ELFT HÍRADÓ

[Janos Asboth <asboth.janos AT wigner.mta.hu>]

Meghívó

BME Elméleti Fizika Szeminárium,

november 15. péntek 10h15,

1111 Budapest, Budafoki út 8., BME F III. magasföldszint 01.,
Elméleti Fizika Tanszék szemináriumi szoba

Orbán Gergő (Wigner FK, Komputációs Tudományok Osztálya):

Efficiently readable codes through humble nonlinearities

A remarkable parallel between artificial and biological learning systems is
that both feature an elementary nonlinearity, the firing threshold: the
linear response of a unit (an artificial or a biological neuron) is fed
through a firing rate nonlinearity, which clips the response of the unit
under a certain threshold. We discuss insights obtained from neuroscience
why this humble nonlinearity is effective in supporting cutting-edge
performance in applications ranging from scientific image analysis to
multi-billion dollar commercial applications. We approach this question
from an appealing idea that is phrased in neuroscience as representational
untangling: the idea that high dimensional signals that are hopelessly
nonlinearly entangled become linearly separable through computational
processes. Signatures of such computations can be identified in the visual
cortex where complex image content such as the identity of faces can be
linearly decoded irrespective of nuisance variables, such as pose,
lighting, or orientation. It remains a burning question, however, what
elementary computations can contribute to representational untangling. We
point out that the  basic but ubiquitous form of local nonlinearity the
firing threshold can achieve untangling under nuisance variable
uncertainty. We argue that the efficiency of this nonlinearity in achieving
easily readable codes lies in its capability to balance between two
computational goals: preservation of information and sparsification of
neural responses. We show through recordings from visual cortical neurons
that the threshold of biological neurons is in a range that is optimal for
representational untangling..


Minden érdeklődőt szeretettel várunk.

Asbóth János
szemináriumi koordinátor