My Digital Art Sampling

No Comments


Self-Replicating Cursors


SurrealWaspTranscendent Entomologies  


Hodkin-Huxley and Pyramidal Neuron Image.pngHodgkin-Huxley Neuron


PoetryUpload.pngYear 1,000,000


JupiterBrain1 - FinalizedJupiter Brain


PigmanSubwayDon’t Spook the Snuffler


3D Cells GraphicEndodermal



Notes on: A Topological Representation of Branching Neuronal Morphologies, Kanari et al. 2017

No Comments


  • Neural morphologies involve complicated branching structures.
  • Traditionally, classification of such morphologies has involved qualitative assessment of microscopic images.
  • Topological data analysis (TDA) can classify geometric structures that are built from well-understood pieces such as spheres, cylinders, and tori. However, this sort of analysis computationally intensive when applied to neurons since their morphologies are so complex.
  • Feature extraction has been applied to quantitatively classify neural morphologies. This is a technique that allows a large, but potentially redundant, dataset to be converted into a reduced set of features that can be more easily analyzed.
  • Neurons are reconstructed as a set of points in connected by edges and then experts identify the relevant morphological characteristics to subject to the feature extraction algorithms. Unfortunately, the initial feature identification is quite subjective and can lead to different results. In addition, a great deal of information loss occurs in the processing of the data.
  • Neither TDA nor feature extraction are ideal for studying neural morphologies.

Topological Morphology Descriptor (TMD)

  • The TMD is an algorithm that maps the original images of dendritic trees to a topological representation with less information loss than other methods, though it does simplify the trees somewhat.
  • The TMD takes a partially ordered set of branch points (“parent” nodes with multiple further connections or “children”) and leaves (parent nodes with no further connections) as input.
  • Branch points and leaves are ordered by the “parent-child” relations.

The Persistence Barcode and Persistence Diagram

  • This output of the TMD is a “persistence” barcode and an equivalent persistence diagram.
  • In the persistence barcode below (left), the distance or “lifetime” along a component (in this case a component is an individual dendritic branch) is given on the horizontal axis. This distance is measured from the component’s “birth” or most distant point from the soma to its “death” or closest point to the soma.
  • In the equivalent example of a persistence diagram below (right), each component’s death distance is plotted against corresponding birth distances.
  • Note that many possible functions f can be used to generate a persistence barcode and diagram, but here, they are ordered by lifetime with units of microns.
  • In the figure below, the highlighted red component is used to exemplify what an individual component looks like when analyzed in this way.

Fig.1 Notes on Topology for Neural Morphologies Paper

  • The TMD method allows analysis of any treelike structure, unlike previous techniques which were not generalizable for the various types of neural morphology. TMD is much more biologically useful and computationally inexpensive.
  • TMD allows a reliability measure to be assigned to proposed groupings of neuronal trees. This helps develop a diversity profile that reflects the morphological variety of neurons.

Details of the TMD Algorithm

  • T denotes a rooted tree embedded in (a tree with a central node or root R).
  • N is the set of nodes in a tree T. Furthermore, N is defined as the union of branch points B and leaves L. Individual nodes are referred to in lowercase as n
  • Nodes with the same parent are called siblings.
  • Sets of nodes are denoted as A.
  • Subtrees are designated Tn and centered at node n. Each subtree possesses leaves Ln. Individual leaves are .
  • The function f could be defined by the radial distance, the branch length, the path distance, or other characteristics. In this case, it is defined as the radial distance from the soma (or root R).
  • The function v (below) is computed by the TMD algorithm. It finds a maximum value for f as a function of x where x belongs to the set of leaves.

TMD v of n function

  • Recall that (here) f is the radial distance from the soma. As such, v(n) will find the most distant leaves.
  • The algorithm starts with the leaves and follows each leaf’s path down towards the soma.
  • Leaves are associated with their respective nodes in the set of nodes A. When a node (with one or more siblings) is encountered, all but the “oldest” sibling is removed from that node. The oldest sibling is the sibling that extends farthest from the soma. After passing through all the nodes on a particular path, the algorithm reaches the soma and designates that path as a component. The component is added to the persistence barcode.
  • Below are some examples of what various paths along the dendritic tree might look like. Once again, this is achieved by a process of removing the younger siblings emerging from each node along a given path.

Fig.2 Notes on Topology for Neural Morphologies Paper

  • Once all the paths have been traversed, a set of components is outputted (recall that a component is an individual branch, not the entire path). The terminus (birth) and parent node (death) of each component is known. The lifetime of each component is the distance between its birth and death. This information allows a persistence barcode and an equivalent persistence diagram to be plotted.
  • To reiterate, the persistence barcode and diagram describe all the individual components in a given neural tree.
  • The TMD algorithm has linear complexity with respect to the number of nodes.
  • If a tree is rotated or translated, the TMD algorithm will still give the same result.

TMD Algorithm

Comparing Persistence Diagrams

  • There are standard topological distance measures between persistence diagrams, but this are not suited to dendritic trees. The authors defined an alternative distance measure called dbar in the space of persistence barcodes. However, there were still issues with this metric because it converted the barcode into a one-dimensional distribution and so lost some important information for classifying neural morphologies.
  • The better method for comparing persistence barcodes is to convert each barcode into a matrix of pixel values (an image) by using a discretization algorithm (in this case a sum of Gaussian Kernels) and then to employ machine learning for classifying the images.


Emotion as a Property of Information Part 2: Engineering Joy

No Comments

     If emotion has a physical basis, then it is possible to engineer emotions to stably maintain more optimal states. While such an endeavor might raise concerns regarding its apparent connection to recreational drug abuse, it is important to understand that recreational drugs represent an extremely crude form of emotional engineering. Using current street drugs to alter one’s mood is like trying to fly by jumping off a cliff while holding a pair of cardboard sheets. More rationally engineered pharmacological agents (i.e. antidepressants) are somewhat better in this respect, but still very primitive. To rationally design more positive emotional states, I propose that human and animal brains should be transferred to a more easily modifiable substrate and undergo cognitive rewiring consistent with joyful sentiments described by an information-structural theory of emotion.

     In this scenario, biological brains would first be gradually replaced by equivalent, engineered neural prosthetics. By maintaining continuous information transfer between neurons and machine components, the uploaded consciousness would not be a mere duplicate as in scan-copy mind uploading. Each neural prosthesis would contain some form of neuromorphic circuitry that simulates the region of the brain undergoing replacement. Such simulations would likely require multi-compartmental models of neurons along with fairly advanced learning rules and perhaps some models to approximate the effects of gene expression alterations due to non-synaptic chemical signaling. Note that a complete synaptic and dendritic connectome for the individual would also be necessary. This might be obtainable by MRI microscopy or similar technologies (this is a complicated topic, so I will cover it in depth elsewhere). The prosthetics would carry a dense array of nanoscale sensory and stimulatory equipment on their surfaces, allowing for detection of neurotransmitter release from presynaptic neurons and for stimulation of postsynaptic neural activity.

Engineering Emotion Figure1

     Once the brain is completely replaced, consciousness might be transferrable to another location by temporarily inactivating part of the computerized cerebrum and connecting the rest to a supercomputer that simulates the inactivated part by an electromagnetic (or similar) conduit. After this connection is established, the remaining part of the brain still housed in its host body could be inactivated and then the supercomputer could turn on a simulation of that same part, completing the brain inside the supercomputer without breaking the continuity of the individual’s conscious experience. Since uploaded minds would exist in a programmable environment, they would be much more amenable to emotional engineering than biological brains.

     In order to emotionally enhance uploaded consciousness, a quantitative understanding of emotion will be necessary. Fortunately, many of the technical challenges associated with mind uploading are similar to the challenges associated with a quantitative theory of emotion. By studying the neural correlates of emotion and relating them to experiences, we might make strides in developing this theory of emotion. I suggest that mapping neural circuitry at the resolution of individual dendritic spines will be sufficient to understand the physical correlates of human emotion. This assumes biologically realistic models of individual neurons, but it is arguable that such models already exist in the form of multicompartmental conductance-based models equipped with learning rules. I would propose that relatively minor improvements are required to model single cells. The more challenging area is likely understanding neural circuitry throughout the human brain as a system. Currently, we can only speculate about universal informational principles that might underlie emotions. For instance (mentioned in part 1), many positive emotions might involve stable positive feedback loops. But by studying the brain, we could achieve enough understanding of emotions to replicate and augment them in a non-biological substrate.

     Emotional engineering might yield spectacular results in the long term. Post-singularity, we might convert entire worlds into computronium, allowing for a simulated paradise to be designed. Organisms living in this simulation could choose to live with minimal capacity to suffer and tremendous capacities for joy, love, and wonder. Though some might claim that this would be a “false bliss,” I argue that existentialist philosophy counters such claims. In existentialism, emotions are real and meaningful by virtue of being felt. If an emotion is truly positive, fulfilling, and lasting, then there is no reason to condemn it simply because it has been engineered. We might explore the universe as well, accumulating more matter to transform into computational joy-substrate. I would suggest that, the more processing power we acquire, the more we could intensify our positive emotions. Eventually, the entirety of the cosmos might be optimized for joy. This would not be a hedonic device, its happiness would possess a spiritual component. The universal brain would act as a technological nirvana, a resplendent celebration, a sonata machine.

Notes on Neocortical Interneurons

No Comments

Parvalbumin-expressing (PV) cells

  • Fast-spiking, GABAergic
  • Two types of PV cells classified by axonal contacts to target pyramidal cells.
  • (1) Axo-somatic or axo-dendritic (proximal dendrites)
  • (2) Axo-axonic: strongly “edits” firing in an inhibitory manner
  • In layer II/III of the primary somatosensory cortex, PV cells have high connectivity both to and from excitatory pyramidal neurons. This facilitates a strong inhibitory loop, damping the region most of the time

Somatostatin-expressing (Sst) cells

  • GABAergic
  • Sst cells target distal excitatory pyramidal cell dendrites, often in layer I
  • Sst cells receive excitatory inputs from nearby pyramidal cells. The unique molecular properties of these synapses cause the Sst cells to hyperpolarize more slowly between inputs, enabling integration of bursts of firing from the pyramidal cells. This is not the case for other types of cortical cells

Sst Cell Spikes Diagram

Vasoactive intestinal peptide-expressing (VIP) cells

  • GABAergic
  • Inhibits other inhibitory neurons. They more frequently inhibit Sst cells than PV cells

Neurogliaform cells

  • GABAergic
  • Neurogliaform cells do not have post-synaptic partners. Instead, they release GABA into the extracellular space, thus increasing IPSPs in their general area