Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract Noise scheduling is important for performance Optimal noise scheduling depends on task Increasing image size shifts optimal noise scheduling to noisier one Scaling input data with fixed noise schedule is a good strategy across image sizes Paper Content Why is noise scheduling important for diffusion models? Diffusion models define a noising process of data....

Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract Many tasks can be broken down into subroutines. Neural networks can achieve impressive performance on vision and language tasks, but it is not known how they do this. One possibility is that neural networks break down tasks into subroutines and compose them into an overall solution. Model pruning techniques are used to investigate this question in vision and language tasks....

Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract Paper Content Introduction Firstly, a compositional framework in which distributional meanings and linguistic structure interact meaningfully and transparently Flexible enough to accommodate compositionality beyond Frege’s bottom-up notion Local meanings of words informed by global textual context Modern machine learning methods involve top-down meaning flows DisCoCat proposed in 2008, enjoyed empirical support DisCoCirc proposed as an improvement of DisCoCat Produce text meanings from sentence meanings Meanings of words may change as text progresses Space of sentence meanings unspecified in DisCoCat Made up from relevant nouns in DisCoCirc Text circuits two-dimensional Hybrid grammar for text Phrase structure, pronominal links, phrase scope, transformational grammar rules Text diagrams from text with hybrid grammar Text circuits generative for text Characterise text circuits as essential meaning connectedness Text circuits conservatively extended to accommodate more features of language Parsing, expanding fragment of language, relationships between text circuits and other grammatical and semantic formalisms Distillation of text circuits from text Text circuits eliminate grammatical bureaucracy Model meanings in circuits with vector spaces, linear maps and tensor product Human language one-dimensional vehicle for higher-dimensional content A hybrid grammar for text Introducing a hybrid grammar that is generative and captures linguistic connectedness Developed in three steps: context-sensitive 3 grammar for simple sentences, pronominal links to identify recurring nouns and pronouns, and rules to fuse together simple sentences Uses ideas from Chomsky’s transformational phrase structure grammars, Lambek’s pregroups, discourse representation theory, and dependency grammars Does not deal with some grammatical phenomena, omits certain grammatical patterns, and only deals with part of language Combines grammar and meaning for efficient tools Preliminaries String diagrams are a graphical mathematical framework for composing input-output boxes....

Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract Cellular automata (CAs) are computational models based on local interactions of cells Graph CAs (GCAs) are similar to Graph Neural Networks (GNNs) Graph Neural CAs (GNCAs) can be trained to approximate the transition rule of any GCA Existing GNCAs are anisotropic, meaning they don’t treat related instances the same E(n)-GNCAs are proposed as a class of isotropic automata that are lightweight and can handle large graphs E(n)-GNCAs are successfully applied to three tasks: pattern formation, graph auto-encoding, and simulation of E(n)-equivariant dynamical systems Paper Content Introduction Collective intelligence is a growing area of machine learning research Collective intelligence is inspired by real biological systems Cellular automata are used to study collective intelligence and morphogenesis Neural nets have been used to learn and design CA rules Neural Cellular Automata have been applied for self-organizing systems, image generation and classification, and reinforcement learning Graph Neural Networks are universal engines for learning transition rules E(n)-GNCAs respect isometries in the state space E(n)-GNCAs are showcased on three different tasks: pattern formation, graph auto-encoding, and simulation of E(n)-equivariant dynamical systems Preliminaries and related work A graph consists of nodes and edges An adjacency matrix can be used to define a graph Each node can have a state, such as location, velocity, and features Each edge can have an attribute Graph (neural) cellular automata Graph Cellular Automata (GCA) is a triple (G, S, τ ) where G is a graph and S is a state space....

Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract Retrieval-augmented language models are powerful but expensive. Some work avoids cost by pre-encoding a text corpus into a memory. LUMEN is a hybrid between these two extremes. LUMEN outperforms pure memory and FiD on multiple question-answering tasks. LUMEN advantage increases with model size. Paper Content Introduction Retrieval-augmented language models such as Fusion-in-Decoder (Izacard & Grave, 2021) achieve strong performance on knowledge intensive tasks....

Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract Labeling systems have been proposed to help monitor work on the UN SDGs. Systems differ in their specificity and sensitivity and have systematic biases. An ensemble model that pools labeling systems can exceed the performance of all currently available systems. Paper Content Results Compare seven labeling systems on metrics and three labeled data sets Assess bias of different SDGs Assess susceptibility of labeling systems to produce false positives based on text length Assess potential of ensemble models to address limitations of individual labeling systems Sdg labeling systems differ in their sensitivity-specificity trade-offs Compared seven labeling systems to generate predicted labels for documents from three labeled data sets Data sets differ in number of words per document and number of SDGs they were evaluated for Titles data set had 63% of documents judged to contain one SDG Abstracts data set had 80....

Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract Language model-based KG embeddings are usually static and difficult to modify after deployment. A new task of editing language model-based KG embeddings is proposed. Four new datasets are built to evaluate existing models and a new model, KGEditor. KGEditor can update specific facts without affecting the rest with low training resources. Paper Content Introduction Knowledge Graphs (KGs) are multi-relational graphs with massive symbolic facts KGs can provide back-end support for knowledge-intensive tasks KG embedding approaches represent KGs in low-dimension vector spaces Traditional KG embedding models are structure-based Recent trend is to apply text descriptions with an expressive black-box model New task of editing language model-based KG embeddings proposed Three principles to evaluate performance of task: knowledge reliability, locality, efficiency Four new datasets built for EDIT and ADD sub-tasks Existing approaches suffer from limited ability to efficiently edit KG embeddings Proposed approach KGEditor can modify incorrect knowledge or add new knowledge while maintaining the others Editing factual knowledge Editable training is an early model-agnostic attempt to quickly edit a trained model....

Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract Motion artifacts can lead to incorrect diagnosis or characterization in imaging studies. A deep network can reduce the joint image-motion parameter search to a search over rigid motion parameters. The network is trained using simulated, motion-corrupted k-space data. Experiments on simulated and realistic 2D fast spin echo brain MRI show high reconstruction fidelity....

Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract State-of-the-art approaches for weather and climate modeling are based on physics-informed numerical models. Data-driven approaches based on machine learning aim to directly solve a downstream forecasting or projection task. These networks are trained using curated and homogeneous climate datasets. ClimaX is a flexible and generalizable deep learning model for weather and climate science....

Link to paper The full paper is available here. You can also find the paper on PapersWithCode here. Abstract Introduces k-planes, a white-box model for radiance fields in arbitrary dimensions Uses d choose 2 planes to represent a d-dimensional scene Planar factorization makes adding dimension-specific priors easy Linear feature decoder with a learned color basis yields similar performance as a nonlinear black-box MLP decoder Yields competitive and often state-of-the-art reconstruction fidelity with low memory usage Paper Content Introduction Recent interest in dynamic radiance fields requires 4D volume representations....