TIdDLe Forum

CerCo Conference - Friday, 16 April 2021, 9:00 AM
Title: Predictive Error-driven Learning in the Brain
Speaker: Randall O'Reilly (Center for Neuroscience, University of California Davis)

Abstract: How do humans learn from raw sensory experience?  Throughout life, but most obviously in infancy, we learn without explicit instruction.  We propose a detailed biological mechanism for the widely-embraced idea that learning is driven by the differences between predictions and actual outcomes (i.e., predictive error-driven learning).  Specifically, numerous weak projections into the pulvinar nucleus of the thalamus generate top-down predictions, and sparse, focal driver inputs from lower areas supply the actual outcome, originating in layer 5 intrinsic bursting (5IB) neurons.  Thus, the outcome representation is only briefly activated, roughly every 100 ms (i.e., 10 Hz, alpha), resulting in a temporal difference error signal, which drives local synaptic changes throughout the neocortex.  This results in a biologically-plausible form of error backpropagation learning.  We implemented these mechanisms in a large-scale model of the visual system, and found that the simulated inferotemporal (IT) pathway learns to systematically categorize 3D objects according to invariant shape properties, based solely on predictive learning from raw visual inputs.  These categories match human judgments on the same stimuli, and are consistent with neural representations in IT cortex in primates.  The same learning mechanisms apply to sensory-motor predictive learning, and are being tested in an immersive rodent-level navigation and foraging environment, and in speech perception and production.

Zoom link:
https://univ-tlse3-fr.zoom.us/j/8504047 … ZPWm02QT09

Meeting ID: 850 4047 0167
Passcode: 616413

Mathieu Chalvidal, ANITI PhD student in the chairs of Thomas Serre and Rufin VanRullen, will present his recent work on:

"A control perspective on parameterization in infinite-depth neural networks''.
Abstract: The perspective of artificial neural modules as mathematical functions describing a vector field over a topological activity space opens interesting perspectives for deep learning: This view interprets the sequential computation of stacked or recurrent neural modules in traditional deep neural networks as the discrete evolution of a dynamical system, such that a particular parametrization of the module gives rise to particular trajectories of hidden state/activity tensor given their initial point (input stimuli). Recently, new approaches have developed further this view suggesting to solve the ordinary differential equation (Neural ODE) associated with the said module function, such that model depth corresponds to integration over time. This interpretation culminates in the assimilation of neural networks as particular flows. Despite their elegant formulation and lightweight memory cost, neural ordinary differential equations (NODEs) suffer from known representatio!
nal limitations. In particular, the single flow learned by NODEs cannot express all homeomorphisms from a given data space to itself, and their static weight parametrization restricts the type of functions they can learn compared to discrete architectures with layer-dependent weights. Here, we describe a new module called neurally- controlled ODE (N-CODE) designed to improve the expressivity of NODEs. The parameters of N-CODE modules are dynamic variables governed by a trainable map from initial or current activation state, resulting in forms of open-loop and closed-loop control, respectively.

Anyone is welcome to join this ANITI seminar in-person at the LAAS (just say you come for the ANITI talk). The seminar will also be accessible through the link:: https://seminar.laas.fr/b/gui-lyf-mp4-6v7 (no need for a password,  just identify yourself with a name or pseudo to enter the virtual room).

Title: Deep Learning: Promise and Challenges - A Bayesian Perspective
By:    Dr. Nidhal Carla Bouaynaya, Rowan University (USA)

Abstract: Within the field of machine learning, deep learning approaches have resulted in state-of-the-art accuracy in visual object detection, speech recognition, and many other domains including transportation. Deep learning techniques hold the promise of emerging technologies, such as autonomous unmanned vehicles, smart cities infrastructure, and cybersecurity. However, deep learning models are deterministic, and as a result are unable to understand or assess their uncertainty, a critical part of any predictive system’s output. This can have disastrous consequences, especially when the output of such models is fed into higher-level decision making procedures, such as autonomous drones and vehicles. This talk is divided into two parts. First, we provide intuitive insights into deep learning models and show their applications in flight safety and operational efficiency. We then introduce Bayesian deep learning to assess the model’s confidence in its prediction and show pioneering results on robustness to noise and artifacts in the data as well as resilience to adversarial attacks.

Location: ENAC amphi Breguet **Valid identification required for entry**

Title: A quick overview of deep learning theory: part 2
By:    François Malgouyres, IMT

Abstract: The recent successful practical results in deep learning have triggered a renewed interest in deep learning theory. In this talk we will briefly review old and new results about optimization algorithms and guarantees for feed-forward neural networks

Location: UPS amphi 1R3-Schwartz

Note: These introductory lectures, organized by IMT and the AOC group, are directed to an audience of deep learning beginners *with an already solid mathematical background*.

Title: An Introduction to Deep Reinforcement Learning
By:    Emmanuel Rachelson, ISAE-SUPAERO.

Abstract: Deep Reinforcement Learning has become a hot topic over the recent years. This talk is designed for those who want to discover the topic from a rigorous (yet illustrative) point of view. Here is a tentative outline :
+Introduction [target duration : 5 minutes],
+Deep RL is RL, but deep. So we need an overview of RL foundations [target duration : 20 minutes],
+Deep Q-networks [target duration : 10 minutes],
+Towards the state of the art / hot topics [target duration : 10 minutes].

Location: UPS amphi 1R3-Schwartz

Note: These introductory lectures, organized by IMT and the AOC group, are directed to an audience of deep learning beginners *with an already solid mathematical background*.

Title: Spike-based computing and learning in brains, machines, and visual systems in particular
By:    Tim Masquelier, CerCo (CNRS).

Abstract: In recent years, deep learning has been a revolution in the field of machine learning, for computer vision in particular. In this approach, a deep (multilayer) artificial neural network (ANN) is usually trained in a supervised manner, using a stochastic gradient descent method known as backpropagation. Huge amounts of labeled examples are required, but the resulting classification accuracy is truly impressive, often outperforming humans. Neurons in an ANN are characterized by a single, static, continuous-valued activation. Yet biological neurons use discrete electrical impulses called spikes to compute and transmit information, and the spike times, in addition to the spike rates, matter. Spiking neural networks (SNNs) are thus more biologically realistic than ANNs, and arguably the only viable option if one wants to understand how the brain computes. SNNs are also more hardware friendly and energy-efficient than ANNs, and are thus appealing for technology, especially for edge computing. However, training deep SNNs remains a challenge. Spiking neurons' transfer function is usually non-differentiable, which prevents using classic backpropagation. I will review recent supervised and unsupervised methods to train deep SNNs, and compare them in terms of accuracy, but also computational cost and hardware friendliness.

Location: ISAE-SUPAERO, room 05.035
Access to the ISAE-SUPAERO campus requires that you send an email to emmanuel.rachelson@isae-supaero.fr at least the day before the seminar.