MG
M. Guichard
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
2 records found
1
Shortest Paths and Hyperbolic Geometry
Predicting shortest paths in incomplete networks through latent geometry
Finding shortest paths in networks is a ubiquitous problem with applications spanning GPS navigation, finding biological interaction pathways, and internet packet routing. Most existing approaches assume the network is fully observed. In incomplete networks, paths must be predicted, a problem composed of two components: extrapolating information about missing links, and applying this to reconstruct paths that would exist in the complete network.
In this thesis we use hyperbolic geometry to address both components, showing that a network’s full latent geometry can be used to improve path prediction. To recover missing-link information, we adapt D-Mercator, a hyperbolic embedder, to account for uniformly missing links in partially observed networks. This improves inference of the inverse-temperature parameter—important for accurately characterizing the geometry of networks—and results in better link prediction performance.
To then reconstruct paths, we first isolate the nodes that compose them: we evaluate several methods for identifying nearly shortest path nodes, introducing four new methods. We find that the existing distance-to-geodesic and proposed two-hop-contribution methods perform best, deriving an analytic relationship between them. Building on this, we compare a range of path-prediction methods, contrasting hyperbolic and network-based link-prediction baselines against a full hyperbolic approach based on surrogate link probability networks. Geometry-based methods consistently outperform their network-based counterparts, and the surrogate link probability networks offer greater predictive accuracy and noise tolerance than hyperbolic link prediction alone.
Prediction is most accurate in the heterogeneous, low-temperature regimes occupied by real networks such as the protein-protein interaction network. These findings demonstrate that hyperbolic geometry can effectively address both components of path prediction through latent link probabilities, though the accuracy attained underlines how challenging path prediction remains in partially observed networks. ...
In this thesis we use hyperbolic geometry to address both components, showing that a network’s full latent geometry can be used to improve path prediction. To recover missing-link information, we adapt D-Mercator, a hyperbolic embedder, to account for uniformly missing links in partially observed networks. This improves inference of the inverse-temperature parameter—important for accurately characterizing the geometry of networks—and results in better link prediction performance.
To then reconstruct paths, we first isolate the nodes that compose them: we evaluate several methods for identifying nearly shortest path nodes, introducing four new methods. We find that the existing distance-to-geodesic and proposed two-hop-contribution methods perform best, deriving an analytic relationship between them. Building on this, we compare a range of path-prediction methods, contrasting hyperbolic and network-based link-prediction baselines against a full hyperbolic approach based on surrogate link probability networks. Geometry-based methods consistently outperform their network-based counterparts, and the surrogate link probability networks offer greater predictive accuracy and noise tolerance than hyperbolic link prediction alone.
Prediction is most accurate in the heterogeneous, low-temperature regimes occupied by real networks such as the protein-protein interaction network. These findings demonstrate that hyperbolic geometry can effectively address both components of path prediction through latent link probabilities, though the accuracy attained underlines how challenging path prediction remains in partially observed networks. ...
Finding shortest paths in networks is a ubiquitous problem with applications spanning GPS navigation, finding biological interaction pathways, and internet packet routing. Most existing approaches assume the network is fully observed. In incomplete networks, paths must be predicted, a problem composed of two components: extrapolating information about missing links, and applying this to reconstruct paths that would exist in the complete network.
In this thesis we use hyperbolic geometry to address both components, showing that a network’s full latent geometry can be used to improve path prediction. To recover missing-link information, we adapt D-Mercator, a hyperbolic embedder, to account for uniformly missing links in partially observed networks. This improves inference of the inverse-temperature parameter—important for accurately characterizing the geometry of networks—and results in better link prediction performance.
To then reconstruct paths, we first isolate the nodes that compose them: we evaluate several methods for identifying nearly shortest path nodes, introducing four new methods. We find that the existing distance-to-geodesic and proposed two-hop-contribution methods perform best, deriving an analytic relationship between them. Building on this, we compare a range of path-prediction methods, contrasting hyperbolic and network-based link-prediction baselines against a full hyperbolic approach based on surrogate link probability networks. Geometry-based methods consistently outperform their network-based counterparts, and the surrogate link probability networks offer greater predictive accuracy and noise tolerance than hyperbolic link prediction alone.
Prediction is most accurate in the heterogeneous, low-temperature regimes occupied by real networks such as the protein-protein interaction network. These findings demonstrate that hyperbolic geometry can effectively address both components of path prediction through latent link probabilities, though the accuracy attained underlines how challenging path prediction remains in partially observed networks.
In this thesis we use hyperbolic geometry to address both components, showing that a network’s full latent geometry can be used to improve path prediction. To recover missing-link information, we adapt D-Mercator, a hyperbolic embedder, to account for uniformly missing links in partially observed networks. This improves inference of the inverse-temperature parameter—important for accurately characterizing the geometry of networks—and results in better link prediction performance.
To then reconstruct paths, we first isolate the nodes that compose them: we evaluate several methods for identifying nearly shortest path nodes, introducing four new methods. We find that the existing distance-to-geodesic and proposed two-hop-contribution methods perform best, deriving an analytic relationship between them. Building on this, we compare a range of path-prediction methods, contrasting hyperbolic and network-based link-prediction baselines against a full hyperbolic approach based on surrogate link probability networks. Geometry-based methods consistently outperform their network-based counterparts, and the surrogate link probability networks offer greater predictive accuracy and noise tolerance than hyperbolic link prediction alone.
Prediction is most accurate in the heterogeneous, low-temperature regimes occupied by real networks such as the protein-protein interaction network. These findings demonstrate that hyperbolic geometry can effectively address both components of path prediction through latent link probabilities, though the accuracy attained underlines how challenging path prediction remains in partially observed networks.
Backpropagating in time-discretized multi-spike spiking neural networks
How are the training accuracy and training speed (in epochs and time) of a spiking neural network affected when numerically integrating with the forward-Euler and Parker-Sochacki methods?
Spiking neural networks have gained traction as both a tool for neuroscience research and a new frontier in machine learning. A plethora of neuroscience literature exists exploring the realistic simulation of neurons, with complex models re- quiring the formulation and integration of ordinary differential equations. Overcoming this challenge has led to the exploration of various numerical integration techniques with the goal of highly stable and accurate simulations. In contrast, training spiking neural networks is often done with simple leaky integrate-and-fire models and rudimentary integration methods such as the forward-Euler method. In this research we explore how more complex numerical integration methods, borrowed from neuroscience research, affect the training of networks based on current-based leaky integrate- and-fire neurons. We derive equations required for the integration process and suggest the use of spike time interpolation. Furthermore, we pro- vide insights into applying backpropagation on numerically integrated networks and highlight possible pitfalls of the process. We conclude that numerically integrated networks can achieve training accuracies close to their theoretical limits, with good convergence and training time characteristics. Specifically, high order integrations achieve robust and computationally viable training. Additionally, we explore the effects of spike time interpolation on network accuracy and use our findings to pro- vide insights into the role of different integration parameters on the effective training of spiking neural networks.
...
Spiking neural networks have gained traction as both a tool for neuroscience research and a new frontier in machine learning. A plethora of neuroscience literature exists exploring the realistic simulation of neurons, with complex models re- quiring the formulation and integration of ordinary differential equations. Overcoming this challenge has led to the exploration of various numerical integration techniques with the goal of highly stable and accurate simulations. In contrast, training spiking neural networks is often done with simple leaky integrate-and-fire models and rudimentary integration methods such as the forward-Euler method. In this research we explore how more complex numerical integration methods, borrowed from neuroscience research, affect the training of networks based on current-based leaky integrate- and-fire neurons. We derive equations required for the integration process and suggest the use of spike time interpolation. Furthermore, we pro- vide insights into applying backpropagation on numerically integrated networks and highlight possible pitfalls of the process. We conclude that numerically integrated networks can achieve training accuracies close to their theoretical limits, with good convergence and training time characteristics. Specifically, high order integrations achieve robust and computationally viable training. Additionally, we explore the effects of spike time interpolation on network accuracy and use our findings to pro- vide insights into the role of different integration parameters on the effective training of spiking neural networks.