This workshop is an attempt to bring together different perspectives on the functional role and organizational principles of the hippocampal formation. It is planned to have talks of 25 minutes each with 15 minutes discussion time, so that there is plenty of opportunity for an exchange of ideas and questions. All the following speakers have confirmed their participation:
slowness to place fields" |
Laurenz Wiskott (Institute for Theoretical Biology, Humboldt-University Berlin, Germany)
We present a model for the self-organized formation of hippocampal place cells based on unsupervised learning on natural visual stimuli. Our model consists of a hierarchy of Slow Feature Analysis (SFA) modules, which were recently shown to be a good model for the early visual system. The system extracts a distributed representation of position, which is transcoded into a place field representation by sparse coding (ICA). We introduce a mathematical framework for determining the solutions of SFA, which accurately predicts the distributed representation of computer simulations.
|09:45-10:10|| "May the hippocampal
formation break the curse of dimensionality?" |
András Lörincz (Department of Information Systems, Eötvös Loránd University, Budapest, Hungary)
The two-phase computational model of the entorhinal cortex - hippocampal (EC-HC) loop of Lorincz and Buzsaki is extended by a neural Kalman-filter model and by novel indepedendent process analysis neural architectures that perform noise filtering and can break combinatorial explosion together. Several falsifying predictions of the Lorincz and Buzsaki model have been reinforced experimentally after those predictions were made. The new model elaborates those predictions and offers a computational explanation for the crucial top-down supervisory-like role of this loop. Realistic robot simulations in a U-shaped and circular labyrinths are in progress. Prelimiary results concerning neural responses at the superficial and deep layers of the EC as well as at the CA1 subfield of the HC show good agreement with the experiments.
"Dynamics and Randomization Drive CA3 Recoding for Sequence Learning
and Prediction" |
William B. Levy (Laboratory of Systems Neurodynamics, University of Virginia, Virginia)
One abstract view of hippocampal function combines the ideas of sequence learning and random recoding. Using the psychological paradigm of trace conditioning, we distinguish four encoding modes and are beginning to elucidate the parametric sensitivities of this problem. Appropriate initialization values are needed but relatively easy to obtain. More critical are the interactions between synaptic modification parameters, average activity, connectivity, and the sequences to be encoded.
Levy, W. B, Hocking, A. B., & Wu, X. B. Interpreting hippocampal function as recoding and forecasting. Neural Networks 18, 2005, 1242-1264.
is the advantage of differentiating CA1 from CA3?" |
Gergely Papp (SISSA-Cognitive Neuroscience, Trieste, Italy)
The subfields of the mammalian hippocampus are strikingly different in their network architecture: after the laminated and extensively recurrent entorhinal cortex, information flows through 3 single layers of principal cells, with DG and CA1 essentially feed-forward, and between them CA3, again massively recurrent. The functional significance of such a salient structural differentiation has remained unclear, and neural activity in CA3 and CA1 shows qualitatively similar features, e.g. place fields in rodents. Many computational models succeed in mimicking hippocampal functions even when equipped with the architecture of CA3 alone. Based on the possibility that the functional advantage of the differentiation may be merely quantitative, we have developed a simulation approach that quantitatively compares the performance of a differentiated with a uniform network model, both comprised of the same number of units and connections (Treves, 2004).
Recent experiments by the Moser lab in Trondheim (and complementary independent results by the Knierim lab in Houston) have discovered a remarkable functional difference between CA3 and CA1 activity patterns: multiple environments with overlapping features are represented by distinct ensembles in CA3, whereas ensembles in CA1 show a correspondingly graded overlap (Leutgeb et al, 2004). The experiments suggest that the core memory operation is accomplished by CA3, in setting up a new, arbitrarily assigned representation to any distinguishable environment or context; whereas CA1, although able to access the multiple CA3 representations held concurrently in storage, tends to also reflect non-orthogonalized entorhinal inputs. Our models, building on Treves and Rolls (1992), indicate an important but quantitative contribution by DG to the orthogonalization of CA3 patterns, when needed, and detail the quantitative effect of the CA3-CA1 differentiation per se, decoupled from the predictable advantage of merely adding more computing resources to an existing CA3 network model. Results indicate that, in the differentiated CA3 network, continuous environments are represented by novel but fractured charts, which are then smoothed by CA1 recoding. Such “interpolation” may help the CA1 representation transmit a higher amount of decodable information, to be read out by cortical networks downstream.
|14:00-14:25|| "Radically different
computational properties of the two main cortical inputs to area CA1 in
single cells and networks" |
Szabolcs Káli (Institute of Experimental Medicine, Hungarian Academy of Sciences, Budapest, Hungary)
Hippocampal area CA1 receives afferent input from area CA3 via the Schaffer collaterals, but it also gets direct input from entorhinal cortex via the perforant path. These two projections innervate distinct domains of the dendritic tree of CA1 pyramidal cells, and may subserve different computational functions. The properties of the two inputs were first investigated in morphologically and biophysically detailed models of CA1 pyramidal neurons. Simulations revealed dramatic differences between the two pathways in terms of their efficacy, rules of summation, and interactions with intracellular events such as dendritic action potentials. The consequences of such differences were then explored in a model of the CA1 network which also included some of the major classes of inhibitory interneuron. We found that the two pathways contributed rather differently to spiking activity in the pyramidal cell population, and also played distinct roles in hippocampal (theta and gamma) oscillations.
|14:40-15:05|| "Electrical coupling between
principal cell axons: a source of very fast oscillations (>70 Hz),
gamma oscillations (30-70 Hz), and beta2 oscillations (20-30 Hz)"
Roger Traub (Department of Physiology and Pharmacology, Suny Downstate Medical Center, Brooklyn, New York)
A variety of experimental evidence supports the existence of gap junctions between principal cell axons, with ultrastructural confirmation being actively pursued. Both experimental and modeling evidence indicate that a network of electrically coupled axons can generate very fast oscillations, by a process resembling percolation on a random graph. Very fast oscillations in a pyramidal cell axonal plexus can in turn lead to gamma oscillations, in the presence of synaptic inhibition, an idea experimentally supported in the hippocampus, neocortex and entorhinal cortex. Finally, a newly discovered beta2 network oscillation occurs in parietal cortex in vitro, confined to large layer 5 pyramidal cells, that does not require chemical synapses, but does require gap junctions. The relevant electrical coupling in beta2 also appears to be axonal, although the morphological substrate remains to be identified.
"Uncertainty, phase and oscillatory hippocampal recall" |
Máté Lengyel (Gatsby Computational Neuroscience Unit, University College London, United Kingdom)
The activity of hippocampal neurons shows structured, dynamical, population behavior such as coordinated oscillations. It has long been observed that such oscillations provide a substrate for representing analog information in the firing phases of neurons relative to the underlying population rhythm. However, it has become increasingly clear that it is essential for neural populations to represent uncertainty about the information they capture, and the substantial recent work on neural codes for uncertainty has omitted any analysis of oscillatory systems. Here, we observe that, since neurons in an oscillatory network need not only fire once in each cycle (or even at all), uncertainty about the analog quantities each neuron represents by its firing phase might naturally be reported through the degree of concentration of the spikes that it fires. We apply this theory to memory in a model of oscillatory associative recall in hippocampal area CA3. Although it is not well treated in the literature, representing and manipulating uncertainty is fundamental to competent memory; our theory enables us to view CA3 as an effective uncertainty-aware, retrieval system.
Work in collaboration with Peter Dayan (Gatsby Computational Neuroscience Unit, University College London), Jeehyun Kwag and Ole Paulsen (Department of Physiology, Anatomy and Genetics, University of Oxford), and Francesco Battaglia (Swammerdam Institute for Life Sciences, University of Amsterdam).
Discussion with some introductary remarks by |
Peter Appleby (Institute for Theoretical Biology, Humboldt-University Berlin, Germany)
Some questions and ideas that came up during the discussion:
There is an established theory for the loop EC-DG-CA3-CA1-EC. What are the direct EC-CA3 and EC-CA1 connections good for?
What is the role of the many different types of inhibitory interneurons?
One should maybe look at the hippocampus more from the evolutionary perspective.
Modelers should work more on complete models of the hippocampal formation rather than on selected substructures. The system as a whole might face problems that we don't see with incomplete models. For instance, one might encounter problems of stability that would require inhibitory interneurons (s.a.).
Most models are on a rather abstract level. There should be some more work that relates these abstract models to detailed neural properties, such as STDP.
It would be good to know more about the statistics of the input into the hippocampus. Following the success(!?) of visual modeling as an adaptation to natural image statistics, a better understanding of the hippocampal input statistics might help us model the hippocampus.
What are the effects / what is the role of attentional control within the hippocampus?
The hippocampus is often modeled only by itself. There should be more considerations of the interaction between hippocampus and cortex. For example, does the hypothesis actually hold that the hippocampus is an intermediate storage site, or is that unrealistic given the small size of the hippocampus relative to the cortex (even if it stores only pointers) and the phenomenon of reconsolidation.