Líneas de
Investigación

RL1 /
Aprendizaje Profundo para Visión y Lenguaje

Nuevas teorías y métodos para continuar desentrañando el potencial del Aprendizaje Profundo para crear sistemas cognitivos avanzados con un enfoque en la visión y el lenguaje.

RL2 /
IA neuro-simbólica

Integración de la IA lógica-probabilística y la basada en el aprendizaje profundo, invocando mutuamente las soluciones de cada parte, inyectando y utilizando la semántica en el aprendizaje profundo

RL3 /
IA inspirada en el cerebro

Reunir a científicos de la neurociencia, la psicología cognitiva y la IA para explotar los conocimientos de las operaciones anatómicas y cognitivas de los cerebros biológicos para iluminar a los investigadores de la IA.

RL4 /
Aprendizaje automático basado en la física

Reunir a matemáticos, físicos y científicos de la IA para explotar los conocimientos de las ciencias físicas para desarrollar modelos de aprendizaje de máquina basados en relaciones causales.

RL5 /
IA centrada en el ser humano

Nuevas tecnologías para un uso justo, seguro y transparente de la IA en la sociedad, así como metodologías para evaluar su impacto en la misma. Promover nuevas herramientas para una IA interpretable y explicable.

Líderes / Marcelo Mendoza

Publicaciones

RL3, Publisher: Heliyon, Link>

AUTHORS

Rodrigo Vergara, Pedro Maldonado, Rocío Loyola-Navarro, Alexandre Hyafil, Cristobal Moenne, Francisco Aboitiz

ABSTRACT

The ability of an organism to voluntarily control the stimuli onset modulates perceptual and attentional functions. Since stimulus encoding is an essential component of working memory (WM), we conjectured that controlling the initiation of the perceptual process would positively modulate WM. To corroborate this proposition, we tested twenty-five healthy subjects in a modified-Sternberg WM task under three stimuli presentation conditions: an automatic presentation of the stimuli, a self-initiated presentation of the stimuli (through a button press), and a self-initiated presentation with random-delay stimuli onset. Concurrently, we recorded the subjects' electroencephalographic signals during WM encoding. We found that the self-initiated condition was associated with better WM accuracy, and earlier latencies of N1, P2 and P3 evoked potential components representing visual, attentional and mental review of the stimuli processes, respectively. Our work demonstrates that self-initiated stimuli enhance WM performance and accelerate early visual and attentional processes deployed during WM encoding. We also found that self-initiated stimuli correlate with an increased attentional state compared to the other two conditions, suggesting a role for temporal stimuli predictability. Our study remarks on the relevance of self-control of the stimuli onset in sensory, attentional and memory updating processing for WM.


62 visualizaciones Ir a la publicación

RL3, Publisher: Springer Series in Computational Neuroscience, Link>

AUTHORS

Pedro Maldonado

ABSTRACT

In Chap. 11, Pedro Maldonado describes his joint work with his PhD advisor, George Gerstein, demonstrating plasticity in receptive field properties, neuronal interactions, and network dynamics in the rat auditory cortex upon electrical intracortical microstimulation.


74 visualizaciones Ir a la publicación

RL3, Publisher: Sustainability, Link>

AUTHORS

Nicolás Szoloch, Mauricio Varas, Pedro Maldonado, Raúl Pezoa, Franco Basso

ABSTRACT

A great deal of research has examined the efficacy of variable message signs (VMS) to induce driver behavior changes, improve safety conditions, and decongest the traffic network. However, there is little literature regarding the most effective ways to display this information on VMS. Furthermore, none of the previous contributions have concentrated on analyzing what impact flashing VMS have on drivers by using real traffic data. This article seeks to bridge this gap, analyzing the effect of incorporating intermittent light stimulation to messages on drivers’ behavior on a Chilean highway, using vehicle-by-vehicle data obtained in a non-intrusive way. In order to do so, an experiment was carried out to measure the responses of drivers when faced with two types of messages: (1) those intended to induce a speed reduction and (2) those aimed at generating lane changes. From the statistical models we obtained several insights. Our results show that flashing messages may increase the effectiveness of VMS depending on environmental and traffic conditions. In particular, for speed moderation messages, we found 12 significant effects, showing, for example, that a flashing message is most effective in the hours of darkness, with low congestion, small spacing, and low average speeds. Additionally, it has a more significant impact on experienced drivers. On the other hand, for lane change messages, we found five significant effects, showing that flashing messaging reduces its effectiveness in situations where a high cognitive load is required, such as in high flow and high average speeds. No particular effects were identified in either case for specific vehicle types.


55 visualizaciones Ir a la publicación

RL3, Publisher: Frontiers in Neuroscience, Link>

AUTHORS

Martin Irani, Marion Inostroza, María A. García - Pérez, Vicente Tiznado, Tamara Bustamante, Pedro Maldonado, José L. Valdés

ABSTRACT

Hippocampal-dependent memories emerge late during postnatal development, aligning with hippocampal maturation. During sleep, the two-stage memory formation model states that through hippocampal-neocortical interactions, cortical slow-oscillations (SO), thalamocortical Spindles, and hippocampal sharp-wave ripples (SWR) are synchronized, allowing for the consolidation of hippocampal-dependent memories. However, evidence supporting this hypothesis during development is still lacking. Therefore, we performed successive object-in-place tests during a window of memory emergence and recorded in vivo the occurrence of SO, Spindles, and SWR during sleep, immediately after the memory encoding stage of the task. We found that hippocampal-dependent memory emerges at the end of the 4th postnatal week independently of task overtraining. Furthermore, we observed that those animals with better performance in the memory task had increased Spindle density and duration and lower density of SWR. Moreover, we observed changes in the SO-Spindle and Spindle-SWR temporal-coupling during this developmental period. Our results provide new evidence for the onset of hippocampal-dependent memory and its relationship to the oscillatory phenomenon occurring during sleep that helps us understand how memory consolidation models fit into the early stages of postnatal development.


62 visualizaciones Ir a la publicación

RL3, Publisher: Scientific Reports, Link>

AUTHORS

Yukako Yamane, Junji Ito, Sonja Grün, Pedro Maldonado, Cristian Joana, Hiroshi Tamura, Ichiro Fujita

ABSTRACT

In natural vision, neuronal responses to visual stimuli occur due to self-initiated eye movements. Here, we compare single-unit activity in the primary visual cortex (V1) of non-human primates to flashed natural scenes (passive vision condition) to when they freely explore the images by self-initiated eye movements (active vision condition). Active vision enhances the number of neurons responding, and the response latencies become shorter and less variable across neurons. The increased responsiveness and shortened latency during active vision were not explained by increased visual contrast. While the neuronal activities in all layers of V1 show enhanced responsiveness and shortened latency, a significant increase in lifetime sparseness during active vision is observed only in the supragranular layer. These findings demonstrate that the neuronal responses become more distinct in active vision than passive vision, interpreted as consequences of top-down predictive mechanisms.


45 visualizaciones Ir a la publicación

RL3, Publisher: eNeuro Link>

AUTHORS

Miriam Schwalm, José I. Egaña, Rodrigo Montefusco-Siegmund, Pedro Maldonado, Christ Devia, Eduardo Rosales Jubal

ABSTRACT

Variations in human behavior correspond to the adaptation of the nervous system to different internal and environmental demands. Attention, a cognitive process for weighing environmental demands, changes over time. Pupillary activity, which is affected by fluctuating levels of cognitive processing, appears to identify neural dynamics that relate to different states of attention. In mice, for example, pupil dynamics directly correlate with brain state fluctuations. Although, in humans, alpha-band activity is associated with inhibitory processes in cortical networks during visual processing, and its amplitude is modulated by attention, conclusive evidence linking this narrowband activity to pupil changes in time remains sparse. We hypothesize that, as alpha activity and pupil diameter indicate attentional variations over time, these two measures should be comodulated. In this work, we recorded the electroencephalographic (EEG) and pupillary activity of 16 human subjects who had their eyes fixed on a gray screen for 1 min. Our study revealed that the alpha-band amplitude and the high-frequency component of the pupil diameter covariate spontaneously. Specifically, the maximum alpha-band amplitude was observed to occur ∼300 ms before the peak of the pupil diameter. In contrast, the minimum alpha-band amplitude was noted to occur ∼350 ms before the trough of the pupil diameter. The consistent temporal coincidence of these two measurements strongly suggests that the subject’s state of attention, as indicated by the EEG alpha amplitude, is changing moment to moment and can be monitored by measuring EEG together with the diameter pupil.


48 visualizaciones Ir a la publicación

RL3, Publisher: Frontiers in Systems Neuroscience Link>

AUTHORS

Pedro Maldonado, Mario Villalobos, Rodrigo Vergara, Sergio Vicencio-Jimenez

ABSTRACT

It is still elusive to explain the emergence of behavior and understanding based on its neural mechanisms. One renowned proposal is the Free Energy Principle (FEP), which uses an information-theoretic framework derived from thermodynamic considerations to describe how behavior and understanding emerge. FEP starts from a whole-organism approach, based on mental states and phenomena, mapping them into the neuronal substrate. An alternative approach, the Energy Homeostasis Principle (EHP), initiates a similar explanatory effort but starts from single-neuron phenomena and builds up to whole-organism behavior and understanding. In this work, we further develop the EHP as a distinct but complementary vision to FEP and try to explain how behavior and understanding would emerge from the local requirements of the neurons. Based on EHP and a strict naturalist approach that sees living beings as physical and deterministic systems, we explain scenarios where learning would emerge without the need for volition or goals. Given these starting points, we state several considerations of how we see the nervous system, particularly the role of the function, purpose, and conception of goal-oriented behavior. We problematize these conceptions, giving an alternative teleology-free framework in which behavior and, ultimately, understanding would still emerge. We reinterpret neural processing by explaining basic learning scenarios up to simple anticipatory behavior. Finally, we end the article with an evolutionary perspective of how this non-goal-oriented behavior appeared. We acknowledge that our proposal, in its current form, is still far from explaining the emergence of understanding. Nonetheless, we set the ground for an alternative neuron-based framework to ultimately explain understanding.


34 visualizaciones Ir a la publicación

RL3, Publisher: The European journal of neuroscience Link>

AUTHORS

Marcos E. Orchard, Pedro Maldonado, Taiki Harada, Rodrigo Vergara, Ismael Jaras

ABSTRACT

It is widely accepted that the brain, like any other physical system, is subjected to physical constraints that restrict its operation. The brain's metabolic demands are particularly critical for proper neuronal function, but the impact of these constraints continues to remain poorly understood. Detailed single-neuron models are recently integrating metabolic constraints, but these models’ computational resources make it challenging to explore the dynamics of extended neural networks, which are governed by such constraints. Thus, there is a need for a simplified neuron model that incorporates metabolic activity and allows us to explore the dynamics of neural networks. This work introduces an energy-dependent leaky integrate-and-fire (EDLIF) neuronal model extension to account for the effects of metabolic constraints on the single-neuron behavior. This simple, energy-dependent model could describe the relationship between the average firing rate and the Adenosine triphosphate (ATP) cost as well as replicate a neuron's behavior under a clinical setting such as amyotrophic lateral sclerosis (ALS). Additionally, EDLIF model showed better performance in predicting real spike trains – in the sense of spike coincidence measure – than the classical leaky integrate-and-fire (LIF) model. The simplicity of the energy-dependent model presented here makes it computationally efficient and, thus, suitable for studying the dynamics of large neural networks.


46 visualizaciones Ir a la publicación

RL2, Publisher: Journal of Machine Learning Research, Link>

AUTHORS

Pablo Barceló, Jorge Pérez, Javier Marinkovic

ABSTRACT

Alternatives to recurrent neural networks, in particular, architectures based on self-attention, are gaining momentum for processing input sequences. In spite of their relevance, the computational properties of such networks have not yet been fully explored. We study the computational power of the Transformer, one of the most paradigmatic architectures exemplifying self-attention. We show that the Transformer with hard-attention is Turing complete exclusively based on their capacity to compute and access internal dense representations of the data. Our study also reveals some minimal sets of elements needed to obtain this completeness result.


247 visualizaciones Ir a la publicación

RL2, Publisher: arXiv, Link>

AUTHORS

Pablo Barceló, Mikaël Monet, Marcelo Arenas, Leopoldo Bertossi

ABSTRACT

In Machine Learning, the SHAP-score is a version of the Shapley value that is used to explain the result of a learned model on a specific entity by assigning a score to every feature. While in general computing Shapley values is an intractable problem, we prove a strong positive result stating that the SHAP-score can be computed in polynomial time over deterministic and decomposable Boolean circuits. Such circuits are studied in the field of Knowledge Compilation and generalize a wide range of Boolean circuits and binary decision diagrams classes, including binary decision trees and Ordered Binary Decision Diagrams (OBDDs). We also establish the computational limits of the SHAP-score by observing that computing it over a class of Boolean models is always polynomially as hard as the model counting problem for that class. This implies that both determinism and decomposability are essential properties for the circuits that we consider. It also implies that computing SHAP-scores is intractable as well over the class of propositional formulas in DNF. Based on this negative result, we look for the existence of fully-polynomial randomized approximation schemes (FPRAS) for computing SHAP-scores over such class. In contrast to the model counting problem for DNF formulas, which admits an FPRAS, we prove that no such FPRAS exists for the computation of SHAP-scores. Surprisingly, this negative result holds even for the class of monotone formulas in DNF. These techniques can be further extended to prove another strong negative result: Under widely believed complexity assumptions, there is no polynomial-time algorithm that checks, given a monotone DNF formula φ and features x,y, whether the SHAP-score of x in φ is smaller than the SHAP-score of y in φ.


288 visualizaciones Ir a la publicación