Kamil Wnuk

Filtering Internet image search results towards keyword based category recognition

In this work we aim to capitalize on the availability of Internet image search engines to automatically create image training sets from user provided queries. This problem is particularly difficult due to the low precision of image search results. Unlike many existing dataset gathering approaches, we do not assume a category model based on a small subset of the noisy data or an ad-hoc validation set. Instead we use a nonparametric measure of strangeness [8] in the space of holistic image representations, and perform an iterative feature elimination algorithm to remove the most strange examples from the category. This is the equivalent of keeping only features that are found to be consistent with others in the class. We show that applying our method to image search data before training improves average recognition performance, and demonstrate that we obtain comparative precision and recall results to the current state of the art, all the while maintaining a significantly simpler approach. In the process we also extend the strangeness-based feature elimination algorithm to automatically select good threshold values and perform filtering of a single class when the background is given.

Spike train driven dynamical models for human actions

We investigate dynamical models of human motion that can support both synthesis and analysis tasks. Unlike coarser discriminative models that work well when action classes are nicely separated, we seek models that have fine-scale representational power and can therefore model subtle differences in the way an action is performed. To this end, we model an observed action as an (unknown) linear time-invariant dynamical model of relatively small order, driven by a sparse bounded input signal. Our motivating intuition is that the time-invariant dynamics will capture the unchanging physical characteristics of an actor, while the inputs used to excite the system will correspond to a causal signature of the action being performed. We show that our model has sufficient representational power to closely approximate large classes of non-stationary actions with significantly reduced complexity. We also show that temporal statistics of the inferred input sequences can be compared in order to recognize actions and detect transitions between them.

Analyzing diving: a dataset for judging action quality

This work presents a unique new dataset and objectives for action analysis. The data presents 3 key challenges: tracking, classification, and judging action quality. The last of these, to our knowledge, has not yet been attempted in the vision literature as applied to sports where technique is scored. This work performs an initial analysis of the dataset with classification experiments, confirming that temporal information is more useful than holistic bag-of-features style analysis in distinguishing dives. Our investigation lays a groundwork of effective tools for working with this type of sports data for future investigations into judging the quality of actions.

Identification of an immune gene expression signature associated with favorable clinical features in Treg-enriched patient tumor samples

Immune heterogeneity within the tumor microenvironment undoubtedly adds several layers of complexity to our understanding of drug sensitivity and patient prognosis across various cancer types. Within the tumor microenvironment, immunogenicity is a favorable clinical feature in part driven by the antitumor activity of CD8+ T cells. However, tumors often inhibit this antitumor activity by exploiting the suppressive function of regulatory T cells (Tregs), thus suppressing the adaptive immune response. Despite the seemingly intuitive immunosuppressive biology of Tregs, prognostic studies have produced contradictory results regarding the relationship between Treg enrichment and survival. We therefore analyzed RNA-seq data of Treg-enriched tumor samples to derive a pan-cancer gene signature able to help reconcile the inconsistent results of Treg studies, by better understanding the variable clinical association of Tregs across alternative tumor contexts. We show that increased expression of a 32-gene signature in Treg-enriched tumor samples (n = 135) is able to distinguish a cohort of patients associated with chemosensitivity and overall survival. This cohort is also enriched for CD8+ T cell abundance, as well as the antitumor M1 macrophage subtype. With a subsequent validation in a larger TCGA pool of Treg-enriched patients (n = 626), our results reveal a gene signature able to produce unsupervised clusters of Treg-enriched patients, with one cluster of patients uniquely representative of an immunogenic tumor microenvironment. Ultimately, these results support the proposed gene signature as a putative biomarker to identify certain Treg-enriched patients with immunogenic tumors that are more likely to be associated with features of favorable clinical outcome.Cancer: Gene expression signature stratifies patients by tumor immune activityA new genetic test could help predict responses to therapy and survival outcomes among cancer patients with tumors that are infiltrated with large numbers of regulatory T cells (Tregs). Kevin B. Givechian of NantOmics in Los Angeles, California, USA, and colleagues measured gene activity levels in tumor samples taken from 135 patients with Treg-enriched cancers of all kinds. They singled out genes with particularly variable expression levels to create a 32-gene signature that revealed two distinct clusters of patients: those who responded to their prescribed drugs and lived longer, and those who were treatment-resistant and died sooner. The researchers also validated the gene panel in a larger, independent cohort of 626 tumor samples, showing that it could identify patients with immunogenic tumors who are more likely have favorable clinical outcomes.

Predicting DNA accessibility in the pan-cancer tumor genome using RNA-seq, WGS, and deep learning

DNA accessibility, chromatin regulation, and genome methylation are key drivers of transcriptional events promoting tumor growth. However, understanding the impact of DNA sequence data on transcriptional regulation of gene expression is a challenge, particularly in noncoding regions of the genome. Recently, neural networks have been used to effectively predict DNA accessibility in multiple specific cell types [14]. These models make it possible to explore the impact of mutations on DNA accessibility and transcriptional regulation. Our work first improved on prior cell-specific accessibility prediction, obtaining a mean receiver operating characteristic (ROC) area under the curve (AUC) = 0.910 and mean precision-recall (PR) AUC = 0.605, compared to the previous mean ROC AUC = 0.895 and mean PR AUC = 0.561 [14]. Our key contribution extended the model to enable accessibility predictions on any new sample for which RNA-seq data is available, without requiring cell-type-specific DNase-seq data for re-training. This new model obtained overall PR AUC = 0.621 and ROC AUC = 0.897 when applied across whole genomes of new samples whose biotypes were held out from training, and PR AUC = 0.725 and ROC AUC = 0.913 on randomly held out new samples whose biotypes were allowed to overlap with training. More significantly, we showed that for promoter and promoter flank regions of the genome our model predicts accessibility to high reliability, achieving PR AUC = 0.838 in held out biotypes and PR AUC = 0.908 in randomly held out samples.This performance is not sensitive to whether the promoter and flank regions fall within genes used in the input RNA-seq expression vector. Finally, we utilize this tool to investigate, for the first time, promoter accessibility patterns across several cohorts from The Cancer Genome Atlas (TCGA) [27].