDxbYLm
[ "Pasan Dissanayake", "Sanghamitra Dutta" ]
Poster
interpretability_and_explainability
Counterfactual explanations provide ways of achieving a favorable model outcome with minimum input perturbation. However, counterfactual explanations can also be leveraged to reconstruct the model by strategically training a surrogate model to give similar predictions as the original (target) model. In this work, we an...
[ "model extraction", "counterfactual explanations", "decision boundary shift", "query complexity" ]
We propose novel performance guarantees and strategies for leveraging counterfactual explanations in model reconstruction.
21,256
2405.05369
title_snapshot
[ -0.032004766166210175, -0.033805962651968, -0.037435293197631836, 0.060126811265945435, 0.04757757857441902, 0.03059309348464012, 0.00971285067498684, -0.010720542632043362, -0.028986100107431412, -0.038265641778707504, -0.00959363579750061, 0.048061564564704895, -0.07630197703838348, -0.0...
Inductive biases of multi-task learning and finetuning: multiple regimes of feature reuse
https://openreview.net/forum?id=UwvjJZWjPT
[ "Samuel Lippl", "Jack Lindsey" ]
Poster
learning_theory
Neural networks are often trained on multiple tasks, either simultaneously (multi-task learning, MTL) or sequentially (pretraining and subsequent finetuning, PT+FT). In particular, it is common practice to pretrain neural networks on a large auxiliary task before finetuning on a downstream task with fewer samples. Desp...
[ "multi-task learning", "implicit regularization", "finetuning", "pretraining", "implicit bias" ]
Analysis of the inductive biases associated with pre-training+finetuning and multi-task learning
21,226
2310.02396
title_snapshot
[ 0.0060838633216917515, -0.017011156305670738, 0.00089879339793697, 0.029880236834287643, 0.02898149937391281, 0.022956637665629387, 0.053330857306718826, 0.010973247699439526, -0.03967786207795143, -0.04168923571705818, 0.0003068803925998509, 0.005405987147241831, -0.06507893651723862, -0....
Communication Efficient Distributed Training with Distributed Lion
https://openreview.net/forum?id=wDirCeTIoz
[ "Bo Liu", "Lemeng Wu", "Lizhang Chen", "Kaizhao Liang", "Jiaxu Zhu", "Chen Liang", "Raghuraman Krishnamoorthi", "qiang liu" ]
Poster
infrastructure
The Lion optimizer has been a promising competitor with the AdamW for training large AI models, with advantages in memory, computation, and sample efficiency. In this paper, we introduce Distributed Lion, an innovative adaptation of Lion for distributed training environments. Leveraging the sign operator in Lion, our D...
[ "Distributed Optimization" ]
We introduce the distributed version of the Lion optimizer with efficient binary/low-precision communication. We provide both theoretical and empirical evidence to demonstrate it is a simple yet strong method.
21,219
2404.00438
title_snapshot
[ -0.011281288228929043, -0.025401746854186058, -0.010956617072224617, 0.022019585594534874, 0.022982895374298096, 0.03473590686917305, 0.030536841601133347, 0.006132020149379969, -0.019868705421686172, -0.04781295359134674, -0.02814244106411934, 0.00906884390860796, -0.07301004976034164, -0...
Aligning LLM Agents by Learning Latent Preference from User Edits
https://openreview.net/forum?id=DlYNGpCuwa
[ "Ge Gao", "Alexey Taymanov", "Eduardo Salinas", "Paul Mineiro", "Dipendra Misra" ]
Poster
natural_language_processing
We study interactive learning of language agents based on user edits made to the agent's output. In a typical setting such as writing assistants, the user interacts with a language agent to generate a response given a context, and may optionally edit the agent response to personalize it based on their latent preference...
[ "NLP", "LLM", "preference learning", "user feedback", "user edits" ]
null
21,194
2404.15269
title_snapshot
[ 0.005227379500865936, -0.00883063767105341, -0.0038590310141444206, 0.039637044072151184, 0.04600970819592476, 0.017138486728072166, 0.012996998615562916, 0.018529405817389488, 0.0001145810674643144, -0.020836051553487778, -0.03278699889779091, 0.05696655064821243, -0.058993589133024216, -...
Can Models Learn Skill Composition from Examples?
https://openreview.net/forum?id=1sLdprsbmk
[ "Haoyu Zhao", "Simran Kaur", "Dingli Yu", "Anirudh Goyal", "Sanjeev Arora" ]
Poster
generative_models
As large language models (LLMs) become increasingly advanced, their ability to exhibit compositional generalization---the capacity to combine learned skills in novel ways not encountered during training---has garnered significant attention. This type of generalization, particularly in scenarios beyond training data, is...
[ "Skill Composition", "Large Language Model" ]
null
21,185
2409.19808
title_snapshot
[ 0.0006136906449683011, 0.00006832759390817955, -0.02211652509868145, 0.041230496019124985, 0.051940739154815674, 0.005879085510969162, 0.046576421707868576, 0.030064158141613007, -0.03173544257879257, -0.014385540969669819, -0.02220636047422886, 0.03581036254763603, -0.06340303272008896, -...
Learning from Snapshots of Discrete and Continuous Data Streams
https://openreview.net/forum?id=GxwnQ8sxkL
[ "Pramith Devulapalli", "Steve Hanneke" ]
Poster
learning_theory
Imagine a smart camera trap selectively clicking pictures to understand animal movement patterns within a particular habitat. These "snapshots", or pieces of data captured from a data stream at adaptively chosen times, provide a glimpse of different animal movements unfolding through time. Learning a continuous-time pr...
[ "Learning Theory; Online Learning; Continuous Processes" ]
This paper builds a theoretical framework for non-adaptive and adaptive algorithms to predict sets of functions from continuous data streams, showing how selective querying supports accurate learning, even with limited observability.
21,178
2412.06079
title_snapshot
[ 0.006984876003116369, -0.02253405749797821, -0.03103313408792019, 0.06677959114313126, 0.051260463893413544, 0.02205502986907959, 0.041589073836803436, 0.0036165255587548018, -0.03788992390036583, -0.029972048476338387, -0.031229402869939804, -0.003476384561508894, -0.09417295455932617, -0...
Generalization Analysis for Label-Specific Representation Learning
https://openreview.net/forum?id=dtPIUXdJHY
[ "Yifan Zhang", "Min-Ling Zhang" ]
Spotlight
learning_theory
Label-specific representation learning (LSRL), i.e., constructing the representation with specific discriminative properties for each class label, is an effective strategy to improve the performance of multi-label learning. However, the generalization analysis of LSRL is still in its infancy. The existing theory bounds...
[ "Learning Theory", "Multi-Label Learning" ]
null
21,176
null
null
[ -0.018148798495531082, -0.013184426352381706, 0.001521041733212769, 0.027849333360791206, 0.019754638895392418, 0.03026387467980385, 0.03384888917207718, -0.02471492439508438, -0.03074183128774166, 0.0017276708967983723, -0.018320107832551003, 0.013658525422215462, -0.09853741526603699, 0....
Smoothie: Label Free Language Model Routing
https://openreview.net/forum?id=pPSWHsgqRp
[ "Neel Guha", "Mayee F Chen", "Trevor Chow", "Ishan S. Khare", "Christopher Re" ]
Poster
natural_language_processing
Large language models (LLMs) are increasingly used in applications where LLM inputs may span many different tasks. Recent work has found that the choice of LLM is consequential, and different LLMs may be good for different input samples. Prior approaches have thus explored how engineers might select an LLM to use for e...
[ "large language models", "weak supervision", "graphical models", "routing" ]
We propose an algorithm for learning LLM routers without labeled data.
21,167
2412.04692
title_snapshot
[ -0.007865016348659992, -0.022190354764461517, -0.027580035850405693, 0.038395386189222336, 0.055092841386795044, 0.023789290338754654, 0.021876877173781395, 0.04677395895123482, -0.018693014979362488, -0.02591036446392536, 0.00750691769644618, 0.028455374762415886, -0.07833235710859299, 0....
Identification of Analytic Nonlinear Dynamical Systems with Non-asymptotic Guarantees
https://openreview.net/forum?id=nF34qXcY0b
[ "Negin Musavi", "Ziyao Guo", "Geir Dullerud", "Yingying Li" ]
Poster
learning_theory
This paper focuses on the system identification of an important class of nonlinear systems: nonlinear systems that are linearly parameterized, which enjoy wide applications in robotics and other mechanical systems. We consider two system identification methods: least-squares estimation (LSE), which is a point estimatio...
[ "set-membership", "least-squares", "nonlinear systems", "non-asymptotic guarantees" ]
This paper generalizes the system estimation conditions for nonlinear control systems under i.i.d. inputs and provides non-asymptotic analysis.
21,147
2411.00656
title_snapshot
[ -0.04005441442131996, -0.007678781636059284, 0.00005252499977359548, 0.011227796785533428, 0.04459715262055397, 0.05137355253100395, 0.020439254119992256, -0.010959918610751629, -0.04829850047826767, -0.021509747952222824, 0.014430313371121883, -0.01710241474211216, -0.06770441681146622, -...
Bisimulation Metrics are Optimal Transport Distances, and Can be Computed Efficiently
https://openreview.net/forum?id=CSjVSnvTbG
[ "Sergio Calo", "Anders Jonsson", "Gergely Neu", "Ludovic Schwartz", "Javier Segovia-Aguas" ]
Poster
reinforcement_learning
We propose a new framework for formulating optimal transport distances between Markov chains. Previously known formulations studied couplings between the entire joint distribution induced by the chains, and derived solutions via a reduction to dynamic programming (DP) in an appropriately defined Markov decision process...
[ "Optimal transport", "Markov chains", "bisimulation metrics", "Markov decision processes" ]
null
21,144
2406.04056
title_snapshot
[ -0.05978070944547653, 0.00215118913911283, -0.015790531411767006, 0.041453033685684204, 0.06990408152341843, 0.013682113960385323, 0.02394251339137554, -0.0062642572447657585, -0.012766651809215546, -0.06944797933101654, 0.025840047746896744, -0.020192964002490044, -0.07441332936286926, 0....
Detecting Brittle Decisions for Free: Leveraging Margin Consistency in Deep Robust Classifiers
https://openreview.net/forum?id=XHCYZNmqnv
[ "Jonas Ngnawe", "Sabyasachi Sahoo", "Yann Batiste Pequignot", "Frederic Precioso", "Christian Gagné" ]
Poster
machine_vision
Despite extensive research on adversarial training strategies to improve robustness, the decisions of even the most robust deep learning models can still be quite sensitive to imperceptible perturbations, creating serious risks when deploying them for high-stakes real-world applications. While detecting such cases may ...
[ "adversarial robustness", "empirical robustness estimation", "classification", "vulnerability detection" ]
We introduce and use margin-consistency for robust deep classifiers the efficient detection of vulnerable instances to adversarial examples via the logit margin..
21,142
2406.18451
title_snapshot
[ -0.0018237695330753922, -0.03213142603635788, -0.01576894335448742, 0.04644101485610008, 0.05971672385931015, 0.022241787984967232, 0.03199700266122818, -0.012152661569416523, -0.012756919488310814, -0.05517728999257088, -0.013706152327358723, 0.01167262252420187, -0.08105062693357468, 0.0...
Schur Nets: exploiting local structure for equivariance in higher order graph neural networks
https://openreview.net/forum?id=HRnSVflpgt
[ "QINGQI ZHANG", "Ruize Xu", "Risi Kondor" ]
Poster
graph_neural_networks
Recent works have shown that extending the message passing paradigm to subgraphs communicating with other subgraphs, especially via higher order messages, can boost the expressivity of graph neural networks. In such architectures, to faithfully account for local structure such as cycles, the local operations must be eq...
[ "graph neural networks", "equivariance", "spectral graph theory", "higher order message passing" ]
We show how to build higher order GNNs that are equivariant to the automorphism groups of subgraphs without actually having to find the automorphism groups.
21,135
null
null
[ -0.004968559369444847, -0.033912766724824905, 0.0023213473614305258, 0.04570484533905983, 0.007323129568248987, 0.028974924236536026, 0.023520883172750473, 0.02747727744281292, -0.013292058371007442, -0.057208944112062454, 0.02739005908370018, -0.03135376051068306, -0.09237074106931686, 0....
Efficient Minimum Bayes Risk Decoding using Low-Rank Matrix Completion Algorithms
https://openreview.net/forum?id=8iPobEKUUA
[ "Firas Trabelsi", "David Vilar", "Mara Finkelstein", "Markus Freitag" ]
Poster
natural_language_processing
Minimum Bayes Risk (MBR) decoding is a powerful decoding strategy widely used for text generation tasks but its quadratic computational complexity limits its practical application. This paper presents a novel approach for approximating MBR decoding using matrix completion techniques, focusing on a machine translation t...
[ "Machine Translation", "Minimum Bayes Risk", "Natural Language Processing", "Low-Rank Matrix Completion" ]
null
21,129
2406.02832
title_snapshot
[ -0.01825401932001114, 0.004057107027620077, 0.0023278184235095978, 0.04895836487412453, 0.03156930208206177, 0.01880721189081669, 0.027659548446536064, 0.0036225351504981518, -0.04018288850784302, -0.028424140065908432, -0.04022594168782234, 0.030072851106524467, -0.046850208193063736, -0....
EGSST: Event-based Graph Spatiotemporal Sensitive Transformer for Object Detection
https://openreview.net/forum?id=cknAewsBhD
[ "Sheng Wu", "Hang Sheng", "Hui Feng", "Bo Hu" ]
Poster
machine_vision
Event cameras provide exceptionally high temporal resolution in dynamic vision systems due to their unique event-driven mechanism. However, the sparse and asynchronous nature of event data makes frame-based visual processing methods inappropriate. This study proposes a novel framework, Event-based Graph Spatiotemporal ...
[ "Event camera", "Graph", "Transformer", "Object detection" ]
The EGSST framework, which is proposed for the rapid and efficient processing of event camera data for object detection, integrates a Graph Transformer structure with adaptive temporal attention, inspired by the dynamic response of human eyes.
21,128
null
null
[ 0.003591674380004406, -0.0214135330170393, 0.03282870724797249, 0.025678377598524094, 0.027685921639204025, 0.03996605426073074, 0.00797246489673853, 0.02883356250822544, -0.0429423451423645, -0.026668008416891098, -0.019923867657780647, -0.014281915500760078, -0.06815861165523529, 0.01412...
The Power of Resets in Online Reinforcement Learning
https://openreview.net/forum?id=7sACcaOmGi
[ "Zakaria Mhammedi", "Dylan J Foster", "Alexander Rakhlin" ]
Spotlight
learning_theory
Simulators are a pervasive tool in reinforcement learning, but most existing algorithms cannot efficiently exploit simulator access -- particularly in high-dimensional domains that require general function approximation. We explore the power of simulators through online reinforcement learning with local simulator acces...
[ "Reinforcement learning", "learning theory", "generative model", "simulator", "coverability" ]
We show that online reinforcement learning with the power to reset to previously visited states unlocks new statistical guarantees that were previously out of reach.
21,125
2404.15417
title_snapshot
[ -0.04738548770546913, -0.022027377039194107, -0.008026334457099438, 0.07320599257946014, 0.06832750886678696, 0.01693573407828808, 0.020258959382772446, -0.021610088646411896, -0.03511518985033035, -0.03938568755984306, -0.0014636516571044922, -0.007623438723385334, -0.06691386550664902, -...
Boosting Sample Efficiency and Generalization in Multi-agent Reinforcement Learning via Equivariance
https://openreview.net/forum?id=MQIET1VfoV
[ "Joshua McClellan", "Naveed Haghani", "John Winder", "Furong Huang", "Pratap Tokekar" ]
Poster
reinforcement_learning
Multi-Agent Reinforcement Learning (MARL) struggles with sample inefficiency and poor generalization [1]. These challenges are partially due to a lack of structure or inductive bias in the neural networks typically used in learning the policy. One such form of structure that is commonly observed in multi-agent scenario...
[ "Equivariant Graph Neural Networks", "Reinforcement Learning", "Multi-agent Reinforcement Learning", "Symmetry", "generalization", "sample efficiency", "MARL" ]
We demonstrate improved sample efficiency and generalization in Multi-Agent Reinforcement Learning (MARL) via using Exploration-enhanced Equivariant Neural Networks instead of traditional function approximators such as MLPs.
21,116
2410.02581
title_snapshot
[ -0.002599879866465926, -0.026418359950184822, 0.014893288724124432, 0.04760744795203209, 0.003010011510923505, 0.020079368725419044, 0.027209162712097168, 0.010408652946352959, -0.03477197512984276, -0.059876374900341034, 0.019100448116660118, -0.02026548981666565, -0.08146386593580246, -0...
Accelerating ERM for data-driven algorithm design using output-sensitive techniques
https://openreview.net/forum?id=yW3tlSwusb
[ "Maria Florina Balcan", "Christopher Seiler", "Dravyansh Sharma" ]
Poster
learning_theory
Data-driven algorithm design is a promising, learning-based approach for beyond worst-case analysis of algorithms with tunable parameters. An important open problem is the design of computationally efficient data-driven algorithms for combinatorial algorithm families with multiple parameters. As one fixes the problem i...
[ "Learning Theory", "Data-driven Algorithm Design" ]
null
21,092
2204.03569
title_snapshot
[ -0.04378378018736839, -0.003800906939432025, -0.01530531793832779, 0.03963657096028328, 0.024160658940672874, 0.0682193860411644, 0.023773156106472015, -0.003848214168101549, -0.0021379769314080477, -0.012117978185415268, 0.004127735737711191, 0.02338089421391487, -0.05310894548892975, 0.0...
ActAnywhere: Subject-Aware Video Background Generation
https://openreview.net/forum?id=ntlFREw59A
[ "Boxiao Pan", "Zhan Xu", "Chun-Hao Paul Huang", "Krishna Kumar Singh", "Yang Zhou", "Leonidas Guibas", "Jimei Yang" ]
Poster
machine_vision
We study a novel problem to automatically generate video background that tailors to foreground subject motion. It is an important problem for the movie industry and visual effects community, which traditionally requires tedious manual efforts to solve. To this end, we propose ActAnywhere, a video diffusion model that t...
[ "Video Background Generation", "Video Generation", "Video Synthesis", "Video Editing" ]
null
21,087
2401.10822
title_snapshot
[ 0.04049377888441086, -0.026759644970297813, 0.022401688620448112, 0.023436591029167175, 0.03182874992489815, -0.0033066102769225836, 0.032751426100730896, -0.0028912180569022894, -0.029984351247549057, -0.039873890578746796, -0.02678023837506771, -0.02195226401090622, -0.036349933594465256, ...
A Compositional Atlas for Algebraic Circuits
https://openreview.net/forum?id=mXlR1FLFDc
[ "Benjie Wang", "Denis Mauá", "Guy Van den Broeck", "YooJung Choi" ]
Poster
probabilistic_methods
Circuits based on sum-product structure have become a ubiquitous representation to compactly encode knowledge, from Boolean functions to probability distributions. By imposing constraints on the structure of such circuits, certain inference queries become tractable, such as model counting and most probable configuratio...
[ "semiring", "probabilistic circuits", "logic circuits", "probabilistic inference", "algebraic" ]
We introduce a unifying framework for deriving algorithms and tractability conditions for complex compositional inference queries, such as marginal MAP, logic programming inference and causal inference.
21,077
2412.05481
title_snapshot
[ -0.012900803238153458, 0.005970000755041838, -0.023350808769464493, 0.05253366753458977, 0.041956353932619095, 0.025433551520109177, 0.017657196149230003, -0.006150845903903246, -0.010555753484368324, -0.015488616190850735, -0.0012958805309608579, 0.013560950756072998, -0.06477540731430054, ...
Low-Rank Optimal Transport through Factor Relaxation with Latent Coupling
https://openreview.net/forum?id=hGgkdFF2hR
[ "Peter Halmos", "Xinhao Liu", "Julian Gold", "Benjamin Raphael" ]
Poster
other
Optimal transport (OT) is a general framework for finding a minimum-cost transport plan, or coupling, between probability distributions, and has many applications in machine learning. A key challenge in applying OT to massive datasets is the quadratic scaling of the coupling matrix with the size of the dataset. [Forrow...
[ "Optimal Transport", "Sinkhorn", "Low-Rank", "Matrix Factorization" ]
A general framework for low rank optimal transport using a latent coupling matrix and relaxed projections.
21,075
2411.10555
title_snapshot
[ -0.03457110747694969, -0.030920330435037613, 0.020562756806612015, 0.03273491933941841, 0.035431794822216034, -0.0003044986224267632, 0.023135993629693985, -0.000043057789298472926, -0.010063707828521729, -0.06807225942611694, 0.004954056348651648, -0.007362661883234978, -0.05145327374339104...
Binding in hippocampal-entorhinal circuits enables compositionality in cognitive maps
https://openreview.net/forum?id=JO6T4rEJ32
[ "Christopher Kymn", "Sonia Mazelet", "Anthony Hitchcock Thomas", "Denis Kleyko", "Edward Paxon Frady", "Friedrich Sommer", "Bruno Olshausen" ]
Poster
neuroscience_and_cognitive_science
We propose a normative model for spatial representation in the hippocampal formation that combines optimality principles, such as maximizing coding range and spatial information per neuron, with an algebraic framework for computing in distributed representation. Spatial position is encoded in a residue number system, w...
[ "cognitive maps", "compositionality", "hippocampus", "entorhinal cortex" ]
null
21,038
2406.18808
title_snapshot
[ -0.02763984352350235, 0.04981695115566254, -0.007471108343452215, 0.017693543806672096, 0.053975947201251984, 0.021032100543379784, 0.02514265663921833, 0.016718750819563866, -0.0720280259847641, -0.057407669723033905, -0.01444521639496088, -0.03933076560497284, -0.059431042522192, 0.00515...
On Tractable $\Phi$-Equilibria in Non-Concave Games
https://openreview.net/forum?id=3CtTMF5zzM
[ "Yang Cai", "Constantinos Costis Daskalakis", "Haipeng Luo", "Chen-Yu Wei", "Weiqiang Zheng" ]
Poster
algorithmic_game_theory
While Online Gradient Descent and other no-regret learning procedures are known to efficiently converge to a coarse correlated equilibrium in games where each agent's utility is concave in their own strategy, this is not the case when utilities are non-concave -- a common scenario in machine learning applications invol...
[ "Non-Concave Games", "$\\Phi$-Equilibrium", "$\\Phi$-Regret Minimization", "Learning in Games" ]
We initiate the study of *tractable* $\Phi$-equilibria in non-concave games and examine several natural families of strategy modifications.
21,037
2403.08171
title_snapshot
[ -0.05425481125712395, -0.026629390195012093, 0.01838585175573826, 0.031833846122026443, 0.04565322399139404, 0.010981790721416473, 0.0005487423040904105, 0.013628434389829636, -0.015833403915166855, -0.06029367446899414, 0.002589937299489975, 0.02311025559902191, -0.06721857190132141, 0.00...
Better by default: Strong pre-tuned MLPs and boosted trees on tabular data
https://openreview.net/forum?id=3BNPUDvqMt
[ "David Holzmüller", "Leo Grinsztajn", "Ingo Steinwart" ]
Poster
machine_learning_for_other_sciences_and_fields
For classification and regression on tabular data, the dominance of gradient-boosted decision trees (GBDTs) has recently been challenged by often much slower deep learning methods with extensive hyperparameter tuning. We address this discrepancy by introducing (a) RealMLP, an improved multilayer perceptron (MLP), and (...
[ "tabular data", "benchmark", "default parameters", "neural networks", "deep learning", "multilayer perceptron", "gradient-boosted decision trees" ]
We propose better default parameters for boosted decision trees and improved neural networks on tabular data, evaluate them on separate large benchmarks, and show that they can achieve excellent results with moderate runtime.
21,034
2407.04491
title_snapshot
[ -0.05094035714864731, -0.02327152155339718, 0.023431630805134773, 0.058656901121139526, 0.0352913960814476, -0.014799462631344795, 0.031770654022693634, -0.015889756381511688, -0.040343623608350754, -0.032298024743795395, -0.04619878903031349, 0.01759633980691433, -0.07326150685548782, -0....
Refusal in Language Models Is Mediated by a Single Direction
https://openreview.net/forum?id=pH3XAQME6c
[ "Andy Arditi", "Oscar Balcells Obeso", "Aaquib Syed", "Daniel Paleka", "Nina Rimsky", "Wes Gurnee", "Neel Nanda" ]
Poster
safety_in_machine_learning
Conversational large language models are fine-tuned for both instruction-following and safety, resulting in models that obey benign requests but refuse harmful ones. While this refusal behavior is widespread across chat models, its underlying mechanisms remain poorly understood. In this work, we show that refusal is me...
[ "mechanistic interpretability", "refusal", "jailbreaks", "language models", "steering vectors", "representation engineering" ]
null
21,012
2406.11717
title_snapshot
[ -0.05387097969651222, -0.03426797688007355, -0.04418359324336052, 0.049338798969984055, 0.02263454534113407, 0.0164719820022583, 0.06965824216604233, -0.01577235572040081, -0.04476318135857582, 0.0034362731967121363, -0.04210876673460007, 0.03753028064966202, -0.07113588601350784, 0.006968...
Trace is the Next AutoDiff: Generative Optimization with Rich Feedback, Execution Traces, and LLMs
https://openreview.net/forum?id=rYs2Dmn9tD
[ "Ching-An Cheng", "Allen Nie", "Adith Swaminathan" ]
Poster
optimization_for_deep_networks
We study a class of optimization problems motivated by automating the design and update of AI systems like coding assistants, robots, and copilots. AutoDiff frameworks, like PyTorch, enable efficient end-to-end optimization of differentiable systems. However, general computational workflows can be non-differentiable an...
[ "Optimization", "Back-Propagation", "Automatic Differentiation", "LLM", "Language Feedback", "Execution Trace" ]
Framework for efficient optimization of heterogenous parameters in general computational workflows
20,983
2406.16218
title_snapshot
[ 0.00740857282653451, -0.02672375738620758, -0.009924614802002907, 0.02822055108845234, 0.06022288277745247, 0.02474343404173851, 0.01570742577314377, -0.009678208269178867, 0.0008990918868221343, -0.06301000714302063, -0.006588545627892017, -0.014396551996469498, -0.055741433054208755, -0....
SIRIUS : Contexual Sparisty with Correction for Efficient LLMs
https://openreview.net/forum?id=5bR2l1b2eh
[ "Yang Zhou", "Zhuoming Chen", "Zhaozhuo Xu", "Xi Victoria Lin", "Beidi Chen" ]
Poster
optimization_for_deep_networks
With the blossom of large language models (LLM), inference efficiency becomes increasingly important. Various approximate methods are proposed to reduce the cost at inference time. Contextual Sparsity (CS) is appealing for its training-free nature and its ability to reach a higher compression ratio seemingly without si...
[ "Contextual Sparsity", "LLM inference", "Knowledge Distillation" ]
null
20,976
null
null
[ -0.01400345005095005, -0.035855982452631, 0.0026823971420526505, 0.034591205418109894, 0.0560922808945179, 0.027252577245235443, 0.011849750764667988, 0.028225446119904518, -0.04472232609987259, -0.028102602809667587, -0.0129078458994627, 0.043876007199287415, -0.045804813504219055, -0.003...
Attack-Aware Noise Calibration for Differential Privacy
https://openreview.net/forum?id=hOcsUrOY0D
[ "Bogdan Kulynych", "Juan Felipe Gomez", "Georgios Kaissis", "Flavio Calmon", "Carmela Troncoso" ]
Poster
privacy
Differential privacy (DP) is a widely used approach for mitigating privacy risks when training machine learning models on sensitive data. DP mechanisms add noise during training to limit the risk of information leakage. The scale of the added noise is critical, as it determines the trade-off between privacy and utility...
[ "differential privacy", "DP-SGD" ]
Direct calibration of noise to attack risk increases utility of DP mechanisms at the same level of risk, compared to calibration of noise to a given level of epsilon.
20,974
2407.02191
title_snapshot
[ 0.006202030461281538, 0.01708999089896679, -0.025042062625288963, 0.04903065785765648, 0.04680977016687393, 0.03458027541637421, 0.04981488734483719, -0.06291592866182327, -0.01604047417640686, -0.03020574152469635, 0.03329813852906227, 0.000734731846023351, -0.056486111134290695, -0.00332...
Fairness in Social Influence Maximization via Optimal Transport
https://openreview.net/forum?id=axW8xvQPkF
[ "Shubham Chowdhary", "Giulia De Pasquale", "Nicolas Lanzetti", "Ana-Andreea Stoica", "Florian Dorfler" ]
Poster
fairness
We study fairness in social influence maximization, whereby one seeks to select seeds that spread a given information throughout a network, ensuring balanced outreach among different communities (e.g. demographic groups). In the literature, fairness is often quantified in terms of the expected outreach within individua...
[ "Fairness", "social influence maximization", "optimal transport" ]
null
20,973
2406.17736
title_snapshot
[ 0.0002015272038988769, -0.029639672487974167, 0.024125777184963226, 0.03147205337882042, 0.021856997162103653, -0.007948126643896103, 0.014650056138634682, 0.02029728889465332, -0.018298305571079254, -0.04899739474058151, 0.01029618177562952, -0.031327806413173676, -0.05465377867221832, -0...
A Critical Evaluation of AI Feedback for Aligning Large Language Models
https://openreview.net/forum?id=FZQYfmsmX9
[ "Archit Sharma", "Sedrick Keh", "Eric Mitchell", "Chelsea Finn", "Kushal Arora", "Thomas Kollar" ]
Poster
natural_language_processing
Learning from AI feedback (LAIF) is a popular paradigm for improving the instruction-following abilities of powerful pre-trained language models. LAIF first performs supervised fine-tuning (SFT) using demonstrations from a teacher model and then further fine-tunes the model with reinforcement learning (RL) or direct pr...
[ "reinforcement learning from human feedback", "ai feedback", "alignment", "direct preference optimization" ]
SFT can be just as effective as AI feedback when using strong teacher models.
20,965
2402.12366
title_snapshot
[ 0.0034141680225729942, -0.041187044233083725, 0.01819465681910515, 0.04630030319094658, 0.0399060919880867, 0.026944361627101898, 0.026043735444545746, 0.01944424770772457, -0.009021945297718048, -0.007202924229204655, -0.012357604689896107, 0.0732458308339119, -0.0649627074599266, -0.0168...
Paths to Equilibrium in Games
https://openreview.net/forum?id=LxxIiInmuF
[ "Bora Yongacoglu", "Gurdal Arslan", "Lacra Pavel", "Serdar Yuksel" ]
Spotlight
algorithmic_game_theory
In multi-agent reinforcement learning (MARL) and game theory, agents repeatedly interact and revise their strategies as new data arrives, producing a sequence of strategy profiles. This paper studies sequences of strategies satisfying a pairwise constraint inspired by policy updating in reinforcement learning, where an...
[ "game theory", "multi-agent reinforcement learning", "strategic dynamics" ]
We study a path connectivity structure of games that is relevant to strategic dynamics and iterative learning algorithms in multi-agent systems. We prove that paths to equilibrium exist in all normal-form games.
20,958
2403.18079
title_snapshot
[ -0.044946812093257904, -0.015051107853651047, -0.011225166730582714, 0.00829605758190155, 0.0458584688603878, 0.0004960076767019928, 0.024231737479567528, 0.02690637670457363, -0.03939200937747955, -0.05409562960267067, -0.01826276257634163, 0.02853906713426113, -0.05308311805129051, -0.02...
Tangent Space Causal Inference: Leveraging Vector Fields for Causal Discovery in Dynamical Systems
https://openreview.net/forum?id=Bj2CpB9Dey
[ "Kurt Butler", "Daniel Waxman", "Petar Djuric" ]
Poster
causal_inference
Causal discovery with time series data remains a challenging yet increasingly important task across many scientific domains. Convergent cross mapping (CCM) and related methods have been proposed to study time series that are generated by dynamical systems, where traditional approaches like Granger causality are unrelia...
[ "causal discovery", "convergent cross mapping", "manifolds", "dynamical systems", "differential geometry" ]
We improve convergent cross mapping by explicitly considering vector fields instead of individual predictions.
20,955
2410.23499
title_snapshot
[ -0.025295237079262733, -0.029675515368580818, -0.02306273765861988, 0.03445048630237579, 0.03607160225510597, 0.028191816061735153, 0.0408804826438427, 0.048500847071409225, -0.020340479910373688, -0.05459074303507805, 0.015577775426208973, 0.0012053558602929115, -0.04866466671228409, 0.00...
Towards Effective Planning Strategies for Dynamic Opinion Networks
https://openreview.net/forum?id=LYivxMp5es
[ "Bharath Chandra Muppasani", "Protik Nag", "Vignesh Narayanan", "Biplav Srivastava", "Michael Huhns" ]
Poster
machine_learning_for_social_sciences
In this study, we investigate the under-explored intervention planning aimed at disseminating accurate information within dynamic opinion networks by leveraging learning strategies. Intervention planning involves identifying key nodes (search) and exerting control (e.g., disseminating accurate/official information thro...
[ "Opinion networks", "Dynamic Planning", "Misinformation Spread", "Network dynamics." ]
Containing misinformation spread through learning-based approaches using Graph Neural Networks
20,952
2410.14091
title_snapshot
[ 0.016515344381332397, -0.048397138714790344, -0.018413497135043144, 0.04028360918164253, 0.020941564813256264, 0.011103925295174122, 0.01457272656261921, -0.005626131780445576, -0.014229861088097095, -0.024285610765218735, 0.02105787582695484, 0.007850926369428635, -0.08324465155601501, -0...
Universality of AdaGrad Stepsizes for Stochastic Optimization: Inexact Oracle, Acceleration and Variance Reduction
https://openreview.net/forum?id=rniiAVjHi5
[ "Anton Rodomanov", "Xiaowen Jiang", "Sebastian U Stich" ]
Poster
optimization
We present adaptive gradient methods (both basic and accelerated) for solving convex composite optimization problems in which the main part is approximately smooth (a.k.a. $(\delta, L)$-smooth) and can be accessed only via a (potentially biased) stochastic gradient oracle. This setting covers many interesting examples ...
[ "convex optimization", "stochastic optimization", "adaptive methods", "universal algorithms", "acceleration", "variance reduction", "AdaGrad", "SVRG", "weakly smooth functions", "Hölder condition", "inexact oracle", "complexity estimates" ]
null
20,951
2406.06398
title_snapshot
[ -0.03440895676612854, -0.002074338961392641, 0.011996236629784107, 0.04197688773274422, 0.043888792395591736, 0.05580956116318703, 0.050311051309108734, 0.0015339888632297516, -0.00914466567337513, -0.035589586943387985, -0.018341129645705223, -0.008954612538218498, -0.050423189997673035, ...
Proportional Fairness in Non-Centroid Clustering
https://openreview.net/forum?id=Actjv6Wect
[ "Ioannis Caragiannis", "Evi Micha", "Nisarg Shah" ]
Poster
fairness
We revisit the recently developed framework of proportionally fair clustering, where the goal is to provide group fairness guarantees that become stronger for groups of data points that are large and cohesive. Prior work applies this framework to centroid-based clustering, where points are partitioned into clusters, an...
[ "clustering", "proportional fairness", "core", "fully justified representation", "auditing" ]
null
20,946
2410.23273
title_snapshot
[ 0.0032233160454779863, -0.016819221898913383, 0.014903733506798744, 0.036888446658849716, 0.0354527086019516, 0.029400765895843506, -0.01369931735098362, -0.015102045610547066, -0.033574193716049194, -0.03371421992778778, -0.011634149588644505, -0.03662317991256714, -0.0548124723136425, -0...
Unified Insights: Harnessing Multi-modal Data for Phenotype Imputation via View Decoupling
https://openreview.net/forum?id=8B3sAX889P
[ "Qiannan Zhang", "Weishen Pan", "Zilong Bai", "Chang Su", "Fei Wang" ]
Poster
machine_learning_for_healthcare
Phenotype imputation plays a crucial role in improving comprehensive and accurate medical evaluation, which in turn can optimize patient treatment and bolster the reliability of clinical research. Despite the adoption of various techniques, multi-modal biological data, which can provide crucial insights into a patient'...
[ "Phenotype Imputation", "Graph Neural Networks", "Biological Multi-modal data" ]
null
20,939
null
null
[ -0.021363945677876472, -0.02862711437046528, -0.01974158175289631, 0.03900105133652687, 0.04799283668398857, 0.015641875565052032, 0.04315904155373573, -0.01611548848450184, -0.024590883404016495, -0.0393444187939167, 0.017895031720399857, 0.0034207741264253855, -0.08702044188976288, 0.013...
Incentivizing Quality Text Generation via Statistical Contracts
https://openreview.net/forum?id=wZgw4CrxwK
[ "Eden Saig", "Ohad Einav", "Inbal Talgam-Cohen" ]
Poster
algorithmic_game_theory
While the success of large language models (LLMs) increases demand for machine-generated text, current pay-per-token pricing schemes create a misalignment of incentives known in economics as moral hazard: Text-generating agents have strong incentive to cut costs by preferring a cheaper model over the cutting-edge one, ...
[ "Contract Theory", "Contract Design", "Moral Hazard", "Natural Language Generation", "LLM evaluation", "Hypothesis Testing" ]
We design pay-for-performance contracts that incentivize the use of high-quality LLMs for text generation.
20,935
2406.11118
title_snapshot
[ -0.016951287165284157, -0.018413806334137917, -0.030166897922754288, 0.08122579008340836, 0.035741351544857025, 0.0173900555819273, 0.012192021124064922, 0.018925294280052185, -0.010603467933833599, -0.027478637173771858, -0.037270016968250275, 0.04815584421157837, -0.05670290067791939, -0...
Dynamic Rescaling for Training GNNs
https://openreview.net/forum?id=IfZwSRpqHl
[ "Nimrah Mustafa", "Rebekka Burkholz" ]
Poster
graph_neural_networks
Graph neural networks (GNNs) with a rescale invariance, such as GATs, can be re-parameterized during optimization through dynamic rescaling of network parameters and gradients while keeping the loss invariant. In this work, we explore dynamic rescaling as a tool to influence GNN training dynamics in two key ways: i) ba...
[ "graph neural network", "rescale invariance", "generalization", "network balance" ]
null
20,930
null
null
[ -0.024327168241143227, -0.028786268085241318, 0.0030036221724003553, 0.027886779978871346, 0.03299852833151817, 0.047916702926158905, 0.027563506737351418, -0.001650497899390757, -0.04238065704703331, -0.05517726391553879, 0.0377575159072876, -0.01183759793639183, -0.07506076246500015, 0.0...
The Fairness-Quality Tradeoff in Clustering
https://openreview.net/forum?id=bUi2xECa7w
[ "Rashida Hakim", "Ana-Andreea Stoica", "Christos Papadimitriou", "Mihalis Yannakakis" ]
Poster
fairness
Fairness in clustering has been considered extensively in the past; however, the trade-off between the two objectives --- e.g., can we sacrifice just a little in the quality of the clustering to significantly increase fairness, or vice-versa? --- has rarely been addressed. We introduce novel algorithms for tracing the ...
[ "clustering", "algorithmic-fairness", "multiobjective-optimization" ]
We propose new algorithms for recovering the Pareto Front of the clustering problem with fairness considerations.
20,929
null
null
[ -0.03173234686255455, -0.019951090216636658, 0.005883748643100262, 0.04940243065357208, 0.03777014836668968, 0.03825542703270912, -0.003548062639310956, 0.01039570290595293, -0.03244810923933983, -0.041813794523477554, -0.022959409281611443, -0.0063863531686365604, -0.05762426182627678, -0...
Controlling Multiple Errors Simultaneously with a PAC-Bayes Bound
https://openreview.net/forum?id=lwpfH9wVkO
[ "Reuben Adams", "John Shawe-Taylor", "Benjamin Guedj" ]
Poster
learning_theory
Current PAC-Bayes generalisation bounds are restricted to scalar metrics of performance, such as the loss or error rate. However, one ideally wants more information-rich certificates that control the entire distribution of possible outcomes, such as the distribution of the test loss in regression, or the probabilities ...
[ "PAC-Bayes", "Generalization", "Statistical Learning Theory" ]
We prove a novel PAC-Bayes bound which provides rich information on the types of errors likely to be made at inference time.
20,923
2202.05560
title_snapshot
[ -0.035626839846372604, 0.03544527292251587, -0.0185762457549572, 0.03886778652667999, 0.04707280546426773, 0.05098563805222511, 0.03810950741171837, -0.04192367568612099, -0.026535306125879288, -0.017462490126490593, -0.03267500922083855, 0.018512139096856117, -0.06559155136346817, -0.0196...
Mutual Information Estimation via Normalizing Flows
https://openreview.net/forum?id=JiQXsLvDls
[ "Ivan Butakov", "Alexander Tolmachev", "Sofia Malanchuk", "Anna Neopryatnaya", "Alexey Frolov" ]
Poster
probabilistic_methods
We propose a novel approach to the problem of mutual information (MI) estimation via introducing a family of estimators based on normalizing flows. The estimator maps original data to the target distribution, for which MI is easier to estimate. We additionally explore the target distributions with known closed-form exp...
[ "Normalizing flows", "information theory", "mutual information" ]
Normalizing flows are used to allow for explicit mutual information estimation via closed-form expressions
20,910
2403.02187
title_snapshot
[ -0.005192322190850973, -0.014494738541543484, 0.0017020065570250154, 0.031501930207014084, 0.054145995527505875, 0.04131563752889633, 0.03281932324171066, 0.0018258688505738974, -0.01366459485143423, -0.05065666139125824, 0.021119652315974236, -0.005297592841088772, -0.06023577228188515, -...
Detecting Bugs with Substantial Monetary Consequences by LLM and Rule-based Reasoning
https://openreview.net/forum?id=hB5NkiET32
[ "Brian Zhang", "ZHUO ZHANG" ]
Poster
machine_learning_for_other_sciences_and_fields
Financial transactions are increasingly being handled by automated programs called *smart contracts*. However, one challenge in the adaptation of smart contracts is the presence of vulnerabilities, which can cause significant monetary loss. In 2024, $247.88 M was lost in 20 smart contract exploits. According to a rec...
[ "LLM", "rule based reasoning", "smart contract", "accounting bugs" ]
We develop ABAuditor, a hybrid LLM and rule-based reasoning system to detect bugs with substantial monetary consequence.
20,891
null
null
[ -0.017725087702274323, -0.011530356481671333, -0.02054310217499733, 0.025404661893844604, 0.08272477239370346, -0.01083402056246996, 0.030616050586104393, 0.0059174420312047005, -0.020481545478105545, -0.008057008497416973, -0.03584864363074303, 0.029027439653873444, -0.06774672865867615, ...
Immiscible Diffusion: Accelerating Diffusion Training with Noise Assignment
https://openreview.net/forum?id=kK23oMGe9g
[ "Yiheng Li", "Heyang Jiang", "Akio Kodaira", "Masayoshi Tomizuka", "Kurt Keutzer", "Chenfeng Xu" ]
Poster
diffusion_based_models
In this paper, we point out that suboptimal noise-data mapping leads to slow training of diffusion models. During diffusion training, current methods diffuse each image across the entire noise space, resulting in a mixture of all images at every point in the noise layer. We emphasize that this random mixture of noise-d...
[ "Diffusion Model", "Training Efficiency" ]
We propose Immiscible Diffusion to increase the training efficiency up to 3x with only one line of code by noise assignment.
20,882
2406.12303
title_snapshot
[ 0.0033094522077590227, -0.014609805308282375, -0.022770635783672333, 0.05054984614253044, 0.0326550118625164, 0.027791280299425125, 0.0088093476369977, -0.027575677260756493, -0.003954698331654072, -0.07407674938440323, 0.008965614251792431, -0.022852616384625435, -0.053265318274497986, 0....
Geometry of naturalistic object representations in recurrent neural network models of working memory
https://openreview.net/forum?id=N2RaC7LO6k
[ "Xiaoxuan Lei", "Takuya Ito", "Pouya Bashivan" ]
Poster
neuroscience_and_cognitive_science
Working memory is a central cognitive ability crucial for intelligent decision-making. Recent experimental and computational work studying working memory has primarily used categorical (i.e., one-hot) inputs, rather than ecologically-relevant, multidimensional naturalistic ones. Moreover, studies have primarily investi...
[ "Working memory", "geometry", "recurrent neural networks" ]
We found that multi-task RNN models of working memory maintain both task-relevant and irrelevant information in orthogonalized subspaces and use rotational dynamics to track past information amidst new inputs.
20,855
2411.02685
title_snapshot
[ -0.015313036739826202, 0.010802045464515686, -0.024122999981045723, 0.03341056406497955, 0.03432324901223183, 0.014387279748916626, 0.05150417238473892, 0.0482533797621727, -0.04949106648564339, -0.041027892380952835, -0.019749509170651436, -0.0015043241437524557, -0.04551679268479347, -0....
Interpreting the Weight Space of Customized Diffusion Models
https://openreview.net/forum?id=DAO2BFzMfy
[ "Amil Dravid", "Yossi Gandelsman", "Kuan-Chieh Wang", "Rameen Abdal", "Gordon Wetzstein", "Alexei A Efros", "Kfir Aberman" ]
Poster
diffusion_based_models
We investigate the space of weights spanned by a large collection of customized diffusion models. We populate this space by creating a dataset of over 60,000 models, each of which is a base model fine-tuned to insert a different person's visual identity. We model the underlying manifold of these weights as a subspace, ...
[ "Weight Space", "Model Editing", "Diffusion Models", "Latent Space", "Personalization" ]
Using a dataset of fine-tuned diffusion models, we define a subspace in diffusion model weight space that enables controllable creation of new diffusion models.
20,829
2406.09413
title_snapshot
[ -0.014397159218788147, -0.02008514665067196, 0.009596520103514194, 0.04008759185671806, 0.05068686231970787, 0.041227225214242935, 0.0057391393929719925, -0.007834735326468945, -0.00800661277025938, -0.059619709849357605, -0.014026974327862263, -0.02833111770451069, -0.06809402257204056, -...
On the Limitations of Fractal Dimension as a Measure of Generalization
https://openreview.net/forum?id=YO6GVPUrKN
[ "Charlie Tan", "Inés García-Redondo", "Qiquan Wang", "Michael M. Bronstein", "Anthea Monod" ]
Poster
learning_theory
Bounding and predicting the generalization gap of overparameterized neural networks remains a central open problem in theoretical machine learning. There is a recent and growing body of literature that proposes the framework of fractals to model optimization trajectories of neural networks, motivating generalization bo...
[ "Generalization", "Optimization", "Persistent Homology", "Fractal Dimension" ]
We experimentally and statistically show that PH dimension does not always correlate with generalization gap.
20,827
2406.02234
title_snapshot
[ -0.03257080167531967, -0.003909683786332607, 0.0001883817312773317, 0.019872285425662994, 0.03439168259501457, 0.02428063191473484, 0.04988141357898712, 0.007210494950413704, -0.031028026714920998, -0.03451147675514221, 0.013032088056206703, -0.018526315689086914, -0.06793282181024551, 0.0...
ProxyFusion: Face Feature Aggregation Through Sparse Experts
https://openreview.net/forum?id=FIs87Iro9j
[ "Bhavin Jawade", "Alexander Stone", "Deen Dayal Mohan", "Xiao Wang", "Srirangaraj Setlur", "Venu Govindaraju" ]
Poster
deep_learning_architectures
Face feature fusion is indispensable for robust face recognition, particularly in scenarios involving long-range, low-resolution media (unconstrained environments) where not all frames or features are equally informative. Existing methods often rely on large intermediate feature maps or face metadata information, makin...
[ "Feature Fusion", "Face Recognition", "Pooling" ]
Efficient face feature fusion using sparse experts for robust recognition in challenging environments.
20,824
null
null
[ 0.012271489948034286, -0.027894604951143265, 0.008542275987565517, 0.01992725022137165, 0.04044967517256737, 0.022366290912032127, 0.021882779896259308, -0.003878822550177574, -0.03973665460944176, -0.06636440008878708, 0.012744741514325142, 0.0015560232568532228, -0.09376002103090286, -0....
How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks
https://openreview.net/forum?id=ez7w0Ss4g9
[ "Etai Littwin", "Omid Saremi", "Madhu Advani", "Vimal Thilak", "Preetum Nakkiran", "Chen Huang", "Joshua M. Susskind" ]
Poster
learning_theory
Two competing paradigms exist for self-supervised learning of data representations. Joint Embedding Predictive Architectures (JEPAs) is a class of architectures in which semantically similar inputs are encoded into representations that are predictive of each other. A recent successful approach that falls under the...
[ "SSL", "JEPA" ]
We show that JEPA architectures are biased towards learning highly influential features
20,813
2407.03475
title_snapshot
[ 0.00858347862958908, 0.003175540594384074, -0.03949838876724243, 0.03501318767666817, 0.022707831114530563, 0.030334018170833588, 0.05281182751059532, -0.028644956648349762, -0.020691048353910446, -0.041269730776548386, -0.0234649907797575, -0.036721520125865936, -0.06700579077005386, 0.02...
Quasi-Bayes meets Vines
https://openreview.net/forum?id=gcpeEg88R3
[ "David Huk", "Yuanhe Zhang", "Ritabrata Dutta", "Mark Steel" ]
Poster
probabilistic_methods
Recently developed quasi-Bayesian (QB) methods \cite{fong2023martingale} proposed a stimulating change of paradigm in Bayesian computation by directly constructing the Bayesian predictive distribution through recursion, removing the need for expensive computations involved in sampling the Bayesian posterior distributi...
[ "Quasi-Bayesian", "Copula", "Vine Copula", "Nonparametric Bayesian", "density estimation" ]
We introduced the Quasi-Bayesian Vine, which combines the strength of recursive joint Bayesian predictive densities with vines for varied tasks in high-dimensions, outperforming benchmark methods.
20,808
2406.12764
title_snapshot
[ -0.01451594103127718, -0.00821402482688427, -0.00004756075577461161, 0.007348527200520039, 0.052467383444309235, 0.04716384783387184, 0.018036669120192528, -0.013740711845457554, -0.024398107081651688, -0.0598808117210865, 0.0027112921234220266, 0.01199144683778286, -0.07061832398176193, 0...
Wasserstein Gradient Boosting: A Framework for Distribution-Valued Supervised Learning
https://openreview.net/forum?id=cuO0DenqMl
[ "Takuo Matsubara" ]
Poster
probabilistic_methods
Gradient boosting is a sequential ensemble method that fits a new weaker learner to pseudo residuals at each iteration. We propose Wasserstein gradient boosting, a novel extension of gradient boosting, which fits a new weak learner to alternative pseudo residuals that are Wasserstein gradients of loss functionals of pr...
[ "Gradient Boosting; Wasserstein Gradient Flow; Uncertainty Quantification" ]
We propose a novel family of gradient boosting, Wasserstein gradient boosting, which returns a set of particles that approximates a target probability distribution assigned at each input.
20,807
2405.09536
title_snapshot
[ 0.006155260372906923, -0.04064222052693367, 0.0062468755058944225, 0.03574584051966667, 0.03446322679519653, 0.0415794663131237, 0.014200982637703419, -0.00024259381461888552, -0.012491485103964806, -0.03375067561864853, -0.019688015803694725, 0.014922263100743294, -0.0635489895939827, -0....
Safe Time-Varying Optimization based on Gaussian Processes with Spatio-Temporal Kernel
https://openreview.net/forum?id=yKvHJJE9le
[ "Jialin Li", "Marta Zagorowska", "Giulia De Pasquale", "Alisa Rupenyan", "John Lygeros" ]
Poster
active_learning
Ensuring safety is a key aspect in sequential decision making problems, such as robotics or process control. The complexity of the underlying systems often makes finding the optimal decision challenging, especially when the safety-critical system is time-varying. Overcoming the problem of optimizing an unknown time-var...
[ "Safe learning", "Bayesian optimization", "Time-varying optimization" ]
null
20,789
2409.18000
title_snapshot
[ -0.019614629447460175, 0.009034937247633934, 0.03463113680481911, 0.02744751051068306, 0.039913490414619446, 0.03633885830640793, 0.026365648955106735, 0.0014939062530174851, -0.0053369407542049885, -0.04732014611363411, -0.03360899165272713, 0.030979862436652184, -0.05846922844648361, 0.0...
Fine-grained Analysis of In-context Linear Estimation: Data, Architecture, and Beyond
https://openreview.net/forum?id=lYPAYmfQqm
[ "Yingcong Li", "Ankit Singh Rawat", "Samet Oymak" ]
Poster
optimization_for_deep_networks
Recent research has shown that Transformers with linear attention are capable of in-context learning (ICL) by implementing a linear estimator through gradient descent steps. However, the existing results on the optimization landscape apply under stylized settings where task and feature vectors are assumed to be IID and...
[ "In-context learning", "linear attention", "state-space model", "optimization", "RAG", "LoRA" ]
We study the loss landscape of in-context learning for single-layer linear attention and state-space models under general linear tasks model while delineating the effect of distributional alignments (e.g., RAG), low-rank constraints, and LoRA.
20,787
2407.10005
title_snapshot
[ 0.0013310444774106145, 0.020413102582097054, 0.002161915646865964, 0.02164074033498764, 0.014163593761622906, 0.018018843606114388, 0.03160907328128815, 0.011188766919076443, -0.014718089252710342, -0.008971082977950573, -0.025482475757598877, 0.0035113159101456404, -0.07022631913423538, -...
Multistep Distillation of Diffusion Models via Moment Matching
https://openreview.net/forum?id=C62d2nS3KO
[ "Tim Salimans", "Thomas Mensink", "Jonathan Heek", "Emiel Hoogeboom" ]
Poster
generative_models
We present a new method for making diffusion models faster to sample. The method distills many-step diffusion models into few-step models by matching conditional expectations of the clean data given noisy data along the sampling trajectory. Our approach extends recently proposed one-step methods to the multi-step case,...
[ "generative modeling", "diffusion", "distillation" ]
null
20,780
2406.04103
title_snapshot
[ 0.008538114838302135, -0.018638910725712776, -0.016097625717520714, 0.07576163858175278, 0.04296114295721054, 0.023785561323165894, 0.028919510543346405, -0.006718856748193502, 0.003287367522716522, -0.04503141716122627, 0.006972637493163347, -0.04579468443989754, -0.04883826524019241, -0....
Group and Shuffle: Efficient Structured Orthogonal Parametrization
https://openreview.net/forum?id=7EQx56YSB2
[ "Mikhail Gorbunov", "Kolya Yudin", "Vera Soboleva", "Aibek Alanov", "Alexey Naumov", "Maxim Rakhuba" ]
Poster
deep_learning_architectures
The increasing size of neural networks has led to a growing demand for methods of efficient finetuning. Recently, an orthogonal finetuning paradigm was introduced that uses orthogonal matrices for adapting the weights of a pretrained model. In this paper, we introduce a new class of structured matrices, which unifies a...
[ "Parameter-efficient finetuning", "PEFT", "orthogonal", "structured matrices", "convolutions" ]
null
20,779
2406.10019
title_snapshot
[ -0.012403479777276516, -0.030918456614017487, 0.021900011226534843, 0.016421668231487274, 0.033009134232997894, 0.04416581243276596, 0.023152871057391167, 0.011862246319651604, -0.00962747447192669, -0.05697886273264885, -0.015297908335924149, -0.003000305499881506, -0.055056046694517136, ...
On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models
https://openreview.net/forum?id=B3rZZRALhk
[ "Tariq Berrada", "Pietro Astolfi", "Melissa Hall", "Reyhane Askari Hemmat", "Yohann Benchetrit", "Marton Havasi", "Matthew J. Muckley", "Karteek Alahari", "Adriana Romero-Soriano", "Jakob Verbeek", "Michal Drozdzal" ]
Poster
diffusion_based_models
Large-scale training of latent diffusion models (LDMs) has enabled unprecedented quality in image generation. However, large-scale end-to-end training of these models is computationally costly, and hence most research focuses either on finetuning pretrained models or experiments at smaller scales. In this work we ai...
[ "Generative Models", "Generative Modeling", "Diffusion", "Latent diffusion", "Computer vision", "text-to-image diffusion" ]
null
20,774
2411.03177
title_snapshot
[ -0.010533253662288189, -0.03224269300699234, -0.01524871401488781, 0.047128766775131226, 0.05497452989220619, 0.015782508999109268, -0.0006975299911573529, -0.0101945661008358, 0.0035615083761513233, -0.03420046344399452, -0.012793267145752907, -0.01905003748834133, -0.030726682394742966, ...
Double-Ended Synthesis Planning with Goal-Constrained Bidirectional Search
https://openreview.net/forum?id=LJNqVIKSCr
[ "Kevin Yu", "Jihye Roh", "Ziang Li", "Wenhao Gao", "Runzhong Wang", "Connor W. Coley" ]
Spotlight
machine_learning_for_other_sciences_and_fields
Computer-aided synthesis planning (CASP) algorithms have demonstrated expert-level abilities in planning retrosynthetic routes to molecules of low to moderate complexity. However, current search methods assume the sufficiency of reaching arbitrary building blocks, failing to address the common real-world constraint whe...
[ "Retrosynthesis", "synthesis planning", "chemistry", "bidirectional search" ]
We propose a neural-guided bidirectional search algorithm for a new starting material-constrained formulation of synthesis planning
20,765
2407.06334
title_snapshot
[ -0.015546227805316448, 0.01558616105467081, -0.0042336503975093365, 0.012855344451963902, 0.04280677065253258, -0.001849574502557516, 0.010861699469387531, -0.00751572847366333, 0.002911269199103117, -0.052156850695610046, 0.012205353006720543, 0.007942563854157925, -0.03425374627113342, 0...
Overfitting Behaviour of Gaussian Kernel Ridgeless Regression: Varying Bandwidth or Dimensionality
https://openreview.net/forum?id=7Sh0XkN1KS
[ "Marko Medvedev", "Gal Vardi", "Nathan Srebro" ]
Poster
learning_theory
We consider the overfitting behavior of minimum norm interpolating solutions of Gaussian kernel ridge regression (i.e. kernel ridgeless regression), when the bandwidth or input dimension varies with the sample size. For fixed dimensions, we show that even with varying or tuned bandwidth, the ridgeless solution is never...
[ "Kernel ridge regression", "benign overfitting", "tempered overfitting", "Gaussian kernel" ]
null
20,763
2409.03891
title_snapshot
[ -0.05405688285827637, -0.036931052803993225, 0.05685148015618324, 0.008296707645058632, 0.04161475598812103, 0.044297996908426285, 0.021665390580892563, 0.0020534449722617865, -0.026920685544610023, -0.024438107386231422, -0.03467779606580734, 0.0344388522207737, -0.07725748419761658, 0.01...
Multi-Label Learning with Stronger Consistency Guarantees
https://openreview.net/forum?id=zAuerb1KGx
[ "Anqi Mao", "Mehryar Mohri", "Yutao Zhong" ]
Poster
learning_theory
We present a detailed study of surrogate losses and algorithms for multi-label learning, supported by $H$-consistency bounds. We first show that, for the simplest form of multi-label loss (the popular Hamming loss), the well-known consistent binary relevance surrogate suffers from a sub-optimal dependency on the number...
[ "multi-label learning", "consistency", "surrogate loss", "hamming loss", "learning theory" ]
null
20,754
2407.13746
title_snapshot
[ -0.018986498937010765, -0.005933315958827734, -0.009321634657680988, 0.061817467212677, 0.024856824427843094, 0.01743486523628235, 0.006411019712686539, -0.018334800377488136, -0.01661860942840576, -0.023618223145604134, -0.022259347140789032, 0.02632494457066059, -0.09521052241325378, -0....
LACIE: Listener-Aware Finetuning for Calibration in Large Language Models
https://openreview.net/forum?id=RnvgYd9RAh
[ "Elias Stengel-Eskin", "Peter Hase", "Mohit Bansal" ]
Poster
natural_language_processing
When answering questions, large language models (LLMs) can convey not only an answer to the question, but a level of confidence about the answer being correct. This includes explicit markers of confidence (e.g. giving a numeric confidence score) as well as implicit markers, like using an authoritative tone or elaborati...
[ "LLM calibration", "uncertainty", "question answering", "pragmatics", "listener-speaker model", "LLM confidence estimation" ]
null
20,752
2405.21028
title_judge
[ 0.0023684180341660976, -0.026641514152288437, -0.002892466727644205, 0.017593858763575554, 0.024826521053910255, 0.03380178287625313, 0.023249933496117592, 0.0280904583632946, -0.024063024669885635, -0.010701761581003666, 0.010015105828642845, 0.07383076846599579, -0.031326234340667725, -0...
Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing
https://openreview.net/forum?id=tPdJ2qHkOB
[ "Ye Tian", "Baolin Peng", "Linfeng Song", "Lifeng Jin", "Dian Yu", "Lei Han", "Haitao Mi", "Dong Yu" ]
Poster
natural_language_processing
Despite the impressive capabilities of Large Language Models (LLMs) on various tasks, they still struggle with scenarios that involves complex reasoning and planning. Self-correction and self-learning emerge as viable solutions, employing strategies that allow LLMs to refine their outputs and learn from self-assessed r...
[ "self-improving", "search", "large language models" ]
null
20,740
2404.12253
title_snapshot
[ -0.03283049911260605, -0.0162674430757761, -0.006977610290050507, 0.029381470754742622, 0.07154574990272522, 0.03776082769036293, 0.02858494222164154, 0.001902503427118063, -0.02612406760454178, -0.002657563192769885, -0.029107745736837387, 0.030335308983922005, -0.04652857780456543, -0.01...