Luis Bettencourt

Hierarchical networks are known to achieve high classification accuracy on difficult machine-learning tasks. For many applications, a clear explanation of why the data was classified a certain way is just as important as the classification itself. However, the complexity of hierarchical networks makes them ill-suited for existing explanation methods. We propose a new method, contribution propagation, that gives per-instance explanations of a trained network's classifications. We give theoretical foundations for the proposed method, and evaluate its correctness empirically. Finally, we use the resulting explanations to reveal unexpected behavior of networks that achieve high accuracy on visual object-recognition tasks using well-known data sets.

In the primate's primary visual cortex (V1), cells are classified in terms of two categories: simple cells and complex cells, given their response properties. While simple cells respond strongly to gating and bar stimuli at a certain phase and location, responses of complex cells are insensitive to small translation of stimulus within the receptive field [1]. Inspired by the response properties of simple and complex cells in the primary visual cortex, we propose a computational network to learn the receptive fields of these cells, and address the development of translation invariance from a temporal sequence of natural images. A generative model with sparseness constraints is devised to minimize the energy of prediction errors. Each simple cell is modulated by a higher layer of complex cells in a multiplicative fashion, where a slowness property and a trace-like rule are enforced on complex cells, as the result of a temporal coherence soft constraint. Furthermore, non-negativity constraints of the latent cell variables and weight matrices are imposed to fit the known neurophysiology. We present an online gradient descent algorithm to train our model from natural image sequences, in which a pre-training strategy is used to initialize the weights. The developed connection weights show that complex cell outputs are directly proportional to quadratic forms of simple cell responses. Each receptive field of simple cells develop a Gabor-like orientation filter, and each complex cell pools similar simple cell receptive fields - in retinotopic and feature space - producing the locally-invariant representation.

Conventional sparse coding learns optimal dictionaries of feature bases to approximate input signals; however, it is not favorable to classify the inputs. Recent research has focused on building discriminative sparse coding models to facilitate the classification tasks. In this paper, we develop a new discriminative sparse coding model via bidirectional flows. Sensory inputs (from bottom-up) and discriminative signals (supervised from top-down) are propagated through a hierarchical network to form sparse representations at each level. The ℓ 0 -constrained sparse coding model allows highly efficient online learning and does not require iterative steps to reach a fixed point of the sparse representation. The introduction of discriminative top-down information flows helps to group reconstructive features belonging to the same class and thus to benefit the classification tasks. Experiments are conducted on multiple data sets including natural images, hand-written digits and 3-D objects with favorable results. Compared with unsupervised sparse coding via only bottom-up directions, the two-way discriminative approach improves the recognition performance significantly.

Neuroscience has revealed many properties of neurons and of the functional organization of visual cortex that are believed to be essential to human vision, but are missing in standard artificial neural networks. Equally important may be the sheer scale of visual cortex requiring ~1 petaflop of computation, while the scale of human visual experience greatly exceeds standard computer vision datasets: the retina delivers ~1 petapixel/year to the brain, driving learning at many levels of the cortical system. We describe work at Los Alamos National Laboratory (LANL) to develop large-scale functional models of visual cortex on LANL's Roadrunner petaflop supercomputer. An initial run of a simple region V1 code achieved 1.144 petaflops during trials at the IBM facility in Poughkeepsie, NY (June 2008). Here, we present criteria for assessing when a set of learned local representations is ¿complete¿ along with general criteria for assessing computer vision models based on their projected scaling behavior. Finally, we extend one class of biologically-inspired learning models to problems of remote sensing imagery.

Social computation, whether in the form of searches performed by swarms of agents or collective predictions of markets, often supplies remarkably good solutions to complex problems. In many examples, individuals trying to solve a problem locally can aggregate their information and work together to arrive at a superior global solution. This suggests that there may be general principles of information aggregation and coordination that can transcend particular applications. Here we show that the general structure of this problem can be cast in terms of information theory and derive mathematical conditions that lead to optimal multi-agent searches. Specifically, we illustrate the problem in terms of local search algorithms for autonomous agents looking for the spatial location of a stochastic source. We explore the types of search problems, defined in terms of the statistical properties of the source and the nature of measurements at each agent, for which coordination among multiple searchers yields an advantage beyond that gained by having the same number of independent searchers. We show that effective coordination corresponds to synergy and that ineffective coordination corresponds to independence as defined using information theory. We classify explicit types of sources in terms of their potential for synergy. We show that sources that emit uncorrelated signals provide no opportunity for synergetic coordination while sources that emit signals that are correlated in some way, do allow for strong synergy between searchers. These general considerations are crucial for designing optimal algorithms for particular search problems in real world settings.

We develop a practical, distributed algorithm to detect events, identify measurement errors, and infer missing readings in ecological applications of wireless sensor networks. To address issues of non-stationarity in environmental data streams, each sensor-processor learns statistical distributions of differences between its readings and those of its neighbors, as well as between its current and previous measurements. Scalar physical quantities such as air temperature, soil moisture, and light flux naturally display a large degree of spatiotemporal coherence, which gives a spectrum of fluctuations between adjacent or consecutive measurements with small variances. This feature permits stable estimation over a small state space. The resulting probability distributions of differences, estimated online in real time, are then used in statistical significance tests to identify rare events. Utilizing the spatio-temporal distributed nature of the measurements across the network, these events are classified as single mode failures-usually corresponding to measurement errors at a single sensor- or common mode events. The event structure also allows the network to automatically attribute potential measurement errors to specific sensors and to correct them in real time via a combination of current measurements at neighboring nodes and the statistics of differences between them. Compared to methods that use Bayesian classification of raw data streams at each sensor, this algorithm is more storage-efficient, learns faster, and is more robust in the face of non-stationary phenomena. Field results from a wireless sensor network (Sensor Web) deployed at Sevilleta National Wildlife Refuge are presented.