[CTN] CTN seminar: Dr. Julien Catanese and Dr. James Bergstra, Dec 18, PAS 2464, 3:30

[Matthijs van der Meer]

Dear all,

Please join us for next Tuesday's CTN seminar (Dec 18), a double-header
featuring James Bergstra and Julien Catanese, both new postdocs at the CTN.

Titles and abstracts of their half-length talks follow below, showing
the exciting experience and perspective they bring to the CTN.

Time and place are the usual, 3.30pm on Tuesday in PAS 2464.

If you would like to meet with the speaker(s), please let me know.

Hope to see you all there!

- Matt


James Bergstra
Banting Postdoctoral Fellow
Computational Neuroscience Research Group, University of Waterloo

Title: For Image Classification, Complex is Better than Simple

Abstract:
Deep networks, popular in machine learning, are composed entirely of
model neurons similar to rate-coding simple cells, but
neurophysiological recordings suggest that very few cells even in area
V1 are truly so simple. One hypothesis for why neurons have more
complex response functions is that this additional flexibility makes
them more useful. In this talk, I will present my doctoral work on the
utility of complex cell models in neural networks for image
classification and image modelling. Neural networks that incorporate
squared filter responses outperform networks whose feature activation
functions use other exponents, when trained by supervised learning.
Notably they outperform activation functions with linear filter
responses, which are typical of multilayer perceptrons and arise
naturally in restricted Boltzmann machines (RBMs). These purely
supervised networks can be improved by two pre-training techniques:
(1) slow feature learning and (2) maximum likelihood training in the
context of the spike-and-slab restricted Boltzmann machine (ssRBM).
Sampled images illustrate that the ssRBM captures important perceptual
elements of natural images, suggesting that it is a good probability
function for modelling natural images, in addition to being a valuable
pre-training algorithm. These results support the hypothesis that
complex cells are a good basis for image classification, and both
supervised and unsupervised learning criteria are consistent with
discriminative objectives and Gabor-like filter properties in visual
area V1.

This work spans several publications, and includes joint work with
Yoshua Bengio, Jerome Louradour, Aaron Courville, and Guillaume
Desjardins.


Julien Catanese
FYSSEN Postdoctoral Fellow
van der Meer lab, University of Waterloo

Title:
Dynamics of decision-related neuronal activity in Hippocampus

Abstract:
Place-selective activity in hippocampal neurons can be modulated by the
trajectory that will be taken in the immediate future ("prospective
coding"), information that could be useful in neural processes
elaborating choices in route planning. To determine if and how
hippocampal prospective neurons are related to decision making, we
measured the time course of the evolution of prospective activity by
recording place responses in rats performing a T-maze alternation task.
After five or seven alternation trials, the routine was unpredictably
interrupted by a photodetector-triggered visual cue as the rat crossed
the middle of central arm, signaling it to suddenly change its intended
choice.

Comparison of the delays between light cue presentation and the onset of
prospective activity for neurons with firing fields at various locations
after the trigger point revealed a 420 ms processing delay. This
surprisingly long delay indicates that prospective activity in the
hippocampus appears much too late to generate planning or decision
signals. This provides yet another example of a prominent brain activity
that is unlikely to play a functional role in the cognitive function
that it appears to represent (planning future trajectories).
Nonetheless, the hippocampus may provide other contextual information to
areas active at the earliest stages of selecting future paths, which
would then return signals that help establish hippocampal prospective
activity.