A naukri.com initiative
ML News
Arxiv
1d
359
Image Credit: Arxiv
SoftCVI: Contrastive variational inference with self-generated soft labels
- Soft Contrastive Variational Inference (SoftCVI) is introduced, allowing a family of variational objectives to be derived through a contrastive estimation framework.
- SoftCVI reframes the inference task as a contrastive estimation problem and does not require positive or negative samples.
- SoftCVI learns by sampling the variational distribution and computing ground truth soft classification labels from the unnormalized posterior itself.
- Empirical investigation shows that SoftCVI can form stable and effective objectives for Bayesian inference tasks, frequently outperforming other variational approaches.
Read Full Article
21 Likes
Arxiv
1d
117
Image Credit: Arxiv
Diffusion-based subsurface CO$_2$ multiphysics monitoring and forecasting
- Carbon capture and storage (CCS) is important for mitigating greenhouse gas emissions from industrial outputs.
- A novel subsurface multiphysics monitoring and forecasting framework using video diffusion models is proposed.
- The proposed method successfully captures complex physical phenomena related to CO2 monitoring.
- It can predict and invert subsurface elastic properties and CO2 saturation with consistency.
Read Full Article
7 Likes
Arxiv
1d
333
Image Credit: Arxiv
LLM Stability: A detailed analysis with some surprises
- LLM (large language model) practitioners commonly notice that outputs can vary for the same inputs under settings expected to be deterministic.
- A systematic investigation into the non-determinism of five LLMs configured to be deterministic was performed.
- Accuracy variations of up to 15% were observed across naturally occurring runs, with a gap of best possible performance to worst possible performance of up to 70%.
- Non-determinism in LLMs is considered essential to the efficient use of compute resources, indicating that this issue will persist.
Read Full Article
20 Likes
Arxiv
1d
318
Image Credit: Arxiv
AI in radiological imaging of soft-tissue and bone tumours: a systematic review evaluating against CLAIM and FUTURE-AI guidelines
- This systematic review evaluates Artificial Intelligence (AI) methods in radiological imaging for the diagnosis and prognosis of soft-tissue and bone tumours.
- The review highlights challenges in clinical translation and assesses the alignment of studies with the CLAIM and FUTURE-AI guidelines.
- Out of 325 evaluated articles, most studies performed moderately on CLAIM but poorly on FUTURE-AI.
- The review suggests that AI developers should focus on design, development, evaluation, and data reproducibility to improve the clinical translation of AI methods.
Read Full Article
19 Likes
Arxiv
1d
200
Image Credit: Arxiv
Learning out-of-time-ordered correlators with classical kernel methods
- Out-of-Time Ordered Correlators (OTOCs) are commonly used to study information scrambling in quantum systems.
- Directly computing OTOCs with classical computers is computationally expensive.
- A study explores the use of classical kernel methods (KMs) to accurately learn OTOCs and related quantities of local one-dimensional quantum systems.
- The proposed method can assist in evaluating OTOC functions of the parameterized quantum systems.
Read Full Article
12 Likes
Arxiv
1d
212
Image Credit: Arxiv
Says Who? Effective Zero-Shot Annotation of Focalization
- Researchers have tested the annotation of focalization in literature using large language models (LLMs).
- The study found that LLMs performed comparable to trained human annotators, achieving an average F1 score of 84.79%.
- The log probabilities output by GPT-family models reflected the difficulty of annotating specific literary excerpts.
- The research highlights the potential of LLMs for computational literary studies and insights into focalization in literature.
Read Full Article
12 Likes
Arxiv
1d
30
Image Credit: Arxiv
Mitigating Covariate Shift in Imitation Learning for Autonomous Vehicles Using Latent Space Generative World Models
- Researchers propose the use of latent space generative world models to address the covariate shift problem in autonomous driving.
- The driving policy can effectively mitigate covariate shift without requiring an excessive amount of training data by leveraging a world model during training.
- The policy learns how to recover from errors by aligning with states observed in human demonstrations during end-to-end training.
- Qualitative and quantitative results demonstrate significant improvements upon prior state of the art in closed-loop testing in the CARLA simulator.
Read Full Article
1 Like
Arxiv
1d
348
Image Credit: Arxiv
DeepLTL: Learning to Efficiently Satisfy Complex LTL Specifications for Multi-Task RL
- A new learning approach has been proposed to efficiently satisfy complex Linear Temporal Logic (LTL) specifications in multi-task reinforcement learning (RL).
- Existing approaches for satisfying LTL specifications suffer from various limitations, such as only being applicable to finite-horizon fragments of LTL, suboptimal solutions, and insufficient handling of safety constraints.
- The proposed method uses B"uchi automata to represent the semantics of LTL specifications and learns policies based on sequences of truth assignments.
- Experiments show that the approach can zero-shot satisfy a wide range of specifications, both finite- and infinite-horizon, and outperforms existing methods in terms of satisfaction probability and efficiency.
Read Full Article
20 Likes
Arxiv
1d
268
Image Credit: Arxiv
ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery
- ScienceAgentBench is a benchmark for evaluating language agents for data-driven scientific discovery.
- It aims to assess the capabilities of large language models (LLMs) in automating scientific discovery tasks.
- The benchmark includes 102 tasks extracted from peer-reviewed publications in four disciplines, with validation from subject matter experts.
- Results show that current language agents have limitations in generating code for data-driven discovery and end-to-end automation of scientific research.
Read Full Article
16 Likes
Arxiv
1d
196
Image Credit: Arxiv
MMIE: Massive Multimodal Interleaved Comprehension Benchmark for Large Vision-Language Models
- Researchers introduce MMIE, a large-scale benchmark for evaluating multimodal comprehension and generation in Large Vision-Language Models (LVLMs).
- MMIE consists of 20K curated multimodal queries covering various categories and subfields.
- The benchmark supports interleaved inputs and outputs, evaluating competencies through multiple-choice and open-ended questions.
- An automated evaluation metric with reduced bias and improved accuracy is proposed.
Read Full Article
11 Likes
Arxiv
1d
155
Image Credit: Arxiv
SensorBench: Benchmarking LLMs in Coding-Based Sensor Processing
- Effective processing, interpretation, and management of sensor data have emerged as a critical component of cyber-physical systems.
- Large Language Models (LLMs) show promising capabilities in processing sensory data, suggesting their potential for developing sensing systems.
- SensorBench, a comprehensive benchmark, is constructed to evaluate LLMs in sensor data processing.
- Results show that LLMs excel in simpler tasks but face challenges in processing compositional tasks compared to engineering experts.
Read Full Article
9 Likes
Arxiv
1d
355
Image Credit: Arxiv
PortLLM: Personalizing Evolving Large Language Models with Training-Free and Portable Model Patches
- PortLLM is a training-free framework for personalizing evolving large language models (LLMs).
- The framework creates an initial lightweight model update patch to capture domain-specific knowledge.
- PortLLM allows for the continual personalization of evolved LLMs at minimal cost.
- Experimental results show that PortLLM achieves comparable performance to fine-tuning with significant reductions in GPU memory usage.
Read Full Article
21 Likes
Arxiv
1d
7
Image Credit: Arxiv
Testing Support Size More Efficiently Than Learning Histograms
- Consider two problems about an unknown probability distribution $p$.
- The best known upper bound for problem (1) uses a general algorithm for learning the histogram of the distribution $p$.
- We show that testing can be done more efficiently than learning the histogram.
- This algorithm also provides a better solution to problem (2), producing larger lower bounds on support size than what follows from previous work.
Read Full Article
Like
Arxiv
1d
340
Image Credit: Arxiv
The Geometry of Concepts: Sparse Autoencoder Feature Structure
- Sparse autoencoders have produced dictionaries of high-dimensional vectors corresponding to the universe of concepts.
- The concept universe shows interesting structure at three levels: 'atomic' small-scale structure, 'brain' intermediate-scale structure, and 'galaxy' scale large-scale structure.
- The atomic structure contains parallelograms or trapezoids, with improved quality when global distractor directions are removed.
- The brain structure exhibits significant spatial modularity, with clusters of co-occurring features spatially clustering together more than expected.
- The large-scale structure follows a power law of eigenvalues with steepest slope in middle layers.
Read Full Article
20 Likes
Arxiv
1d
306
Image Credit: Arxiv
ADMM for Structured Fractional Minimization
- This paper introduces FADMM, the first Alternating Direction Method of Multipliers tailored for a class of structured fractional minimization problems.
- FADMM decouples the original problem into linearized proximal subproblems, featuring two variants: FADMM-D and FADMM-Q.
- The authors establish that FADMM converges to ε-approximate critical points of the problem within an oracle complexity of O(1/ε^3).
- Experiments on synthetic and real-world datasets demonstrate the effectiveness of FADMM in various applications, including sparse Fisher discriminant analysis and robust sparse recovery.
Read Full Article
18 Likes
For uninterrupted reading, download the app