SpikeAI: laureat du Défi Biomimétisme (2019)
- Le projet SpikeAI est lauréat de l’appel à projets 2019 Biomimétisme :
The SpikeAI project targets analog computing for artificial intelligence in the form of Spiking Neural Networks (SNNs). Computer vision systems widely rely on artificial intelligence and especially neural network based machine learning, which recently gained huge visibility. The training stage for deep convolutional neural networks is time-consuming and yields enormous energy consumption. In contrast, the human brain has the ability to perform visual tasks with unrivaled computational and energy efficiency. It is believed that one major factor of this efficiency is the fact that information is vastly represented by short pulses (spikes) at analog –not discrete– times. However, computer vision algorithms using such representation still lack in practice, and its high potential is largely underexploited. Inspired from biology, the project addresses the scientific question of developing a low-power, end-to-end analog sensing and processing architecture. This will be applied on the particular context of a field programmable analog array (FPAA) applied to a stereovision system dedicated to coastal surveillance using an aerial robot of 3D visual scenes, running on analog devices, without a central clock and to validate them in real-life situations. The ambitious long-term vision of the project is to develop the next generation AI paradigm that will at term compete with deep learning. We believe that neuromorphic computing, mainly studied in EU countries, will be a key technology in the next decade. It is therefore both a scientific and strategic challenge for France and EU to foster this technological breakthrough. This call will help kickstart collaboration within this European consortium to help leverage the chance to successfully apply to future large-scale grant proposals (e.g. ANR, CHIST-ERA, ERC).
- Get more information.
In computer vision, the visual search task consists in extracting a scarce and specific visual information (the target) from a large and crowded visual display. This task is usually implemented by scanning the different possible target identities at all possible spatial positions, hence with strong computational load. The human visual system employs a different strategy, combining a foveated sensor with the capacity to rapidly move the center of fixation using saccades. Saccade-based visual exploration can be idealized as an inference process, assuming that the target position and category are independently drawn from a common generative process. Knowing that process, visual processing is then separated in two specialized pathways, the where pathway mainly conveying information about target position in peripheral space, and the what pathway mainly conveying information about the category of the target. We consider here a dual neural network architecture learning independently where to look and then at what to see. This allows in particular to infer target position in retinotopic coordinates, independently to its category. This framework was tested on a simple task of finding digits in a large, cluttered image. Simulation results demonstrate the benefit of specifically learning where to look before actually knowing the target category. The approach is also energy-efficient as it includes the strong compression rate performed at the sensor level, by retina and V1 encoding, which is preserved up to the action selection level, highlighting the advantages of bio-mimetic strategies with regards to traditional computer vision when computing resources are at stake.