R Sigala

Learning features of intermediate complexity for the recognition of biological motion

Humans can recognize biological motion from strongly impoverished stimuli, like point-light displays. Although the neural mechanism underlying this robust perceptual process have not yet been clarified, one possible explanation is that the visual system extracts specific motion features that are suitable for the robust recognition of both normal and degraded stimuli. We present a neural model for biological motion recognition that learns robust mid-level motion features in an unsupervised way using a neurally plausible memory-trace learning rule. Optimal mid-level features were learnt from image motion sequences containing a walker with, or without background motion clutter. After learning of the motion features, the detection performance of the model substantially increases, in particular in presence of clutter. The learned mid-level motion features are characterized by horizontal opponent motion, where this feature type arises more frequently for the training stimuli without motion clutter. The learned features are consistent with recent psychophysical data that indicates that opponent motion might be critical for the detection of point light walkers.

Modeling and simulating the adaptive electrical properties of stochastic polymeric 3D networks

Memristors are passive two-terminal circuit elements that combine resistance and memory. Although in theory memristors are a very promising approach to fabricate hardware with adaptive properties, there are only very few implementations able to show their basic properties. We recently developed stochastic polymeric matrices with a functionality that evidences the formation of self-assembled three-dimensional (3D) networks of memristors. We demonstrated that those networks show the typical hysteretic behavior observed in the ‘one input-one output’ memristive configuration. Interestingly, using different protocols to electrically stimulate the networks, we also observed that their adaptive properties are similar to those present in the nervous system. Here, we model and simulate the electrical properties of these selfassembled polymeric networks of memristors, the topology of which is defined stochastically. First, we show that the model recreates the hysteretic behavior observed in the real experiments. Second, we demonstrate that the networks modeled indeed have a 3D instead of a planar functionality. Finally, we show that the adaptive properties of the networks depend on their connectivity pattern. Our model was able to replicate fundamental qualitative behavior of the real organic 3D memristor networks; yet, through the simulations, we also explored other interesting properties, such as the relation between connectivity patterns and adaptive properties. Our model and simulations represent an interesting tool to understand the very complex behavior of self-assembled memristor networks, which can finally help to predict and formulate hypotheses for future experiments.

Stochastic hybrid 3D matrix: learning and adaptation of electrical properties

Memristive devices are electronic elements with memory properties. This feature marks them out as possible candidates for mimicking synapse properties. Development of systems capable of performing simple brain operations demands a high level of integration of elements and their 3D organization into networks. Here, we demonstrate the formation and electrical properties of stochastic polymeric matrices. Several features of the network revealed similarities with those of the nervous system. In particular, applying different training protocols, we obtained two kinds of learning comparable to the “baby” and “adult” learning in animals and humans. To mimic “adult” learning, multi-task training was applied simultaneously resulting in the formation of few parallel pathways for a given task, modifiable by successive training. To mimic “baby” learning (imprinting), single task training was applied at one time, resulting in the formation of multiple parallel signal pathways, scarcely influenced by successive training.

NSL/ASL: simulation of neural based visuomotor systems

Through experimentation and simulation scientists are able to get an understanding of the underlying biological mechanisms involved in living organisms. These mechanisms, both structural and behavioral, serve as inspiration in the modeling of neural based architectures as well as in the implementation of robotic systems. Among these, we are particularly motivated in studying animals such as toads, frogs, salamanders and praying mantis that rely on visuomotor coordination. In order to deal with the underlying complexity of these systems, we have developed the NSL/ASL simulation system to enable modeling and simulation at different levels of granularity.

Adaptive Properties of Stochastic Memristor Networks: A Computational Study

Abstract A ‘memristor’ is a passive two-terminal circuit element the electric resistance of which depends on the history of the charge that has passed through it. We implemented a platform to simulate adaptive properties of stochastic memristor networks. We showed that such networks follow a stable behavior that diverges from its initial state depending on the history of stimulation. Additionally, we observed that the connectivity patterns of the networks influence their adaptive properties. These results confirm the adaptive properties of statistical memristor networks and suggest that they can be potentially used as complex and self-assembled ‘learning machines’.

Visual Neuroscience: Face-Encoding Mechanisms Revealed by Adaptation

Faces convey a great variety of information, for example about the species, gender, age, identity and even mood or intentions. A recent study sheds light on the neural mechanisms for encoding a faces gaze direction.

Learning mid-level motion features for the recognition of body movements

Robust Recognition of Body Movements based on Critical Mid-Level Features • Humans can recognize body movements (e.g. walking and running) accurately and robustly. • This robustness is demonstrated by the ability of subjects to recognize body movements from strongly impoverished stimuli, like Point-like walkers (PLW), which consist only of a small number of illuminated dots that move like the joints of a human actor [6]. Subjects can even recognize gender, emotions, and identity from these stimuli. • A possible explanation for this robust generalization is that the brain extracts specific motion features (of intermediate complexity) that are shared by both stimuli classes (normal walker and PLW). • The nature of these features is unknown, and it has been discussed whether they are based predominantly on motion or form information [7]. In a recent study, combining methods from image statistics and psychophysical experiments, it was shown that robust recognition can be accomplished based on mid-level motion features [2]. Supported by National Institute of Health grant 2R01-EY-07861-11, Volkswagenstiftung, DFG and HFSP.