A Neuro-Inspired Computational Model for a Visually Guided Robotic Lamprey Using Frame and Event Based Cameras

The computational load associated with computer vision is often prohibitive, and limits the capacity for on-board image analysis in compact mobile robots. Replicating the kind of feature detection and neural processing that animals excel at remains a challenge in most biomimetic aquatic robots. Event-driven sensors use a biologically inspired sensing strategy to eliminate the need for complete frame capture. Systems employing event-driven cameras enjoy reduced latencies, power consumption, bandwidth, and benefit from a large dynamic range. However, to the best of our knowledge, no work has been done to evaluate the performance of these devices in underwater robotics. This work proposes a robotic lamprey design capable of supporting computer vision, and uses this system to validate a computational neuron model for driving anguilliform swimming. The robot is equipped with two different types of cameras: frame-based and event-based cameras. These were used to stimulate the neural network, yielding goal-oriented swimming. Finally, a study is conducted comparing the performance of the computational model when driven by the two different types of camera. It was observed that event-based cameras improved the accuracy of swimming trajectories and led to significant improvements in the rate at which visual inputs were processed by the network.

Biologically plausible unsupervised learning in shallow and deep neural networks

Bernd Albert Illing

The way our brain learns to disentangle complex signals into unambiguous concepts is fascinating but remains largely unknown. There is evidence, however, that hierarchical neural representations play a key role in the cortex. This thesis investigates biologically plausible models of unsupervised learning of hierarchical representations as found in the brain and modern computer vision models. We use computational modeling to address three main questions at the intersection of artificial intelligence (AI) and computational neuroscience.The first question is: What are useful neural representations and when are deep hierarchical representations needed? We approach this point with a systematic study of biologically plausible unsupervised feature learning in a shallow 2-layer networks on digit (MNIST) and object (CIFAR10) classification. Surprisingly, random features support high performance, especially for large hidden layers. When combined with localized receptive fields, random feature networks approach the performance of supervised backpropagation on MNIST, but not on CIFAR10. We suggest that future models of biologically plausible learning should outperform such random feature benchmarks on MNIST, or that such models should be evaluated in different ways.The second question is: How can hierarchical representations be learned with mechanisms supported by neuroscientific evidence? We cover this question by proposing a unifying Hebbian model, inspired by common models of V1 simple and complex cells based on unsupervised sparse coding and temporal invariance learning. In shallow 2-layer networks, our model reproduces learning of simple and complex cell receptive fields, as found in V1. In deeper networks, we stack multiple layers of Hebbian learning but find that it does not yield hierarchical representations of increasing usefulness. From this, we hypothesise that standard Hebbian rules are too constrained to build increasingly useful representations, as observed in higher areas of the visual cortex or deep artificial neural networks.The third question is: Can AI inspire learning models that build deep representations and are still biologically plausible? We address this question by proposing a learning rule that takes inspiration from neuroscience and recent advances in self-supervised deep learning. The proposed rule is Hebbian, i.e. only depends on pre- and post-synaptic neuronal activity, but includes additional local factors, namely predictive dendritic input and widely broadcasted modulation factors. Algorithmically, this rule applies self-supervised contrastive predictive learning to a causal, biological setting using saccades. We find that networks trained with this generalised Hebbian rule build deep hierarchical representations of images, speech and video.We see our modeling as a potential starting point for both, new hypotheses, that can be tested experimentally, and novel AI models that could benefit from added biological realism.

EPFL2021

Dynamic Perception after Visually Guided Grasping by a Human-Like Autonomous Robot

Kuniaki Noda, Mototaka Suzuki

We will explore dynamic perception following the visually guided grasping of several objects by a human-like autonomous robot. This competency serves for object categorization. Physical interaction with the hand-held object gives the neural network of the robot the rich, coherent and multi-modal sensory input. Multi-layered self-organizing maps are designed and examined in static and dynamic conditions. The results of the tests in the former condition show its capability of robust categorization against noise. The network also shows better performance than a single-layered map does. In the latter condition we focus on shaking behavior by moving only the forearm of the robot. In some combinations of grasping style and shaking radius the network is capable of categorizing two objects robustly. The results show that the network capability to achieve the task largely depends on how to grasp and how to move the objects. These results together with a preliminary simulation are promising toward the self-organization of a high degree of autonomous dynamic object categorization.

2006

Constructive Training Methods for Feedforward Neural Networks with Binary Weights

Frédéric Aviolat

Quantization of the parameters of a Perceptron is a central problem in hardware implementation of neural networks using a numerical technology. A neural model with each weight limited to a small integer range will require little surface of silicon. Moreover, according to Occam's razor principle, better generalization abilities can be expected from a simpler computational model. The price to pay for these benefits lies in the difficulty to train these kind of networks. This paper proposes essentially two new ideas for constructive training algorithms, and demonstrates their efficiency for the generation of feedforward networks composed of Boolean threshold gates with discrete weights. A proof of the convergence of these algorithms is given. Some numerical experiments have been carried out and the results are presented in terms of the size of the generated networks and of their generalization abilities.

1996

Biologically plausible unsupervised learning in shallow and deep neural networks

Bernd Albert Illing

EPFL2021