-
A Neural Network Model of Continual Learning with Cognitive Control
Authors:
Jacob Russin,
Maryam Zolfaghar,
Seongmin A. Park,
Erie Boorman,
Randall C. O'Reilly
Abstract:
Neural networks struggle in continual learning settings from catastrophic forgetting: when trials are blocked, new learning can overwrite the learning from previous blocks. Humans learn effectively in these settings, in some cases even showing an advantage of blocking, suggesting the brain contains mechanisms to overcome this problem. Here, we build on previous work and show that neural networks e…
▽ More
Neural networks struggle in continual learning settings from catastrophic forgetting: when trials are blocked, new learning can overwrite the learning from previous blocks. Humans learn effectively in these settings, in some cases even showing an advantage of blocking, suggesting the brain contains mechanisms to overcome this problem. Here, we build on previous work and show that neural networks equipped with a mechanism for cognitive control do not exhibit catastrophic forgetting when trials are blocked. We further show an advantage of blocking over interleaving when there is a bias for active maintenance in the control signal, implying a tradeoff between maintenance and the strength of control. Analyses of map-like representations learned by the networks provided additional insights into these mechanisms. Our work highlights the potential of cognitive control to aid continual learning in neural networks, and offers an explanation for the advantage of blocking that has been observed in humans.
△ Less
Submitted 3 November, 2022; v1 submitted 9 February, 2022;
originally announced February 2022.
-
Do grid codes afford generalization and flexible decision-making?
Authors:
Linda Q. Yu,
Seongmin A. Park,
Sarah C. Sweigart,
Erie D. Boorman,
Matthew R. Nassar
Abstract:
Behavioral flexibility is learning from previous experiences and planning appropriate actions in a changing or novel environment. Successful behavioral adaptation depends on internal models the brain builds to represent the relational structure of an abstract task. Emerging evidence suggests that the well-known roles of the hippocampus and entorhinal cortex (HC-EC) in integrating spatial relations…
▽ More
Behavioral flexibility is learning from previous experiences and planning appropriate actions in a changing or novel environment. Successful behavioral adaptation depends on internal models the brain builds to represent the relational structure of an abstract task. Emerging evidence suggests that the well-known roles of the hippocampus and entorhinal cortex (HC-EC) in integrating spatial relationships into cognitive maps can be extended to map the transition structure between states in non-spatial abstract tasks. However, what the EC grid-codes actually compute to afford generalization remains elusive. We introduce two non-exclusive ideas regarding what grid-codes may represent to afford higher-level cognition. One idea is that grid-codes are eigenvectors of the successor representation (SR) learned online during a task. This view assumes that the grid codes serve as an efficient basis function for learning and representing experienced relationships between entities. Subsequently, the grid codes facilitate generalization in novel contexts such as when the goal changes. The second idea is that the grid-codes reflect the inferred global task structure. This view assumes that the grid-code represents a structural code that is factorized from specific sensory content, enabling structural information to be transferred across tasks. Subsequently, the brain could afford one-shot inferences without requiring experience. The ability to generalize experiences and make appropriate decisions in novel situations is critical for both animals and machines. Here we review proposed computations of the grid-code in the brain, which is potentially critical to behavioral flexibility.
△ Less
Submitted 30 June, 2021;
originally announced June 2021.
-
Complementary Structure-Learning Neural Networks for Relational Reasoning
Authors:
Jacob Russin,
Maryam Zolfaghar,
Seongmin A. Park,
Erie Boorman,
Randall C. O'Reilly
Abstract:
The neural mechanisms supporting flexible relational inferences, especially in novel situations, are a major focus of current research. In the complementary learning systems framework, pattern separation in the hippocampus allows rapid learning in novel environments, while slower learning in neocortex accumulates small weight changes to extract systematic structure from well-learned environments.…
▽ More
The neural mechanisms supporting flexible relational inferences, especially in novel situations, are a major focus of current research. In the complementary learning systems framework, pattern separation in the hippocampus allows rapid learning in novel environments, while slower learning in neocortex accumulates small weight changes to extract systematic structure from well-learned environments. In this work, we adapt this framework to a task from a recent fMRI experiment where novel transitive inferences must be made according to implicit relational structure. We show that computational models capturing the basic cognitive properties of these two systems can explain relational transitive inferences in both familiar and novel environments, and reproduce key phenomena observed in the fMRI experiment.
△ Less
Submitted 19 May, 2021;
originally announced May 2021.
-
Experimental results of the laserwire emittance scanner for LINAC4 at CERN
Authors:
T. Hofmann,
G. E. Boorman,
A. Bosco,
E. Bravin,
S. M. Gibson,
K. O. Kruchinin,
U. Raich,
F. Roncarolo,
F. Zocca
Abstract:
Within the framework of the LHC Injector Upgrade (LIU), the new LINAC4 is currently being commissioned to replace the existing LINAC2 proton source at CERN. After the expected completion at the end of 2016, the LINAC4 will accelerate H- ions to 160 MeV. To measure the transverse emittance of the H- beam, a method based on photo-detachment is proposed. This system will operate using a pulsed laser…
▽ More
Within the framework of the LHC Injector Upgrade (LIU), the new LINAC4 is currently being commissioned to replace the existing LINAC2 proton source at CERN. After the expected completion at the end of 2016, the LINAC4 will accelerate H- ions to 160 MeV. To measure the transverse emittance of the H- beam, a method based on photo-detachment is proposed. This system will operate using a pulsed laser with light delivered via an optical fibre and subsequently focused through a thin slice of the H- beam. The laser photons have sufficient energy to detach the outer electron and create H0/e- pairs. In a downstream dipole, the created H0 particles are separated from the unstripped H- ions and their distribution is measured with a dedicated detector. By scanning the focused laser across the H- beam, the transverse emittance of the H- beam can be reconstructed. This paper will first discuss the concept, design and simulations of the laser
△ Less
Submitted 18 November, 2015;
originally announced November 2015.
