Showing papers by "Université de Montréal published in 2010"

Understanding the difficulty of training deep feedforward neural networks

Abstract: Whereas before 2006 it appears that deep multilayer neural networks were not successfully trained, since then several algorithms have been shown to successfully train them, with experimental results showing the superiority of deeper vs less deep architectures. All these experimental results were obtained with new initialization or training mechanisms. Our objective here is to understand better why standard gradient descent from random initialization is doing so poorly with deep neural networks, to better understand these recent relative successes and help design better algorithms in the future. We first observe the influence of the non-linear activations functions. We find that the logistic sigmoid activation is unsuited for deep networks with random initialization because of its mean value, which can drive especially the top hidden layer into saturation. Surprisingly, we find that saturated units can move out of saturation by themselves, albeit slowly, and explaining the plateaus sometimes seen when training neural networks. We find that a new non-linearity that saturates less can often be beneficial. Finally, we study how activations and gradients vary across layers and during training, with the idea that training may be more difficult when the singular values of the Jacobian associated with each layer are far from 1. Based on these considerations, we propose a new initialization scheme that brings substantially faster convergence. 1 Deep Neural Networks Deep learning methods aim at learning feature hierarchies with features from higher levels of the hierarchy formed by the composition of lower level features. They include Appearing in Proceedings of the 13 International Conference on Artificial Intelligence and Statistics (AISTATS) 2010, Chia Laguna Resort, Sardinia, Italy. Volume 9 of JMLR: WC Weston et al., 2008). Much attention has recently been devoted to them (see (Bengio, 2009) for a review), because of their theoretical appeal, inspiration from biology and human cognition, and because of empirical success in vision (Ranzato et al., 2007; Larochelle et al., 2007; Vincent et al., 2008) and natural language processing (NLP) (Collobert & Weston, 2008; Mnih & Hinton, 2009). Theoretical results reviewed and discussed by Bengio (2009), suggest that in order to learn the kind of complicated functions that can represent high-level abstractions (e.g. in vision, language, and other AI-level tasks), one may need deep architectures. Most of the recent experimental results with deep architecture are obtained with models that can be turned into deep supervised neural networks, but with initialization or training schemes different from the classical feedforward neural networks (Rumelhart et al., 1986). Why are these new algorithms working so much better than the standard random initialization and gradient-based optimization of a supervised training criterion? Part of the answer may be found in recent analyses of the effect of unsupervised pretraining (Erhan et al., 2009), showing that it acts as a regularizer that initializes the parameters in a “better” basin of attraction of the optimization procedure, corresponding to an apparent local minimum associated with better generalization. But earlier work (Bengio et al., 2007) had shown that even a purely supervised but greedy layer-wise procedure would give better results. So here instead of focusing on what unsupervised pre-training or semi-supervised criteria bring to deep architectures, we focus on analyzing what may be going wrong with good old (but deep) multilayer neural networks. Our analysis is driven by investigative experiments to monitor activations (watching for saturation of hidden units) and gradients, across layers and across training iterations. We also evaluate the effects on these of choices of activation function (with the idea that it might affect saturation) and initialization procedure (since unsupervised pretraining is a particular form of initialization and it has a drastic impact).

Human Papillomavirus and Survival of Patients with Oropharyngeal Cancer

Abstract: Background Oropharyngeal squamous-cell carcinomas caused by human papillomavirus (HPV) are associated with favorable survival, but the independent prognostic significance of tumor HPV status remains unknown. Methods We performed a retrospective analysis of the association between tumor HPV status and survival among patients with stage III or IV oropharyngeal squamous-cell carcinoma who were enrolled in a randomized trial comparing accelerated-fractionation radiotherapy (with acceleration by means of concomitant boost radiotherapy) with standard-fractionation radiotherapy, each combined with cisplatin therapy, in patients with squamous-cell carcinoma of the head and neck. Proportional-hazards models were used to compare the risk of death among patients with HPV-positive cancer and those with HPV-negative cancer. Results The median follow-up period was 4.8 years. The 3-year rate of overall survival was similar in the group receiving accelerated-fractionation radiotherapy and the group receiving standard-fractionation radiotherapy (70.3% vs. 64.3%; P = 0.18; hazard ratio for death with accelerated-fractionation radiotherapy, 0.90; 95% confidence interval [CI], 0.72 to 1.13), as were the rates of high-grade acute and late toxic events. A total of 63.8% of patients with oropharyngeal cancer (206 of 323) had HPV-positive tumors; these patients had better 3-year rates of overall survival (82.4%, vs. 57.1% among patients with HPV-negative tumors; P<0.001 by the log-rank test) and, after adjustment for age, race, tumor and nodal stage, tobacco exposure, and treatment assignment, had a 58% reduction in the risk of death (hazard ratio, 0.42; 95% CI, 0.27 to 0.66). The risk of death significantly increased with each additional packyear of tobacco smoking. Using recursive-partitioning analysis, we classified our patients as having a low, intermediate, or high risk of death on the basis of four factors: HPV status, pack-years of tobacco smoking, tumor stage, and nodal stage. Conclusions Tumor HPV status is a strong and independent prognostic factor for survival among patients with oropharyngeal cancer. (ClinicalTrials.gov number, NCT00047008.)

Emerging applications of stimuli-responsive polymer materials

Abstract: Responsive polymer materials can adapt to surrounding environments, regulate transport of ions and molecules, change wettability and adhesion of different species on external stimuli, or convert chemical and biochemical signals into optical, electrical, thermal and mechanical signals, and vice versa. These materials are playing an increasingly important part in a diverse range of applications, such as drug delivery, diagnostics, tissue engineering and 'smart' optical systems, as well as biosensors, microelectromechanical systems, coatings and textiles. We review recent advances and challenges in the developments towards applications of stimuli-responsive polymeric materials that are self-assembled from nanostructured building blocks. We also provide a critical outline of emerging developments.

Biological, clinical and population relevance of 95 loci for blood lipids

Abstract: Plasma concentrations of total cholesterol, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol and triglycerides are among the most important risk factors for coronary artery disease (CAD) and are targets for therapeutic intervention. We screened the genome for common variants associated with plasma lipids in >100,000 individuals of European ancestry. Here we report 95 significantly associated loci (P < 5 x 10(-8)), with 59 showing genome-wide significant association with lipid traits for the first time. The newly reported associations include single nucleotide polymorphisms (SNPs) near known lipid regulators (for example, CYP7A1, NPC1L1 and SCARB1) as well as in scores of loci not previously implicated in lipoprotein metabolism. The 95 loci contribute not only to normal variation in lipid traits but also to extreme lipid phenotypes and have an impact on lipid traits in three non-European populations (East Asians, South Asians and African Americans). Our results identify several novel loci associated with plasma lipids that are also associated with CAD. Finally, we validated three of the novel genes-GALNT2, PPP1R3B and TTC39B-with experiments in mouse models. Taken together, our findings provide the foundation to develop a broader biological understanding of lipoprotein metabolism and to identify new therapeutic opportunities for the prevention of CAD.

A distance-based framework for measuring functional diversity from multiple traits

Abstract: A new framework for measuring functional diversity (FD) from multiple traits has recently been proposed. This framework was mostly limited to quantitative traits without missing values and to situations in which there are more species than traits, although the authors had suggested a way to extend their framework to other trait types. The main purpose of this note is to further develop this suggestion. We describe a highly flexible distance-based framework to measure different facets of FD in multidimensional trait space from any distance or dissimilarity measure, any number of traits, and from different trait types (i.e., quantitative, semi-quantitative, and qualitative). This new approach allows for missing trait values and the weighting of individual traits. We also present a new multidimensional FD index, called functional dispersion (FDis), which is closely related to Rao's quadratic entropy. FDis is the multivariate analogue of the weighted mean absolute deviation (MAD), in which the weights are species relative abundances. For unweighted presence-absence data, FDis can be used for a formal statistical test of differences in FD. We provide the "FD" R language package to easily implement our distance-based FD framework.

Association analyses of 249,796 individuals reveal 18 new loci associated with body mass index

Abstract: Obesity is globally prevalent and highly heritable, but its underlying genetic factors remain largely elusive. To identify genetic loci for obesity susceptibility, we examined associations between body mass index and similar to 2.8 million SNPs in up to 123,865 individuals with targeted follow up of 42 SNPs in up to 125,931 additional individuals. We confirmed 14 known obesity susceptibility loci and identified 18 new loci associated with body mass index (P < 5 x 10(-8)), one of which includes a copy number variant near GPRC5B. Some loci (at MC4R, POMC, SH2B1 and BDNF) map near key hypothalamic regulators of energy balance, and one of these loci is near GIPR, an incretin receptor. Furthermore, genes in other newly associated loci may provide new insights into human body weight regulation.

Genome-wide meta-analysis increases to 71 the number of confirmed Crohn's disease susceptibility loci

Abstract: We undertook a meta-analysis of six Crohn's disease genome-wide association studies (GWAS) comprising 6,333 affected individuals (cases) and 15,056 controls and followed up the top association signals in 15,694 cases, 14,026 controls and 414 parent-offspring trios. We identified 30 new susceptibility loci meeting genome-wide significance (P < 5 × 10⁻⁸). A series of in silico analyses highlighted particular genes within these loci and, together with manual curation, implicated functionally interesting candidate genes including SMAD3, ERAP2, IL10, IL2RA, TYK2, FUT2, DNMT3A, DENND1B, BACH2 and TAGAP. Combined with previously confirmed loci, these results identify 71 distinct loci with genome-wide significant evidence for association with Crohn's disease.

Word Representations: A Simple and General Method for Semi-Supervised Learning

Abstract: If we take an existing supervised NLP system, a simple and general way to improve accuracy is to use unsupervised word representations as extra word features. We evaluate Brown clusters, Collobert and Weston (2008) embeddings, and HLBL (Mnih & Hinton, 2009) embeddings of words on both NER and chunking. We use near state-of-the-art supervised baselines, and find that each of the three word representations improves the accuracy of these baselines. We find further improvements by combining different word representations. You can download our word features, for off-the-shelf use in existing NLP systems, as well as our code, here: http://metaoptimize.com/projects/wordreprs/

Why Does Unsupervised Pre-training Help Deep Learning?

Abstract: Much recent research has been devoted to learning algorithms for deep architectures such as Deep Belief Networks and stacks of auto-encoder variants, with impressive results obtained in several areas, mostly on vision and language data sets. The best results obtained on supervised learning tasks involve an unsupervised learning component, usually in an unsupervised pre-training phase. Even though these new algorithms have enabled training deep models, many questions remain as to the nature of this difficult learning problem. The main question investigated here is the following: how does unsupervised pre-training work? Answering this questions is important if learning in deep architectures is to be further improved. We propose several explanatory hypotheses and test them through extensive simulations. We empirically show the influence of pre-training with respect to architecture depth, model capacity, and number of training examples. The experiments confirm and clarify the advantage of unsupervised pre-training. The results suggest that unsupervised pre-training guides the learning towards basins of attraction of minima that support better generalization from the training data set; the evidence from these results supports a regularization explanation for the effect of pre-training.

Hundreds of variants clustered in genomic loci and biological pathways affect human height

Abstract: Most common human traits and diseases have a polygenic pattern of inheritance: DNA sequence variants at many genetic loci influence the phenotype. Genome-wide association (GWA) studies have identified more than 600 variants associated with human traits, but these typically explain small fractions of phenotypic variation, raising questions about the use of further studies. Here, using 183,727 individuals, we show that hundreds of genetic variants, in at least 180 loci, influence adult height, a highly heritable and classic polygenic trait. The large number of loci reveals patterns with important implications for genetic studies of common human diseases and traits. First, the 180 loci are not random, but instead are enriched for genes that are connected in biological pathways (P = 0.016) and that underlie skeletal growth defects (P < 0.001). Second, the likely causal gene is often located near the most strongly associated variant: in 13 of 21 loci containing a known skeletal growth gene, that gene was closest to the associated variant. Third, at least 19 loci have multiple independently associated variants, suggesting that allelic heterogeneity is a frequent feature of polygenic traits, that comprehensive explorations of already-discovered loci should discover additional variants and that an appreciable fraction of associated loci may have been identified. Fourth, associated variants are enriched for likely functional effects on genes, being over-represented among variants that alter amino-acid structure of proteins and expression levels of nearby genes. Our data explain approximately 10% of the phenotypic variation in height, and we estimate that unidentified common variants of similar effect sizes would increase this figure to approximately 16% of phenotypic variation (approximately 20% of heritable variation). Although additional approaches are needed to dissect the genetic architecture of polygenic human traits fully, our findings indicate that GWA studies can identify large numbers of loci that implicate biologically relevant genes and pathways.

Cardiac-Resynchronization Therapy for Mild-to-Moderate Heart Failure

Abstract: Background Cardiac-resynchronization therapy (CRT) benefits patients with left ventricular systolic dysfunction and a wide QRS complex. Most of these patients are candidates for an implantable cardioverter–defibrillator (ICD). We evaluated whether adding CRT to an ICD and optimal medical therapy might reduce mortality and morbidity among such patients. Methods We randomly assigned patients with New York Heart Association (NYHA) class II or III heart failure, a left ventricular ejection fraction of 30% or less, and an intrinsic QRS duration of 120 msec or more or a paced QRS duration of 200 msec or more to receive either an ICD alone or an ICD plus CRT. The primary outcome was death from any cause or hospitalization for heart failure. Results We followed 1798 patients for a mean of 40 months. The primary outcome occurred in 297 of 894 patients (33.2%) in the ICD–CRT group and 364 of 904 patients (40.3%) in the ICD group (hazard ratio in the ICD–CRT group, 0.75; 95% confi dence interval [CI], 0.64 to 0.87; P<0.001). In the ICD–CRT group, 186 patients died, as compared with 236 in the ICD group (hazard ratio, 0.75; 95% CI, 0.62 to 0.91; P = 0.003), and 174 patients were hospitalized for heart failure, as compared with 236 in the ICD group (hazard ratio, 0.68; 95% CI, 0.56 to 0.83; P<0.001). However, at 30 days after device implantation, adverse events had occurred in 124 patients in the ICD-CRT group, as compared with 58 in the ICD group (P<0.001). Conclusions Among patients with NYHA class II or III heart failure, a wide QRS complex, and left ventricular systolic dysfunction, the addition of CRT to an ICD reduced rates of death and hospitalization for heart failure. This improvement was accompanied by more adverse events. (Funded by the Canadian Institutes of Health Research and Medtronic of Canada; ClinicalTrials.gov number, NCT00251251.)

ARID1A mutations in endometriosis-associated ovarian carcinomas.

Abstract: Background Ovarian clear-cell and endometrioid carcinomas may arise from endometriosis, but the molecular events involved in this transformation have not been described. Methods We sequenced the whole transcriptomes of 18 ovarian clear-cell carcinomas and 1 ovarian clear-cell carcinoma cell line and found somatic mutations in ARID1A (the AT-rich interactive domain 1A [SWI-like] gene) in 6 of the samples. ARID1A encodes BAF250a, a key component of the SWI–SNF chromatin remodeling complex. We sequenced ARID1A in an additional 210 ovarian carcinomas and a second ovarian clear-cell carcinoma cell line and measured BAF250a expression by means of immunohistochemical analysis in an additional 455 ovarian carcinomas. Results ARID1A mutations were seen in 55 of 119 ovarian clear-cell carcinomas (46%), 10 of 33 endometrioid carcinomas (30%), and none of the 76 high-grade serous ovarian carcinomas. Seventeen carcinomas had two somatic mutations each. Loss of the BAF250a protein correlated strongly with the ovarian c...

Sentinel-lymph-node resection compared with conventional axillary-lymph-node dissection in clinically node-negative patients with breast cancer: overall survival findings from the NSABP B-32 randomised phase 3 trial

Abstract: Summary Background Sentinel-lymph-node (SLN) surgery was designed to minimise the side-effects of lymph-node surgery but still offer outcomes equivalent to axillary-lymph-node dissection (ALND). The aims of National Surgical Adjuvant Breast and Bowel Project (NSABP) trial B-32 were to establish whether SLN resection in patients with breast cancer achieves the same survival and regional control as ALND, but with fewer side-effects. Methods NSABP B-32 was a randomised controlled phase 3 trial done at 80 centres in Canada and the USA between May 1, 1999, and Feb 29, 2004. Women with invasive breast cancer were randomly assigned to either SLN resection plus ALND (group 1) or to SLN resection alone with ALND only if the SLNs were positive (group 2). Random assignment was done at the NSABP Biostatistical Center (Pittsburgh, PA, USA) with a biased coin minimisation approach in an allocation ratio of 1:1. Stratification variables were age at entry (≤49 years, ≥50 years), clinical tumour size (≤2·0 cm, 2·1–4·0 cm, ≥4·1 cm), and surgical plan (lumpectomy, mastectomy). SLN resection was done with a blue dye and radioactive tracer. Outcome analyses were done in patients who were assessed as having pathologically negative sentinel nodes and for whom follow-up data were available. The primary endpoint was overall survival. Analyses were done on an intention-to-treat basis. All deaths, irrespective of cause, were included. The mean time on study for the SLN-negative patients with follow-up information was 95·6 months (range 70·1–126·7). This study is registered with ClinicalTrials.gov, number NCT00003830. Findings 5611 women were randomly assigned to the treatment groups, 3989 had pathologically negative SLN. 309 deaths were reported in the 3986 SLN-negative patients with follow-up information: 140 of 1975 patients in group 1 and 169 of 2011 in group 2. Log-rank comparison of overall survival in groups 1 and 2 yielded an unadjusted hazard ratio (HR) of 1·20 (95% CI 0·96–1·50; p=0·12). 8-year Kaplan-Meier estimates for overall survival were 91·8% (95% CI 90·4–93·3) in group 1 and 90·3% (88·8–91·8) in group 2. Treatment comparisons for disease-free survival yielded an unadjusted HR of 1·05 (95% CI 0·90–1·22; p=0·54). 8-year Kaplan-Meier estimates for disease-free survival were 82·4% (80·5–84·4) in group 1 and 81·5% (79·6–83·4) in group 2. There were eight regional-node recurrences as first events in group 1 and 14 in group 2 (p=0·22). Patients are continuing follow-up for longer-term assessment of survival and regional control. The most common adverse events were allergic reactions, mostly related to the administration of the blue dye. Interpretation Overall survival, disease-free survival, and regional control were statistically equivalent between groups. When the SLN is negative, SLN surgery alone with no further ALND is an appropriate, safe, and effective therapy for breast cancer patients with clinically negative lymph nodes. Funding US Public Health Service, National Cancer Institute, and Department of Health and Human Services.

Neural Mechanisms for Interacting with a World Full of Action Choices

Abstract: The neural bases of behavior are often discussed in terms of perceptual, cognitive, and motor stages, defined within an information processing framework that was originally inspired by models of human abstract problem solving. Here, we review a growing body of neurophysiological data that is difficult to reconcile with this influential theoretical perspective. As an alternative foundation for interpreting neural data, we consider frameworks borrowed from ethology, which emphasize the kinds of real-time interactive behaviors that animals have engaged in for millions of years. In particular, we discuss an ethologically-inspired view of interactive behavior as simultaneous processes that specify potential motor actions and select between them. We review how recent neurophysiological data from diverse cortical and subcortical regions appear more compatible with this parallel view than with the classical view of serial information processing stages.

From External Regulation to Self‐Regulation: Early Parenting Precursors of Young Children’s Executive Functioning

Abstract: In keeping with proposals emphasizing the role of early experience in infant brain development, this study investigated the prospective links between quality of parent–infant interactions and subsequent child executive functioning (EF), including working memory, impulse control, and set shifting. Maternal sensitivity, mind-mindedness and autonomy support were assessed when children were 12 to 15 months old (N = 80). Child EF was assessed at 18 and 26 months. All three parenting dimensions were found to relate to child EF. Autonomy support was the strongest predictor of EF at each age, independent of general cognitive ability and maternal education. These findings add to previous results on child stress-response systems in suggesting that parent–child relationships may play an important role in children’s developing self-regulatory capacities.

Comparison of antiarrhythmic drug therapy and radiofrequency catheter ablation in patients with paroxysmal atrial fibrillation: A randomized controlled trial

Abstract: Context Antiarrhythmic drugs are commonly used for prevention of recurrent atrial fibrillation (AF) despite inconsistent efficacy and frequent adverse effects. Catheter ablation has been proposed as an alternative treatment for paroxysmal AF. Objective To determine the efficacy of catheter ablation compared with antiarrhythmic drug therapy (ADT) in treating symptomatic paroxysmal AF. Design, Setting, and Participants A prospective, multicenter, randomized (2:1), unblinded, Bayesian-designed study conducted at 19 hospitals of 167 patients who did not respond to at least 1 antiarrhythmic drug and who experienced at least 3 AF episodes within 6 months before randomization. Enrollment occurred between October 25, 2004, and October 11, 2007, with the last follow-up on January 19, 2009. Intervention Catheter ablation (n = 106) or ADT (n = 61), with assessment for effectiveness in a comparable 9-month follow-up period. Main Outcome Measures Time to protocol-defined treatment failure. The proportion of patients who experienced major treatment-related adverse events within 30 days of catheter ablation or ADT was also reported. Results At the end of the 9-month effectiveness evaluation period, 66% of patients in the catheter ablation group remained free from protocol-defined treatment failure compared with 16% of patients treated with ADT. The hazard ratio of catheter ablation to ADT was 0.30 (95% confidence interval, 0.19-0.47; P Conclusion Among patients with paroxysmal AF who had not responded to at least 1 antiarrhythmic drug, the use of catheter ablation compared with ADT resulted in a longer time to treatment failure during the 9-month follow-up period. Trial Registration clinicaltrials.gov Identifier: NCT00116428

Theano: A CPU and GPU Math Compiler in Python

Abstract: Theano is a compiler for mathematical expressions in Python that combines the convenience of NumPy's syntax with the speed of optimized native machine language. The user composes mathematical expressions in a high-level description that mimics NumPy's syntax and semantics, while being statically typed and functional (as opposed to imperative). These expressions allow Theano to provide symbolic differentiation. Before performing computation, Theano optimizes the choice of expressions, translates them into C++ (or CUDA for GPU), compiles them into dynamically loaded Python modules, all automatically. Common machine learn- ing algorithms implemented with Theano are from 1:6 to 7:5 faster than competitive alternatives (including those implemented with C/C++, NumPy/SciPy and MATLAB) when compiled for the CPU and between 6:5 and 44 faster when compiled for the GPU. This paper illustrates how to use Theano, outlines the scope of the compiler, provides benchmarks on both CPU and GPU processors, and explains its overall design.

Meta-analysis identifies 13 new loci associated with waist-hip ratio and reveals sexual dimorphism in the genetic basis of fat distribution

Abstract: Waist-hip ratio (WHR) is a measure of body fat distribution and a predictor of metabolic consequences independent of overall adiposity. WHR is heritable, but few genetic variants influencing this trait have been identified. We conducted a meta-analysis of 32 genome-wide association studies for WHR adjusted for body mass index (comprising up to 77,167 participants), following up 16 loci in an additional 29 studies (comprising up to 113,636 subjects). We identified 13 new loci in or near RSPO3, VEGFA, TBX15-WARS2, NFE2L3, GRB14, DNM3-PIGC, ITPR2-SSPN, LY86, HOXC13, ADAMTS9, ZNRF3-KREMEN1, NISCH-STAB1 and CPEB4 (P = 1.9 × 10⁻⁹ to P = 1.8 × 10⁻⁴⁰) and the known signal at LYPLAL1. Seven of these loci exhibited marked sexual dimorphism, all with a stronger effect on WHR in women than men (P for sex difference = 1.9 × 10⁻³ to P = 1.2 × 10⁻¹³). These findings provide evidence for multiple loci that modulate body fat distribution independent of overall adiposity and reveal strong gene-by-sex interactions.

Controlling low-level image properties: the SHINE toolbox.

Abstract: Visual perception can be influenced by top-down processes related to the observer’s goals and expectations, as well as by bottom-up processes related to low-level stimulus attributes, such as luminance, contrast, and spatial frequency. When using different physical stimuli across psychological conditions, one faces the problem of disentangling the contributions of low- and high-level factors. Here, we make available the SHINE (spectrum, histogram, and intensity normalization and equalization) toolbox for MATLAB, which we have found useful for controlling a number of image properties separately or simultaneously. The toolbox features functions for specifying the (rotational average of the) Fourier amplitude spectra, for normalizing and scaling mean luminance and contrast, and for exact histogram specification optimized for perceptual visual quality. SHINE can thus be employed for parametrically modifying a number of image properties or for equating them across stimuli to minimize potential low-level confounds in studies on higher level processes.

Observation of a centrality-dependent dijet asymmetry in lead-lead collisions at √sNN=2.76 Tev with the ATLAS detector at the LHC

Abstract: By using the ATLAS detector, observations have been made of a centrality-dependent dijet asymmetry in the collisions of lead ions at the Large Hadron Collider. In a sample of lead-lead events with a per-nucleon center of mass energy of 2.76 TeV, selected with a minimum bias trigger, jets are reconstructed in fine-grained, longitudinally segmented electromagnetic and hadronic calorimeters. The transverse energies of dijets in opposite hemispheres are observed to become systematically more unbalanced with increasing event centrality leading to a large number of events which contain highly asymmetric dijets. This is the first observation of an enhancement of events with such large dijet asymmetries, not observed in proton-proton collisions, which may point to an interpretation in terms of strong jet energy loss in a hot, dense medium.