Nisheeth Srivastava

Model-Based and Data-Driven Fault Detection Performance for a Small UAV

Fault detection and identification algorithms may rely on knowledge of underlying system dynamics while some eschew this modeling in favor of data-driven anomaly detection. This paper considers model-based residual generation and data-driven anomaly detection for a small, low-cost unmanned aerial vehicle using both types of approaches and applies those algorithms to experimental faulted and unfaulted flight-test data. The model-based fault detection strategy uses robust linear filtering methods to reject exogenous disturbances, e.g., wind, and provide robustness to model errors. The data-driven algorithm is developed to operate exclusively on raw flight-test data without detailed system knowledge. The detection performance of these complementary, but different, methods is compared.

Modeling the repertoire of true tumor-specific MHC I epitopes in a human tumor.

DNA replication has a finite measurable error rate, net of repair, in all cells. Clonal proliferation of cancer cells leads therefore to accumulation of random mutations. A proportion of these mutational events can create new immunogenic epitopes that, if processed and presented by an MHC allele, may be recognized by the adaptive immune system. Here, we use probability theory to analyze the mutational and epitope composition of a tumor mass in successive division cycles and create a double Pölya model for calculating the number of truly tumor-specific MHC I epitopes in a human tumor. We deduce that depending upon tumor size, the degree of genomic instability and the degree of death within a tumor, human tumors have several tens to low hundreds of new, truly tumor-specific epitopes. Parenthetically, cancer stem cells, due to the asymmetry in their proliferative properties, shall harbor significantly fewer mutations, and therefore significantly fewer immunogenic epitopes. As the overwhelming majority of the mutations in cancer cells are unrelated to malignancy, the mutation-generated epitopes shall be specific for each individual tumor, and constitute the antigenic fingerprint of each tumor. These calculations highlight the benefits for personalization of immunotherapy of human cancer, and in view of the substantial pre-existing antigenic repertoire of tumors, emphasize the enormous potential of therapies that modulate the anti-cancer immune response by liberating it from inhibitory influences.

Contextual Anomaly Detection in Text Data

We propose using side information to further inform anomaly detection algorithms of the semantic context of the text data they are analyzing, thereby considering both divergence from the statistical pattern seen in particular datasets and divergence seen from more general semantic expectations. Computational experiments show that our algorithm performs as expected on data that reflect real-world events with contextual ambiguity, while replicating conventional clustering on data that are either too specialized or generic to result in contextual information being actionable. These results suggest that our algorithm could potentially reduce false positive rates in existing anomaly detection systems.

Measuring spontaneous devaluations in user preferences

Spontaneous devaluation in preferences is ubiquitous, where yesterdays hit is todays affliction. Despite technological advances facilitating access to a wide range of media commodities, finding engaging content is a major enterprise with few principled solutions. Systems tracking spontaneous devaluation in user preferences can allow prediction of the onset of boredom in users potentially catering to their changed needs. In this work, we study the music listening histories of Last.fm users focusing on the changes in their preferences based on their choices for different artists at different points in time. A hazard function, commonly used in statistics for survival analysis, is used to capture the rate at which a user returns to an artist as a function of exposure to the artist. The analysis provides the first evidence of spontaneous devaluation in preferences of music listeners. Better understanding of the temporal dynamics of this phenomenon can inform solutions to the similarity-diversity dilemma of recommender systems.

Theoretically Optimal Distributed Anomaly Detection

A novel general framework for distributed anomaly detection with theoretical performance guarantees is proposed. Our algorithmic approach combines existing anomaly detection procedures with a novel method for computing global statistics using local sufficient statistics. Under a Gaussian assumption, our distributed algorithm is guaranteed to perform as well as its centralized counterpart, a condition we call ‘zero information loss’. We further report experimental results on synthetic as well as real-world data to demonstrate the viability of our approach.

Learning What to Want: Context-Sensitive Preference Learning.

We have developed a method for learning relative preferences from histories of choices made, without requiring an intermediate utility computation. Our method infers preferences that are rational in a psychological sense, where agent choices result from Bayesian inference of what to do from observable inputs. We further characterize conditions on choice histories wherein it is appropriate for modelers to describe relative preferences using ordinal utilities, and illustrate the importance of the influence of choice history by explaining all major categories of context effects using them. Our proposal clarifies the relationship between economic and psychological definitions of rationality and rationalizes several behaviors heretofore judged irrational by behavioral economists.

Characterizing the Internet's Sense of Humor

In this paper, we report some results from the first internet content-based investigation of the underlying causes of humor. For this purpose, we developed a methodology for extracting semantic distance from tags associated with YouTube videos manually identified as humorous or not by their existing community of users. We found that a novel quantification of episodic incongruity, operationalized via our technique, proves to be a necessary but not sufficient condition for the existence of humor-inducing stimuli in associated videos. Our results represent the first internet-based validation of incongruity-based characterizations of humor, and open up exciting new theoretical and applied possibilities in the use of social computing to discover intrinsic factors responsible for human behaviors like humor, interest and engagement.

An Evolutionarily Motivated Model of Decision-Making Under Uncertainty

There has been a recent shift in decision theory research from attempting to fix classical compensatory models towards adopting non-compensatory models of decision-making. This shift has arisen largely because of the inability of compensatory expected-utility based approaches to explain a large number of cognitive biases reliably observed in human subjects on experimental decision tasks. We show, using a jointly evolutionary and information-theoretic argument, that these so-called biases are, in fact, completely rational, if rationality is defined as minimizing the subjects cognitive effort in making a satisfactorily accurate decision. In this paper, we formalize this intuition in the form of a compensatory model of decision-making, and show that this extremely simple and interpretable model can generatively replicate three classic experimental studies spanning distinctive families of cognitive biases, viz. probabilistic sub-additivity leading to a fourfold pattern of risk aversion, confirmatory positive hypothesis selection and serial ordering effects. We suggest that this unified explanation for hitherto unconnected cognitive phenomena provides evidence for the existence of a fundamental information-theoretic optimality principle in the nature of human intelligence.

Perceptual and cognitive limitations interact in multiple object tracking

Limitations on higher-order cognitive resources have been proposed as constraints on human performance in multiple object tracking (MOT). We report results from experiments testing a possible account of these cognitive limitations - observers rationally allocate attention to object locations to reduce the probability of spatial interference, they accomplish this reduction via increased spatial precision at attended locations, and drop targets within trials when their attention resource pool is overwhelmed. Our experiments used a standard MOT display following (Alvarez & Franconeri, 2007), but made two modifications to the standard design that allowed us to (i) probe the spatial precision with which subjects track individual objects during MOT, and (ii) whether subjects were aware that they had dropped any targets during trials where they made errors. We also conducted computational experiments, augmenting an existing model (Vul et al, 2009) with a central controller that assigns a limited attention resource to low-level object trackers, proportionately reducing their perceptual noise to minimize the chances of local perceptual confusion. We report three main findings. First, our computational model replicates the pattern of human errors on a per-trial basis to a considerably greater extent than a simple spatial interference model. Second, observers can localize targets with greater precision than non-targets, and that, contra simple spatial interference arguments, they localize targets with greater precision in crowded positions than when they are in open space. Third, observers were significantly less likely to report surprise (implying they knew they had dropped at least one target) for trials where our model predicted greater instances of attention resource scarcity. These results together support the case for rational attention allocation attenuating perceptual confusion during MOT trials, and limitations on this attention pool causing at least some of the difficulties that human observers encounter in multiple object tracking. Meeting abstract presented at VSS 2015.

Learning What to Want: Data-Driven Microfoundations

We have developed a method for directly learning relative preferences from histories of comparison information without an intermediate utility computation. Our method infers preferences that are rational in a psychological sense, where agent choices result from Bayesian inference of what to do from observable inputs. We further characterize conditions wherein it is appropriate for modelers to describe relative preferences as scalar utilities, and illustrate the importance of option availability in supplying auxiliary information by explaining all major categories of context effects, as well as predicting novel context effects. Applying our theory to predicting choices under uncertainty leads to good fits with empirical data and endogenous explanations for a number of economic behaviors. By retrieving economic rationality as a special case of psychologically rational preference formation, this work clarifies theoretical connections between economic and psychological definitions of rationality.