Neural computations with precise spiking motifs for dynamic vision

Overview & Synopsis

Our brains are extremely efficient at solving highly complex visual tasks. In a few hundred milliseconds, we are able to recognise different objects invariant to various characteristics, such as their size or orientation. Recently, artificial neural networks have made great strides in solving the tasks faced by biological systems. They draw on knowledge from neuroscience to form biologically realistic learning architectures that could provide us with interesting insights into how the human brain works. But these architectures still face a number of challenges: the models are not always interpretable, they do not necessarily seem to use the same strategies as their biological equivalents and they are very energy-intensive. We believe that one of the reasons why the visual system is so efficient is that it uses short pulses to represent information: the action potentials, or spikes, emitted by neurons. Using a neuromorphic approach, the aim of this thesis project is to develop visual information processing models using representations based on spikes, binary events described only by their time and origin. We have chosen to use a dynamic signal, captured by an event-based camera, which transcribes a visual scene using only events, or spikes. We solve visual cognitive tasks using the temporal code formed by precise sequences of events that we call spiking motifs. A large body of experimental evidence suggests that the temporal code carried by these patterns is a strategy used by the brain to encode visual information. We will see that the use of these patterns makes it possible to develop local and biologically realistic learning methods while dynamically and asynchronously processing the events characterising a visual scene. We show that these algorithms can solve an object recognition task and a motion estimation task ultra-fast and efficiently. We also observe the emergence of an organisation of receptive fields similar to that of biological systems, suggesting that a similar strategy may be employed by the brain. In the final part of this work, we will detail the development of a new algorithm for detecting this type of activity in recordings of real neurons.