Deep neural networks excel at finding hierarchical representations that solve complex tasks over large datasets. How can we humans understand these learned representations? In this work, we present network dissection, an analytic framework to systematically identify the semantics of individual hidden units within image classification and image generation networks. First, we analyze a convolutional neural network (CNN) trained on scene classification and discover units that match a diverse set of object concepts. We find evidence that the network has learned many object classes that play crucial roles in classifying scene classes. Second, we use a similar analytic method to analyze a generative adversarial network (GAN) model trained to generate scenes. By analyzing changes made when small sets of units are activated or deactivated, we find that objects can be added and removed from the output scenes while adapting to the context. Finally, we apply our analytic framework to understanding adversarial attacks and to semantic image editing. <<<

The correct use of model evaluation, model selection, and algorithm selection techniques is vital in academic machine learning research as well as in many industrial settings. This article reviews different techniques that can be used for each of these three subtasks and discusses the main advantages and disadvantages of each technique with references to theoretical and empirical studies. Further, recommendations are given to encourage best yet feasible practices in research and applications of machine learning. Common methods such as the holdout method for model evaluation and selection are covered, which are not recommended when working with small datasets. Different flavors of the bootstrap technique are introduced for estimating the uncertainty of performance estimates, as an alternative to confidence intervals via normal approximation if bootstrapping is computationally feasible. Common cross-validation techniques such as leave-one-out cross-validation and k-fold cross-validation are reviewed, the bias-variance trade-off for choosing k is discussed, and practical tips for the optimal choice of k are given based on empirical evidence. Different statistical tests for algorithm comparisons are presented, and strategies for dealing with multiple comparisons such as omnibus tests and multiple-comparison corrections are discussed. Finally, alternative methods for algorithm selection, such as the combined F-test 5x2 cross-validation and nested cross-validation, are recommended for comparing machine learning algorithms when datasets are small. <<<

This document aims to be a self-contained, mathematically precise overview of transformer architectures and algorithms (*not* results). It covers what transformers are, how they are trained, what they are used for, their key architectural components, and a preview of the most prominent models. The reader is assumed to be familiar with basic ML terminology and simpler neural network architectures such as MLPs. <<<

Discovering causal structure among a set of variables is a fundamental problem in many empirical sciences. Traditional score-based casual discovery methods rely on various local heuristics to search for a Directed Acyclic Graph (DAG) according to a predefined score function. While these methods, e.g., greedy equivalence search, may have attractive results with infinite samples and certain model assumptions, they are usually less satisfactory in practice due to finite data and possible violation of assumptions. Motivated by recent advances in neural combinatorial optimization, we propose to use Reinforcement Learning (RL) to search for the DAG with the best scoring. Our encoder-decoder model takes observable data as input and generates graph adjacency matrices that are used to compute rewards. The reward incorporates both the predefined score function and two penalty terms for enforcing acyclicity. In contrast with typical RL applications where the goal is to learn a policy, we use RL as a search strategy and our final output would be the graph, among all graphs generated during training, that achieves the best reward. We conduct experiments on both synthetic and real datasets, and show that the proposed approach not only has an improved search ability but also allows a flexible score function under the acyclicity constraint. <<<

The highest accuracy object detectors to date are based on a two-stage approach popularized by R-CNN, where a classifier is applied to a sparse set of candidate object locations. In contrast, one-stage detectors that are applied over a regular, dense sampling of possible object locations have the potential to be faster and simpler, but have trailed the accuracy of two-stage detectors thus far. In this paper, we investigate why this is the case. We discover that the extreme foreground-background class imbalance encountered during training of dense detectors is the central cause. We propose to address this class imbalance by reshaping the standard cross entropy loss such that it down-weights the loss assigned to well-classified examples. Our novel Focal Loss focuses training on a sparse set of hard examples and prevents the vast number of easy negatives from overwhelming the detector during training. To evaluate the effectiveness of our loss, we design and train a simple dense detector we call RetinaNet. Our results show that when trained with the focal loss, RetinaNet is able to match the speed of previous one-stage detectors while surpassing the accuracy of all existing state-of-the-art two-stage detectors. Code is at: this https URL. <<<

Genome instability and aberrant alterations of transcriptional programs both play important roles in cancer. Single-cell RNA sequencing (scRNA-seq) has the potential to investigate both genetic and nongenetic sources of tumor heterogeneity in a single assay. Here we present a computational method, Numbat, that integrates haplotype information obtained from population-based phasing with allele and expression signals to enhance detection of copy number variations from scRNA-seq. Numbat exploits the evolutionary relationships between subclones to iteratively infer single-cell copy number profiles and tumor clonal phylogeny. Analysis of 22 tumor samples, including multiple myeloma, gastric, breast and thyroid cancers, shows that Numbat can reconstruct the tumor copy number profile and precisely identify malignant cells in the tumor microenvironment. We identify genetic subpopulations with transcriptional signatures relevant to tumor progression and therapy resistance. Numbat requires neither sample-matched DNA data nor a priori genotyping, and is applicable to a wide range of experimental settings and cancer types. <<<

Identification of alleles that improve crop production and lead to higher-yielding varieties are needed for food security. Here we show that the quantitative trait locus WFP (WEALTHY FARMER'S PANICLE) encodes OsSPL14 (SQUAMOSA PROMOTER BINDING PROTEIN-LIKE 14, also known as IPA1). Higher expression of OsSPL14 in the reproductive stage promotes panicle branching and higher grain yield in rice. OsSPL14 controls shoot branching in the vegetative stage and is affected by microRNA excision. We also demonstrate the feasibility of using the OsSLP14(WFP) allele to increase rice crop yield. Introduction of the high-yielding OsSPL14(WFP) allele into the standard rice variety Nipponbare resulted in increased rice production. <<<

AbstractIn this paper, we present a deep learning method, DDMReg, for accurate registration between diffusion MRI (dMRI) datasets. In dMRI registration, the goal is to spatially align brain anatomical structures while ensuring that local fiber orientations remain consistent with the underlying white matter fiber tract anatomy. DDMReg is a novel method that uses joint whole-brain and tract-specific information for dMRI registration. Based on the successful VoxelMorph framework for image registration, we propose a novel registration architecture that leverages not only whole brain information but also tract-specific fiber orientation information. DDMReg is an unsupervised method for deformable registration between pairs of dMRI datasets: it does not require nonlinearly pre-registered training data or the corresponding deformation fields as ground truth. We perform comparisons with four state-of-the-art registration methods on multiple independently acquired datasets from different populations (including teenagers, young and elderly adults) and different imaging protocols and scanners. We evaluate the registration performance by assessing the ability to align anatomically corresponding brain structures and ensure fiber spatial agreement between different subjects after registration. Experimental results show that DDMReg obtains significantly improved registration performance compared to the state-of-the-art methods. Importantly, we demonstrate successful generalization of DDMReg to dMRI data from different populations with varying ages and acquired using different acquisition protocols and different scanners. <<<

Transcranial magnetic stimulation (TMS) is used to treat multiple psychiatric and neurological conditions by manipulating activity in particular brain networks and circuits, but individual responses are highly variable. In clinical settings, TMS coil placement is typically based on either group average functional maps or scalp heuristics. Here, we found that this approach can inadvertently target different functional networks in depressed patients due to variability in their functional brain organization. More precise TMS targeting should be feasible by accounting for each patient's unique functional neuroanatomy. To this end, we developed a targeting approach, termed targeted functional network stimulation (TANS). The TANS approach improved stimulation specificity in silico in 8 highly sampled patients with depression and 6 healthy individuals and in vivo when targeting somatomotor functional networks representing the upper and lower limbs. Code for implementing TANS and an example dataset are provided as a resource. <<<

Recent methods for embodied instruction following are typically trained end-to-end using imitation learning. This often requires the use of expert trajectories and low-level language instructions. Such approaches assume that neural states will integrate multimodal semantics to perform state tracking, building spatial memory, exploration, and long-term planning. In contrast, we propose a modular method with structured representations that (1) builds a semantic map of the scene and (2) performs exploration with a semantic search policy, to achieve the natural language goal. Our modular method achieves SOTA performance (24.46 %) with a substantial (8.17 % absolute) gap from previous work while using less data by eschewing both expert trajectories and low-level instructions. Leveraging low-level language, however, can further increase our performance (26.49 %). Our findings suggest that an explicit spatial memory and a semantic search policy can provide a stronger and more general representation for state-tracking and guidance, even in the absence of expert trajectories or low-level instructions. <<<

Alessa R Ringel, Quentin Szabo, Andrea M Chiariello, Konrad Chudzik, Robert Schöpflin, Patricia Rothe, Alexandra L Mattei, Tobias Zehnder, Dermot Harnett, Verena Laupert, Simona Bianco, Sara Hetzel, Juliane Glaser, Mai H Q Phan, Magdalena Schindler, Daniel M Ibrahim, Christina Paliou, Andrea Esposito, Cesar A Prada-Medina, Stefan A Haas, Peter Giere, Martin Vingron, Lars Wittler, Alexander Meissner, Mario Nicodemi, Giacomo Cavalli, Frédéric Bantignies, Stefan Mundlos, Michael I Robson <<<

Regulatory landscapes drive complex developmental gene expression, but it remains unclear how their integrity is maintained when incorporating novel genes and functions during evolution. Here, we investigated how a placental mammal-specific gene, Zfp42, emerged in an ancient vertebrate topologically associated domain (TAD) without adopting or disrupting the conserved expression of its gene, Fat1. In ESCs, physical TAD partitioning separates Zfp42 and Fat1 with distinct local enhancers that drive their independent expression. This separation is driven by chromatin activity and not CTCF/cohesin. In contrast, in embryonic limbs, inactive Zfp42 shares Fat1's intact TAD without responding to active Fat1 enhancers. However, neither Fat1 enhancer-incompatibility nor nuclear envelope-attachment account for Zfp42's unresponsiveness. Rather, Zfp42's promoter is rendered inert to enhancers by context-dependent DNA methylation. Thus, diverse mechanisms enabled the integration of independent Zfp42 regulation in the Fat1 locus. Critically, such regulatory complexity appears common in evolution as, genome wide, most TADs contain multiple independently expressed genes. <<<

Triangle completion is a task widely used to study human path integration, an important navigation method relying on idiothetic cues. Systematic biases (compression patterns in the inbound responses) have been well documented in human triangle completion. However, the sources of systematic biases remain controversial. We used cross-validation modeling to compare three plausible theoretical models that assume that systematic errors occur in the encoding outbound path solely (encoding-error model), executing the inbound responses solely (execution-error model), and both (bi-component model), respectively. The data for cross-validation modeling are from a previous study (Qi et al., 2021), in which participants learned three objects’ locations (one at the path origin, that is, home) very well before walking each outbound path and then pointed to the objects’ original locations after walking the outbound path. The modeling algorithm used one inbound response (i.e., response to the home) or multiple inbound responses (i.e., responses to two non-home locations and the home) for each outbound path. The algorithm of using multiple inbound responses demonstrated that the bi-component model outperformed the other models in accounting for the systematic errors. This finding suggests that both encoding the outbound path and executing the inbound responses contribute to the systematic biases in human path integration. In addition, the results showed that the algorithm using only the home response could not distinguish among these three models, suggesting that the typical triangle-completion task with only the home response for each outbound path cannot determine the sources of the systematic biases. <<<

Time series constitute a challenging data type for machine learning algorithms, due to their highly variable lengths and sparse labeling in practice. In this paper, we tackle this challenge by proposing an unsupervised method to learn universal embeddings of time series. Unlike previous works, it is scalable with respect to their length and we demonstrate the quality, transferability and practicability of the learned representations with thorough experiments and comparisons. To this end, we combine an encoder based on causal dilated convolutions with a novel triplet loss employing time-based negative sampling, obtaining general-purpose representations for variable length and multivariate time series. <<<

Active inference offers a principled account of behavior as minimizing average sensory surprise over time. Applications of active inference to control problems have heretofore tended to focus on finite-horizon or discounted-surprise problems, despite deriving from the infinite-horizon, average-surprise imperative of the free-energy principle. Here we derive an infinite-horizon, average-surprise formulation of active inference from optimal control principles. Our formulation returns to the roots of active inference in neuroanatomy and neurophysiology, formally reconnecting active inference to optimal feedback control. Our formulation provides a unified objective functional for sensorimotor control and allows for reference states to vary over time. <<<

The brain regulates the body by anticipating its needs and attempting to meet them before they arise - a process called allostasis. Allostasis requires a model of the changing sensory conditions within the body, a process called interoception. In this paper, we examine how interoception may provide performance feedback for allostasis. We suggest studying allostasis in terms of control theory, reviewing control theory's applications to related issues in physiology, motor control, and decision making. We synthesize these by relating them to the important properties of allostatic regulation as a control problem. We then sketch a novel formalism for how the brain might perform allostatic control of the viscera by analogy to skeletomotor control, including a mathematical view on how interoception acts as performance feedback for allostasis. Finally, we suggest ways to test implications of our hypotheses. <<<

Probabilistic programming languages (PPLs) are an expressive means of representing and reasoning about probabilistic models. The computational challenge of probabilistic inference remains the primary roadblock for applying PPLs in practice. Inference is fundamentally hard, so there is no one-size-fits all solution. In this work, we target scalable inference for an important class of probabilistic programs: those whose probability distributions are discrete . Discrete distributions are common in many fields, including text analysis, network verification, artificial intelligence, and graph analysis, but they prove to be challenging for existing PPLs. We develop a domain-specific probabilistic programming language called Dice that features a new approach to exact discrete probabilistic program inference. Dice exploits program structure in order to factorize inference, enabling us to perform exact inference on probabilistic programs with hundreds of thousands of random variables. Our key technical contribution is a new reduction from discrete probabilistic programs to weighted model counting (WMC). This reduction separates the structure of the distribution from its parameters, enabling logical reasoning tools to exploit that structure for probabilistic inference. We (1) show how to compositionally reduce Dice inference to WMC, (2) prove this compilation correct with respect to a denotational semantics, (3) empirically demonstrate the performance benefits over prior approaches, and (4) analyze the types of structure that allow Dice to scale to large probabilistic programs. <<<

The first genomic scar-based homologous recombination deficiency (HRD) measures were produced using SNP arrays. As array-based technology has been largely replaced by next generation sequencing approaches, it has become important to develop algorithms that derive the same type of genomic scar scores from next generation sequencing (whole exome "WXS", whole genome "WGS") data. In order to perform this analysis, we introduce here the scarHRD R package and show that using this method the SNP array-based and next generation sequencing-based derivation of HRD scores show good correlation (Pearson correlation between 0.73 and 0.87 depending on the actual HRD measure) and that the NGS-based HRD scores distinguish similarly well between BRCA mutant and BRCA wild-type cases in a cohort of triple-negative breast cancer patients of the TCGA data set. <<<

Recent technological advancements have enabled spatially resolved transcriptomic profiling but at multi-cellular pixel resolution, thereby hindering the identification of cell-type-specific spatial patterns and gene expression variation. To address this challenge, we develop STdeconvolve as a reference-free approach to deconvolve underlying cell types comprising such multi-cellular pixel resolution spatial transcriptomics (ST) datasets. Using simulated as well as real ST datasets from diverse spatial transcriptomics technologies comprising a variety of spatial resolutions such as Spatial Transcriptomics, 10X Visium, DBiT-seq, and Slide-seq, we show that STdeconvolve can effectively recover cell-type transcriptional profiles and their proportional representation within pixels without reliance on external single-cell transcriptomics references. STdeconvolve provides comparable performance to existing reference-based methods when suitable single-cell references are available, as well as potentially superior performance when suitable single-cell references are not available. STdeconvolve is available as an open-source R software package with the source code available at https://github.com/JEFworks-Lab/STdeconvolve . <<<

Sachet A Shukla , Michael S Rooney , Mohini Rajasagi , Grace Tiao , Philip M Dixon , Michael S Lawrence , Jonathan Stevens , William J Lane , Jamie L Dellagatta , Scott Steelman , Carrie Sougnez , Kristian Cibulskis , Adam Kiezun , Nir Hacohen , Vladimir Brusic , Catherine J Wu , Gad Getz <<<

Detection of somatic mutations in human leukocyte antigen (HLA) genes using whole-exome sequencing (WES) is hampered by the high polymorphism of the HLA loci, which prevents alignment of sequencing reads to the human reference genome. We describe a computational pipeline that enables accurate inference of germline alleles of class I HLA-A, B and C genes and subsequent detection of mutations in these genes using the inferred alleles as a reference. Analysis of WES data from 7,930 pairs of tumor and healthy tissue from the same patient revealed 298 nonsilent HLA mutations in tumors from 266 patients. These 298 mutations are enriched for likely functional mutations, including putative loss-of-function events. Recurrence of mutations suggested that these 'hotspot' sites were positively selected. Cancers with recurrent somatic HLA mutations were associated with upregulation of signatures of cytolytic activity characteristic of tumor infiltration by effector lymphocytes, supporting immune evasion by altered HLA function as a contributory mechanism in cancer. <<<

We introduce PyClone, a statistical model for inference of clonal population structures in cancers. PyClone is a Bayesian clustering method for grouping sets of deeply sequenced somatic mutations into putative clonal clusters while estimating their cellular prevalences and accounting for allelic imbalances introduced by segmental copy-number changes and normal-cell contamination. Single-cell sequencing validation demonstrates PyClone's accuracy. <<<