Learning and Neuroscience

Neural networks: heuristics and theory

Understanding our human brain better is in itself one of the greatest challenges of our century, but at the same time mimicking the efficiency and robustness by which it represents information has become a core challenge in artificial intelligence research.
After more than twenty years of limited success, the increasing computational power of modern computers is at the present day showing how “back-propagation” algorithms on deep architectures (i.e. networks made of more than two layers of interconnected neurons) can achieve state-of-art performances in most of the tasks where “intelligence” is involved, such as speech and image recognition, where men still outperform computers.
In this talk I will try to give an overview of the mainstream approach in this research field, and then advance some motivations for the slightly different viewpoint adopted by our group, describing briefly the work which we are carrying out in the binary-valued setting.

Thu, 06/11/2014 - 14:30
Luca Saglietti
HuGeF, Via Nizza 52, 1st floor, old building.

BP for the diluted perceptron

dilperc.tgz14.1 KB

This is code for biological data analysis using a multi-state diluted discrete perceptron.
It contains also a matlab mex implementation.

Automated brain image segmentation with convolutional networks: a review

The image segmentation problem, applied to neural tissues, consists in identifying the individual structures (cell bodies, dendrites and axons) in a 3-dimensional scannerized image of a brain portion. Having a good automatic procedure for performing this task (i.e. as good as a human expert is) with good scaling properties would be useful to be able to reconstruct the connection structure of large portions of the brain (the so called "connectome"), which is supposedly a crucial part of the information needed to understand the brain functioning.

Wed, 21/10/2009 - 16:30
Carlo Baldassi
Syndicate content