There is an ongoing debate over the capabilities of hierarchical neural feed-forward architectures for performing real-world invariant object recognition. Although a variety of hierarchical models exists, appropriate supervised and unsupervised learning methods are still an issue of intense research. We propose a feedforward model for recognition that shares components like weightsharing, pooling stages, and competitive nonlinearities with earlier approaches, but focus on new methods for learning optimal featuredetecting cells in intermediate stages of the hierarchical network.

Abstract not found

Abstract not found

plausibility proof

We present a two-layer dynamic generative model of the statistical structure of natural image sequences. The second layer of the model is a linear mapping from simple-cell outputs to pixel values, as in most work on natural image statistics. The first layer models the dependencies of the activity levels (amplitudes or variances) of the simple cells, using a multivariate autoregressive model. The second layer shows the emergence of basis vectors that are localized, oriented and have different scales, just like in previous work. But in our new model, the first layer learns connections between the simple cells that are similar to complex cell pooling: connections are strong among cells with similar preferred location, frequency and orientation. In contrast to previous work in which one of the layers needed to be fixed in advance, the dynamic model enables us to estimate both of the layers simultaneously fromnatural data.

Learning in neural networks is usually applied to parameters related to linear kernels and keeps the nonlinearity of the model fixed. Thus, for successful models properties and parameters of the nonlinearity have to be specified using a priori knowledge, that is, however, often missing. Here, we investigate adapting the nonlinearity simultaneously with the linear kernel. We use natural visual stimuli for training a simple model of the visual system. Many of the neurons converge to an energy detector matching existing models of complex cells. The overall distribution of the parameter describing the nonlinearity matches well recent physiological results. Controls with randomly shuffled natural stimuli and pink noise demonstrate that the match of simulation and experimental results depends on the higher order statistical properties of natural stimuli.

An emerging paradigm analyses in what respect the properties of the nervous system reflect properties of natural scenes. It is hypothesized that neurons form sparse representations of natural stimuli: Each neuron should respond strongly to some stimuli while being inactive upon presentation of most others. For a given network, sparse representations need fewest spikes, and thus consume the least energy. To obtain optimally sparse responses the receptive fields of simulated neurons are optimized. Algorithmically this is identical to searching for basis functions that allow coding for the stimuli with sparse coefficients. The problem is thus identical to maximising the log likelihood of a generative model with prior knowledge of natural images. It is found that the resulting simulated neurons share most properties of simple cells found in primary visual cortex. Thus, forming optimally sparse representations is the most compact approach to described simple cell properties. Many ways of defining sparse responses exist and it is widely believed that the particular choice of the sparse prior of the generative model does not significantly influence the estimated basis functions. Here we examine this assumption more closely. We include the constraint of unit variance of neuronal activity, used in most studies, into the objective functions. We then analyse learning on a database of natural (cat-cam™) visual stimuli. We show that the effective objective functions are largely dominated by the constraint, and are therefore very similar. The resulting receptive fields show some similarities but also qualitative differences. Even in the region where the objective functions are dissimilar, the distributions of coefficients are similar and do not match the priors of the assumed generative model. In conclusion, the specific choice of the sparse prior is relevant, as is the choice of additional constraints, such as normalization of variance.

Many biological and artificial neural networks require the parallel extraction of multiple features, and meet this requirement with distinct populations of neurons that are selective to one property of the stimulus while being non-selective to another property. In this way, several populations can resolve a set of features independently of each other, and thus achieve a parallel mode of processing. This raises the question how an initially homogeneous population of neurons segregates into groups with distinct and complementary response properties. Using a colour image sequence recorded from a camera mounted to the head of a freely behaving cat, we train a network of neurons to achieve optimally stable responses, that is, responses that change minimally over time. This objective leads to the development of colour selective neurons. Adding a second objective, de-correlating activity within the network, a subpopulation of neurons develops with achromatic response properties. Colour selective neurons tend to be non-oriented while achromatic neurons are orientation-tuned. The proposed objective thus successfully leads to the segregation of neurons into complementary populations that are either selective for colour or orientation.

Lierhaus (2005) reported results of an eye tracking experiment with rhesus macaques. He concluded there was a twenty percentage increase in the effect of luminance contrast formonkeys in the reported experiment, as compared to (Reinagel & Zador, 1999) and a decreased influence of spatial contrast. We conducted an eye-tracking experiment under conditions as close as possible to the monkeys’ earlier experiment and compared the results. Influences of luminance and texture contrasts were increased in humans, the effect varying however in different image categories. The most obvious differences we found for eye movements: monkeys tended to look rather down and made faster, wider, and longer lasting saccades as compared to humans. A bootstrap comparison of monkey and human fixation maps showed significant differences between the species.

In natural behaviour we actively attend to parts of a visual scene by moving our eyes. Models of such overt attention combine different local features in a bottom‐up process. Here, we empirically study this integration process and investigate the interaction of luminance, luminance contrast, texture contrast, edges and colour contrast during free viewing of natural stimuli. (1) We describe how the saliency of a feature varies with the feature values: Some features correlate with saliency in a linear and others in a non‐linear fashion. (2) We describe the interaction of features, how the saliency of one feature varies in the context of a second feature. We find that in general knowing the value of one feature does not give much information about the saliency of another feature, i.e. different features contribute independently to a joint saliency map. (3) We use additive and multiplicative integration processes to model feature integration. Additive integration of feature saliencies gives a good description of the data for all analyzed feature combinations within and across feature channels. In summary, we argue that selection of attended regions can be explained by the additive integration of independently processed single feature saliency functions that have suitable linear and non‐linear