Gast ::
| Hauptseite > Publikationsdatenbank > Spike patterns in nature and AI |
| Hauptseite > Publikationsdatenbank > Spike patterns in nature and AI |
| Talk (non-conference) (Invited) | FZJ-2025-03437 |
2024
Abstract: The energy consumption of present AI systems is unsustainable and undemocratic. Understanding the energy efficiency of the brain may uncover pathways out of the dilemma. A promising observation is that unlike artificial neural networks, mammalian brains are sparse in space and time. Spatial sparseness means that neurons are only connected to a tiny fraction of the other neurons, temporal sparseness means that on the few connections there is activity only a few times per second. This architecture suggests that the brain relies on the coordinated activity of populations of neurons to process sensory information and organize behavior.In the past three decades computational neuroscience has learned how to extract and interpret the dynamics of spatio-temporal patterns in massively parallel spike trains. Despite this progress it is still unclear how function and in particular learning is implemented by the mammalian cortex. Reasons may be the undersampling of the system and the limited capability to observe the system during learning.Ironically in artificial neural networks we have a similar situation. These networks can learn complex tasks with high accuracy. But our understanding of what has been learned and how the trained network solves the task is limited. Progress has already been made by transferring ideas of generic learning algorithms for artificial neural networks to spiking systems. This gives us benchmarks on what we can find out about such a system with statistical tools, but the analysis may also give inspiration on the working principles of the natural system.The talk first reviews the state-of-the-art of the time-resolved analysis of spike patterns required for undersampled brain data. Subsequently, the talk reports on the analysis of a spike-based artificial neural network trained on the MNIST data set, which exhibits precise spike timing.
|
The record appears in these collections: |