Limits...
Modeling place field activity with hierarchical slow feature analysis.

Schönfeld F, Wiskott L - Front Comput Neurosci (2015)

Bottom Line: Simulations are based on a hierarchical Slow Feature Analysis (SFA) network topped by a principal component analysis (ICA) output layer.The slowness principle is shown to account for the main findings of the presented experimental studies.Future iterations of the model will thus have to incorporate additional information, such as path integration and grid cell activity, in order to be able to also replicate studies that take place during darkness.

View Article: PubMed Central - PubMed

Affiliation: Theory of Neural Systems Group, Institut für Neuroinformatik, Ruhr Universität Bochum Bochum, Germany.

ABSTRACT
What are the computational laws of hippocampal activity? In this paper we argue for the slowness principle as a fundamental processing paradigm behind hippocampal place cell firing. We present six different studies from the experimental literature, performed with real-life rats, that we replicated in computer simulations. Each of the chosen studies allows rodents to develop stable place fields and then examines a distinct property of the established spatial encoding: adaptation to cue relocation and removal; directional dependent firing in the linear track and open field; and morphing and scaling the environment itself. Simulations are based on a hierarchical Slow Feature Analysis (SFA) network topped by a principal component analysis (ICA) output layer. The slowness principle is shown to account for the main findings of the presented experimental studies. The SFA network generates its responses using raw visual input only, which adds to its biological plausibility but requires experiments performed in light conditions. Future iterations of the model will thus have to incorporate additional information, such as path integration and grid cell activity, in order to be able to also replicate studies that take place during darkness.

No MeSH data available.


Related in: MedlinePlus

Overview of our hierarchical Slow Feature Analysis (SFA) network. SFA nodes in the lowest layer directly work on our software-generated image frames. Each node scans a predetermined receptive field that overlaps 50% with directly adjacent ones. The second layer works on the pre-processed output of the first layer. The third layer consists of a single SFA node integrating the output of all layer-2 SFA nodes. The last layer is one node performing independent component analysis (ICA) to perform sparse coding on the final SFA output.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4441153&req=5

Figure 1: Overview of our hierarchical Slow Feature Analysis (SFA) network. SFA nodes in the lowest layer directly work on our software-generated image frames. Each node scans a predetermined receptive field that overlaps 50% with directly adjacent ones. The second layer works on the pre-processed output of the first layer. The third layer consists of a single SFA node integrating the output of all layer-2 SFA nodes. The last layer is one node performing independent component analysis (ICA) to perform sparse coding on the final SFA output.

Mentions: Figure 1 depicts the structure of the SFA network used in our simulations. The initial three layers are formed from three individual SFA nodes that are each cloned into multiple copies to form their respective layer. The cloned nodes will each perform SFA on their respective input data but share the initial training data. This can be done since we do not expect significant differences in the input statistics at any specific area of the visual field, and further reduces training time of the network. Of the three SFA layers, the initial layer consists of a two dimensional array of 63 by 9 clone nodes featuring a 10 by 8 pixel wide receptive field and 32 output channels each. It works directly on the raw image frames and extracts the most slowly varying signals to relay them to the next layer. Berkes and Wiskott (2005) have shown that the features learned by such a layer resemble those measured in cells of the primary visual cortex. The second layer is defined as an 8 by 2 array of nodes (14 by 6 input channels each) and processes the abstract features provided by the previous layer. The third SFA layer consists of a single SFA instance that integrates all of the features detected over the whole input range by the previous two layers. It can be mapped to the entorhinal cortex and yields the most abstract features extracted from the input data stream. The output signals of this node, however, do not yet resemble clearly localized place field firing patterns. The three initial SFA layers are thus followed by a fourth layer, made from a single MDP node performing sparse coding via independent component analysis (ICA) (Hyvärinen, 1999). This additional step is linked to the dentate gyrus, the hippocampal area one synapse downstream of the entorhinal cortex, where experimental studies have reported sparse firing patterns (Jung and McNaughton, 1993; Chawla et al., 2005).


Modeling place field activity with hierarchical slow feature analysis.

Schönfeld F, Wiskott L - Front Comput Neurosci (2015)

Overview of our hierarchical Slow Feature Analysis (SFA) network. SFA nodes in the lowest layer directly work on our software-generated image frames. Each node scans a predetermined receptive field that overlaps 50% with directly adjacent ones. The second layer works on the pre-processed output of the first layer. The third layer consists of a single SFA node integrating the output of all layer-2 SFA nodes. The last layer is one node performing independent component analysis (ICA) to perform sparse coding on the final SFA output.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4441153&req=5

Figure 1: Overview of our hierarchical Slow Feature Analysis (SFA) network. SFA nodes in the lowest layer directly work on our software-generated image frames. Each node scans a predetermined receptive field that overlaps 50% with directly adjacent ones. The second layer works on the pre-processed output of the first layer. The third layer consists of a single SFA node integrating the output of all layer-2 SFA nodes. The last layer is one node performing independent component analysis (ICA) to perform sparse coding on the final SFA output.
Mentions: Figure 1 depicts the structure of the SFA network used in our simulations. The initial three layers are formed from three individual SFA nodes that are each cloned into multiple copies to form their respective layer. The cloned nodes will each perform SFA on their respective input data but share the initial training data. This can be done since we do not expect significant differences in the input statistics at any specific area of the visual field, and further reduces training time of the network. Of the three SFA layers, the initial layer consists of a two dimensional array of 63 by 9 clone nodes featuring a 10 by 8 pixel wide receptive field and 32 output channels each. It works directly on the raw image frames and extracts the most slowly varying signals to relay them to the next layer. Berkes and Wiskott (2005) have shown that the features learned by such a layer resemble those measured in cells of the primary visual cortex. The second layer is defined as an 8 by 2 array of nodes (14 by 6 input channels each) and processes the abstract features provided by the previous layer. The third SFA layer consists of a single SFA instance that integrates all of the features detected over the whole input range by the previous two layers. It can be mapped to the entorhinal cortex and yields the most abstract features extracted from the input data stream. The output signals of this node, however, do not yet resemble clearly localized place field firing patterns. The three initial SFA layers are thus followed by a fourth layer, made from a single MDP node performing sparse coding via independent component analysis (ICA) (Hyvärinen, 1999). This additional step is linked to the dentate gyrus, the hippocampal area one synapse downstream of the entorhinal cortex, where experimental studies have reported sparse firing patterns (Jung and McNaughton, 1993; Chawla et al., 2005).

Bottom Line: Simulations are based on a hierarchical Slow Feature Analysis (SFA) network topped by a principal component analysis (ICA) output layer.The slowness principle is shown to account for the main findings of the presented experimental studies.Future iterations of the model will thus have to incorporate additional information, such as path integration and grid cell activity, in order to be able to also replicate studies that take place during darkness.

View Article: PubMed Central - PubMed

Affiliation: Theory of Neural Systems Group, Institut für Neuroinformatik, Ruhr Universität Bochum Bochum, Germany.

ABSTRACT
What are the computational laws of hippocampal activity? In this paper we argue for the slowness principle as a fundamental processing paradigm behind hippocampal place cell firing. We present six different studies from the experimental literature, performed with real-life rats, that we replicated in computer simulations. Each of the chosen studies allows rodents to develop stable place fields and then examines a distinct property of the established spatial encoding: adaptation to cue relocation and removal; directional dependent firing in the linear track and open field; and morphing and scaling the environment itself. Simulations are based on a hierarchical Slow Feature Analysis (SFA) network topped by a principal component analysis (ICA) output layer. The slowness principle is shown to account for the main findings of the presented experimental studies. The SFA network generates its responses using raw visual input only, which adds to its biological plausibility but requires experiments performed in light conditions. Future iterations of the model will thus have to incorporate additional information, such as path integration and grid cell activity, in order to be able to also replicate studies that take place during darkness.

No MeSH data available.


Related in: MedlinePlus