Tuesday, July 2015, 21st
"New Ideas in Theoretical Neuroscience"
The symposium will take place at Salle Langevin, 29 rue d'Ulm, starting at 14:00.
14:00 - 14:45 'Reading the mind of the worm'
14:45 - 15:30 'Sequence generating recurrent neural
networks: a novel view of cortex, thalamus,
and the basal ganglia'
15:30 - 15:45 Coffee break
15:45 - 16:30 'Contextual modulation of gamma rhythms in
inhibition stabilized cortical networks'
Reading the mind of the worm
EMBO Fellow, Zimmer Group, IMP, Vienna
If we could read the activity of every neuron of an
organism’s brain at the same time, could we decipher its
thoughts? More tangibly, we could be in a position to ask
how holistic behavior arises from dynamic brain activity.
While individual motor actions have been correlated with
dynamical activities of neuronal sub-networks, it remains
unknown how the brain coordinates these activities to
produce coherent action sequences. Here, we perform
brain-wide single-cell resolution imaging in Caenorhabditis
elegans and find that the neural population exhibits a
widely shared, low-dimensional signal that evolves
cyclically on an attractor-like manifold, the geometry of
which defines the assembly of commands into action
sequences, including decisions between alternative actions.
This study establishes, for the first time in any animal, a
continuous real-time mapping between neural dynamics and
behavioral dynamics on a single-trial basis.
Sequence generating recurrent neural networks: a novel view
of cortex, thalamus, and the basal ganglia
Columbia University, New York
One view of all complex cognitions and behaviors is that
they are sequences of simpler, more stereotyped components.
For example, peeling a banana involves combining a set of
simple reaches and grasps. In this view, sequence generation
is a fundamental computational task that the brain must
perform. Additionally, the same neural hardware may be
reused for each component of a sequence: evidence from
primary motor cortex reveals that the same network of
neurons can drive multiple different kinds of reaches. If we
map each component of a sequence onto an activity "state" of
a neural network, then sequence generation poses the
following questions: 1) how can networks maintain the
current state, 2) how can they switch between states at
appropriate times, and 3) can the computational functions
underlying sequence generation be mapped onto specific brain
structures? In this talk, I will present a multistate
recurrent neural network (RNN) model that can perform the
computation of sequence generation and that addresses each
of these three questions. The model architecture maps nicely
onto the cortex-basal ganglia-thalamus loop, thus offering
novel insights into the functions of these brain structures,
which have long been known to play a role in sequencing.
Interestingly, this specific architecture has recently been
proposed in the machine learning community as the
"multiplicative RNN", and has been shown to have certain
advantages over standard RNNs which will be discussed. Thus,
it may be proposed that brain implements a multiplicative
RNN as a way to harness these computational advantages.
Contextual modulation of gamma rhythms in inhibition
stabilized cortical networks
University of Oregon, Eugene; Columbia University, New York
Cortical networks feature strong recurrent excitation,
posing them near potential instability. By and large, models
of cortical dynamics have relied on single neuronal
saturation to overcome such instability. However, throughout
the cortical dynamic range, neurons' activity tends to
remain well below their saturation levels, and
correspondingly their empirically measured input-output
functions remain convex and supralinear. Such expansive
nonlinearities at first appear to aggravate the problem of
stability. Nevertheless we have recently shown that strong
recurrent inhibition is sufficient to stabilize cortical
networks against runaway excitation, without relying on
single neuronal saturation (Ahmadian et al. 2013, Rubin et
al. 2015). Moreover, as a consequence, such Stabilized
Supralinear Networks (SSN) provide a robust and parsimonious
mechanistic explanation for a plethora of contextual
modulation phenomena observed across sensory cortical areas.
These include surround suppression and divisive
normalization, recently dubbed a canonical brain computation
(Carandini & Heeger, 2011).
In this talk I will first review these published results.
In the second half, I will focus on ongoing work using the
SSN to model aspects of time-dependent cortical dynamics.
Gamma rhythms are a robust feature of cortical dynamics, and
have been hypothesized to play a central role in various
cognitive tasks. Gamma rhythm characteristics such as their
power and peak frequency, however, exhibit strong
dependencies on stimulus and contextual parameters (it has
in turn been argued that such dependencies may invalidate
some hypothesized computational functions for gamma). I will
describe how SSN is able to robustly account for such
modulations in gamma characteristics. In particular, I will
show how the model explains the particular dependence of
gamma peak frequency on local stimulus contrast and stimulus
size observed in the visual cortex. Time allowing, I will
also elaborate on two possible mechanisms for attentional
modulation of rates in SSN, which lead to opposite effects
on gamma power, as observed, respectively, in V1 vs. higher
visual cortical areas.