Extraction and Characterization of Essential Discharge Patterns from Multisite Recordings of Spiking Ongoing Activity Authors: Riccardo Storchi, Gabriele E. M. Biella, Diego Liberati, Giuseppe Baselli Published: 2009, PLOS ONE DOI: 10.1371/journal.pone.0004299 ABSTRACT Neural activation patterns proceed often by schemes or motifs distributed across the involved cortical networks. As neurons are correlated, the estimate of all possible dependencies quickly goes out of control. The complex nesting of different oscillation frequencies and their high non-stationariety further hamper any quantitative evaluation of spiking network activities. The problem is exacerbated by the intrinsic variability of neural patterns. Our technique introduces two important novelties and enables to insulate essential patterns on larger sets of spiking neurons and brain activity regimes. First, the sampling procedure over N units is based on a fixed spike number k in order to detect N-dimensional arrays (k-sequences), whose sum over all dimension is k. Then k-sequences variability is greatly reduced by a hierarchical separative clustering, that assigns large amounts of distinct k-sequences to few classes. Iterative separations are stopped when the dimension of each cluster comes to be smaller than a certain threshold. As threshold tuning critically impacts on the number of classes extracted, we developed an effective cost criterion to select the shortest possible description of our dataset. Finally we described three indexes (C,S,R) to evaluate the average pattern complexity, the structure of essential classes and their stability in time. We validated this algorithm with four kinds of surrogated activity, ranging from random to very regular patterned. Then we characterized a selection of ongoing activity recordings. By the S index we identified unstable, moderatly and strongly stable patterns while by the C and the R indices we evidenced their non-random structure. Our algorithm seems able to extract interesting and non-trivial spatial dynamics from multisource neuronal recordings of ongoing and potentially stimulated activity. Combined with time-frequency analysis of LFPs could provide a powerful multiscale approach linking population oscillations with multisite discharge patterns. INTRODUCTION In the last twenty years, studies on information encoding in the nervous system have provided fundamental insights into the nature of neural inner dynamics and of sensorimotor representation and coding of the external world. Powerful and flexible statistical techniques have grown in time [1] – [5] , improving the analyses of stimulus-response experimental paradigms along with their corollary complexities. Scant attention has been paid to other dynamic features such as spontaneous or ongoing activity. Nevertheless, this feature represents 90% percent of the whole metabolic exertion of the brain [6] . An exhaustive description of ongoing activity as a kind of substrate intermingling with signals generated by external sources, could provide fundamental insights into nervous system dynamics. Spontaneous neuronal population activities from many sites in the central nervous system present complex combinations of different oscillation frequencies [7] – [9] . Regular and repetitive motifs nest within the frames of these global rhythms. Furthermore, stereotyped patterns of specific neuron subsets have been evidenced in many different experimental conditions such as at the onset of UP states [10] or in in vitro recordings with calcium imaging [11] , [12] . Because of the surging number of items, the statistical evaluation of all possible discharge patterns in multisite recordings of even a few channels quickly goes out of control. Nevertheless, pioneering studies and further advancements have developed algorithms capable of recognizing small, repetitive and well-timed patterns in tonic activity regimes [13] – [17] . Still, timing is likely not the only encoding means valid in the CNS. Repetitive spontaneous and evoked neuronal activation patterns evidenced time warping features, namely duration independent activity episodes, in most diverse behavioral conditions [18] – [20] . These features may represent an additional piece of the multifarious presence of noise, a pervasive trait throughout the bottom-up (or top-down) scaling of brain analyses [21] , that affects the immediate detection of accessible regular and repetitive patterns. In any case the crucial role of noise in brain operations, as it has become evident in recent years, has gone largely unknown and is a current subject of wide debate. In the context of a lack of comfortable theoretical background for the present scenario, we tried to design an efficient algorithm to detect the presence of regular motifs in all ongoing activity regimes and for most variable numbers of spikes or recording traces. We then developed and evaluated a set of measures to characterize such motifs in terms of spatial structure, complexity and temporal evolution. Our approach is conceptually different from previous works as we introduced several important novel features. First, in addition to constant time bin sampling, we introduced a conditional sampling based on keeping the number of spikes constant in each sample. This procedure allows the rejection of common mode frequency modulations. Second, we applied a clustering procedure in order to extract a reasonable number of classes from the wide variety of distinct patterns and to reduce the effect of noise. Then, following a cost criterion, we extracted a selection of the most frequently occurring classes that we called essential classes (EC). Finally, we developed three indexes, labeled C , S and R , evaluating respectively the average pattern complexity (C), the structure of the ECs ( S ) and their stability in time ( R ). We applied such indexes both to simulated and to real data. We mainly focused on the results obtained by using conditional sampling. However the results obtained with constant time bin sampling will be reported as well and commented for comparison. In the Methods we describe the whole algorithm flow obtained by cascading constant time bin sampling or conditional sampling, the clustering procedure and the extraction of ECs (respectively subsections 1,2 and 3). We then describe the indexes C , R and S (subsection 4) and provide a detailed explanatory example of the whole procedure (subsection 5), in order to clarify the most critical steps of the algorithm. In the Results we evaluate the cost criterion performances (subsection 1) and apply the indexes C , R and S to several kinds of simulated activities (subsection 2) and to real data (subsection 3) In subsections 2 and 3 we also compare the results obtained on simulations and on real data and the information provided by constant time bin and conditional sampling. In the Discussion we briefly discuss advantages and drawbacks of the techniques developed. MATERIALS AND METHODS Given k spikes and N sources, we define a k -sequence as an N -dimensional vector whose values, on each dimension, are the number of spikes counted on each source, the sum over all dimensions being k . By fixing k and N it is possible to extract a finite set of k -sequences from each multisite recording ( Fig. 1A ). A) Conditional sampling. This procedure may provide adaptive time windows and detect as equal time-warped patterns with the same spatial distributions over the spiking sources. In this case note that by setting k = 8 three equal k -sequences are detected, as evidenced by the red ellipse). B) Constant time bin sampling applied to the same pattern. This is the most classical sampling procedure and is draw for comparison. Consider a set of N spiking sources and let x 1 , …, x N be the number of spikes emitted by them in a variable time period. Then setting a constant sum Σ n x n = k and collecting sets of k -sequences, we sample from the conditional probability distribution K -sequences are sampled along the multichannel recording raster-plot on adjacent non-overlapping windows of variable duration ( Fig. 1A ). Conditional sampling does not rely on an “external” clock-the time of recording, but simply on the spike occurrences across the multiple emitting sources. This peculiar property makes this procedure, to a great extent, insensitive to the common mode frequency modulations that affect all the sampled spiking sources in the same way. A simple example of this property is provided in Document S1 . Unlike the classical constant time-bin sampling (reported in Fig. 1B for comparison), conditional sampling is robust relative to the time-warp, enabling us to consider patterns of different duration and equal spatial distribution just like a same pattern (see red circled samples in Fig. 1A ). In analogy with k -sequences, we also define t -sequences as samples collected by using the classical constant time bin sampling procedure. In this context t represents the number of time-units Δt , being this term the largest time interval in which a source can emit at most one spike (typically Δt = 1 ms). More generally, when we will refer to both k -sequences and t -sequences, we will simply write sequences. THE CLUSTERING PROCEDURE The clustering procedure follows a bisecting divisive partitioning algorithm. At each iteration, the cluster bisection is obtained by applying in cascade the Principal Direction Divisive Partitioning (PDDP, see [22] ) and the Bisecting K-means (BK) algorithms. Iterative separations are stopped when the dimension of each cluster is smaller than a predefined threshold D max . The cluster dimension D j of a j th cluster, is computed as and X i is the set of values of the i th dimension associated with the sequences belonging to the j th cluster. For each cluster j , D i can be interpreted as the number of all its possible distinct sequences and D max as the maximum number for any cluster. At the first step the whole dataset is bisected. Then, before any successive bisection, clusters are ranked by their dimension D and bisection is performed on the largest cluster. This criterion allows the obtaining of clusters which are roughly comparable in their cluster dimension. Once we have chosen the cluster to bisect, we apply PDDP with the following steps. First, we subtract to each i th dimension (of the chosen cluster) its mean value. Then, we compute the covariance matrix of the k-sequences belonging to the chosen cluster and we extract the eigenvector v associated with the largest eigenvalue. Finally, given two subclusters G 1 and G 2 , we assign each sequence x to one of them by evaluating the score A = x T v . When A is larger than 0, we assign x to G 1 . Otherwise we assign x to G 2 . After preliminary bisection by PDDP we compute the centroids w 1 and w 2 on, respectively, G 1 and G 2 , and use them to initialize the BK algorithm. The BK algorithm is iterative and is composed of two steps. First, each item is assigned to the nearest centroid. Then, the centroids are recalculated on the base of the last assignment stage. K -means always converge so that after a number of iterations centroids no longer change their positions. For a more detailed treatment of PDDP and BK refer to [22] – [24] . As shown by Savaresi and Boley [23] , the PDDP algorithm provides a wise centroid initialization while the BK algorithm refines the partitioning. The distinct clusters obtained by the clustering procedure constitute the sequence classes. Decreasing the values of D max we obtain a more detailed description of the dataset, with a larger number of lower dimension sequence classes. However, too small D max values may lead to overfitted descriptions failing to capture the salient properties of the noisy discharge patterns. A solution to this problem is proposed in the following paragraph. D MAX SELECTION AND ESSENTIAL PATTERN EXTRACTION In the present subsection and in the next we propose a general evaluation method applicable to both k -sequences and t -sequences. Thus, although we will always refer to k and k -sequences, the same analysis could be applied to t -sequences simply by k to t label substitution. To choose the best D max value we associate each D max with a cost. The cost is evaluated taking into account the encoding length (in bits) needed to represent a selected model and, using a model-dependent encoding, the data. Being c min (k,D max ) the smallest cost associated with a selected D max value, we define as the smallest cost, given k , of a k -sequence dataset. The encoding algorithm relies on a re-elaboration of a scheme proposed by Willems [25] and is described in detail in Document S1 . This scheme has a simple and direct interpretation in terms of information theory, being straight on implementable as a true compression algorithm. In brief, the k -sequence classes are ordered in a database by the number of their occurrences. The database represents the model. Among the possible subsets constituted by the first M classes, the one accomplishing the shortest encoding length is selected. To exploit the model knowledge about an i th class we need first 2Nlog(k+1) bits to store the class in the model and then log(i)+log(D max ) bits at each occurrence of a k -sequence belonging to that class (for details see Document S1 ). Because log(i) is an increasing function of the class position in the database, classes occurring at high frequency have a smaller cost than rare ones. Not all classes are necessarily stored in the model. Some of them may not occur enough times to be conveniently stored. The classes contained in the model constitute the essential patterns we are looking for to describe the most regular part of our finite dataset. The total cost of encoding the data without the database is and n tot (k) is the overall number of k -sequences. Nlog(k+1) represents an upper limit for the number of possible k -sequences not accounting for the fact that the spike sum over all dimensions is k . This is done for the sake of generality. In fact, in the present form, the cost criterion can be used with any kind of sampling procedure (like constant time bin sampling). This approach is equivalent to the use of a common encoder without any a priori knowledge on the sampling procedure used. To select the best model we start to test the first class by computing and, iteratively, test an increasing class number. At each iteration i , we evaluate the differential cost Δc i,i−1 (k,D max ) of adding a new i th class. Such cost is divided into three components and n(i) indicates the number of the i th class occurrences in the dataset. Δc 1,i,i−1 (k,D max ) represents the differential database cost, that is the number of bits we need to store a class in the database. Δc 2,i,i−1 (k,D max ) represents the gain, achievable by encoding n(i) k -sequences once stored the i th class in the database. Δc 3,i,i−1 (k,D max ) represents the differential flag cost (see Document S1 ). The database cost constitutes a regret term for too short datasets and fades away for long sequences, while the flag cost increases linearly with the length of the dataset thus constituting a constant regret term for the number of classes. For increasing values of i , the i th class will be accepted if the differential cost Δc i,i−1 (k,D max ) is negative. Iterations are stopped when this is not verified. The i th class is rejected and, setting M = i−1 , c min (k,D max ) = c M (k,D max ) . The asymptotic upper bound for the average cost of a k -sequence is (for details see Document S1 ) where p is the probability distribution of the class occurrence (all classes, not only the essentials), H(p) its entropy and log(log((k+1) N +1)+1) an upper limit for the flag cost log(log(M+1)+1) . CHARACTERIZATION OF ESSENTIAL PATTERNS The application of the cost criterion described in the previous paragraph enables us to extract, for any k , distinct classes of essential patterns. Such patterns are associated with the D max value for which c min (k,D max ) = c min (k) . To characterize the ECs we use three measures: C(k) , S(k) and R(k) . C(k) stands for complexity, is a general descriptor of a k -sequence dataset and is given by the ratio S(k) stands for segregation and is computed as follows being x i (z) the i th dimension of the z th k -sequence in the dataset and, respectively, E k and n ess (k) the set of indexes and the number of k -sequences belonging to ECs. S(k) enables us to discriminate k -sequences where spikes are significantly segregated in some subset of sources, from k -sequences where spikes are homogeneously distributed among them. Finally, R(k) (reiteracy) concerns the way successive k -sequences shift among different classes. By the evaluation of R(k) we can detect the presence of stable patterns. Namely, given a symbol string associated with a multichannel recording, we estimate the probability P R (k) that successive k -sequences belong to the same EC where z is the z th symbol of a string associated with the z th k -sequence of the dataset and A k is the set of symbols representing all the distinct ECs (only the essential, not all classes). If P R (k) is significantly higher than expected by chance, reiteracy is detected and we set R(k) = 1 , otherwise R(k) = 0 . An approximated P R0 (k) distribution under the null hypothesis is obtained by N S shuffles on the original symbol sequence. We call N < the number of estimates for which P R0 (k)
(1−N < /N S ) .
The indexes C , R and S may be meaningfully applied to both k -sequences and t -sequences. In case of k -sequences, C(k) represents the spatial complexity connected to the number and the diversity of all the possible configurations collected from a multichannel recording regardless of their duration. R(k) , the shift indicator between two different k -sequence classes, determines the switching properties of the system. Positive reiteracy on k -sequences may be due to small reverberant circuits, to intrinsic single neuron spiking dynamics or, more generally, to any stable differential activity that emerges over common mode modulations of the considered channels. Finally, S(k) represents the average euclidean distance from the homogeneous sequence, i.e. that sequence where all spikes are equally distributed among the sources.
The significance of C , R and S , when dealing with t-sequences (respectively C(t) , R(t) and S(t) ) is less unambiguous. In fact, while C(k) , R(k) and S(k) values simply evaluate spatial regularities, C(t) , R(t) and S(t) values mix spatial and temporal information. Thus, a t -sequence may have a low C(t) because of either stereotyped spatial configurations or regular frequency modulations (or both). The same remark holds true for R(t) and S(t) . Although positive reiteracy may be due to stable spatial patterns and significant segregation to a sparse source recruitment, both R(t) and S(t) may be substantially affected by frequency modulations slower than the sampling frequency.
AN EXPLANATORY EXAMPLE
To gain a better understanding of the whole procedure let's consider the example data shown in Fig. 1A . After conditional sampling, the associated k -sequence set, given k = 5, is displayed in Fig. 1A (bottom) . As this dataset is a bit too short for our analysis we build a larger one simply by repeating six times each item. Then, we add some noise deleting one spike from each k -sequence and reassigning it randomly to one of the 3 channels of the same k -sequence. The resulting set is
2 2 1 1 1 2 4 2 3 3 3 0 0 1
0 1 3 4 2 2 1 0 0 0 1 3 3 1
3 2 1 0 2 1 0 3 2 2 1 2 2 3
2 4 1 3 2 3 0 0 0 1 3 1 2 2
2 1 2 0 0 0 3 4 2 3 2 2 1 0
1 0 2 2 3 2 2 1 3 1 0 2 2 3
2 1 1 1 2 3 3 3 1 3 0 0 0 3
1 3 3 1 2 2 0 1 1 1 3 4 3 1
2 1 1 3 1 0 2 1 3 1 2 1 2 1
3 2 2 2 2 0 1 1 2 4 2 3
2 2 1 0 1 4 4 2 1 0 2 1
0 1 2 3 2 1 0 2 2 1 1 1
We apply the clustering algorithm with three different values of D max , respectively 4,12,48,75.
By associating a symbol with each distinct cluster we get the following strings:
abcdefgahhillafgehahldecmebabccafmhiaildlimfbabddebgfi
aabbccdaaadbbacdcaaabbcbdcaaabbacdadadbbbddcaaabbcadcd
aabbbbaaaaabbababaaabbbbabaaabbabaaaaabbbaabaaabbbaaba
aabbbbaaaaabbababaaabbbbabaaabbabaaaaabbbaabaaabbbaaba
Because D max value is critical, we select the best description by using the evaluation criterion described in paragraph 3 and in Document S1 .
The description associated with D max = 12 is selected as the most synthetic, by our criterion, being the complexity C(5) = 0.95.
The dimension of the clusters a , b , c , d are respectively 12,12,9,12. The cost c 0 (5) is 418.76 bits, while c min (5) = 398.84 bits, being the differential database cost Δc 1,3,0 (5,12) = 2MNlog(k+1) = 2*3*3*log(6) = 46.53 bits, the differential gain Δc 2,3,0 = Σ i (Nlog(k+1)−log(i)−log(D max ))n(i) = −152.34 bits and the differential flag cost Δc 3,3,0( 5,12) = t(k)log(log(M+1)+1) = 85.59 bits.
K -sequence classes a , b , c are defined by our criterion as essential, while d does not occur enough times. Writing ‘ - ’ for non-ECs the final string is
aabb--daaadbba-d-aaabb-bd-aaabba-dadadbbbdd-aaabb-ad-d
On this string we finally compute, on ECs, reiteracy and segregation ( S = 2.24). To evaluate R(5) , we calculate P R = 0.31 on the original final string, then we shuffle the final string N S = 10000 times in order to obtain P R0 ( Fig. S1 ). Having set α = 0.05, and being (1−N < /N S ) = 0.046 we conclude that reiteracy is significant ( R = 1) at level 0.05 (but not at level 0.01).
We can repeat the same procedure with constant time bin sampling. To obtain a reasonable dataset we replicate 6 times the data in Fig. 1B and add noise as in the previous case. The resulting set is
2 2 1 0 2 2 0 4 3 2 3 0 0 1
0 2 1 5 2 1 1 1 0 1 2 3 6 1
3 2 1 3 1 2 0 2 2 2 1 0 2 3
2 0 4 4 1 3 1 0 1 1 0 4 4 2
1 0 0 0 1 1 1 5 3 3 1 0 0 1
2 1 3 1 3 2 1 3 1 1 0 3 1 2
2 0 0 1 2 0 4 3 2 2 0 1 1 2
2 1 5 1 3 1 0 1 1 2 3 6 3 3
2 2 3 3 0 0 3 1 2 2 0 1 1 0
1 3 4 1 3 0 0 0 2 0 4 2
0 2 0 0 0 3 5 2 3 1 1 0
0 2 1 4 3 0 3 3 0 0 2 3
Then we apply the whole algorithm again. With D max = 4,12,48,75 we obtain the following strings:
abcdbbefgbhjkiblgmifcdjjegmbbndioeghbbjkjolhmagjdnoefa
abcdbbefgbfhdabcgfafcdhhegfbbedahegfbbhdhhcffaghdehefa
aaabaacddadcbaaaddadabcccddaacbaccddaacbccaddadcbcccda
aaabaabaaaabbaaaaaaaabbbbaaaabbabbaaaabbbbaaaaabbbbbaa
Because of the larger number of possible t -sequences in respect to k -sequences, we expect the optimal D max value to increase. In fact the final string is associated with D max = 75
aaa-aa-aaaa--aaaaaaaa----aaaa--a--aaaa----aaaaa-----aa
being the dimension of the class a equal to 75 and the values of C,S equal to respectively 0.9966, 1.6016. To use the cost criterion described, we set k at the largest value collected on a single channel during constant time bin sampling ( k = 6).
Being (1−N < /N S ) = 0.0406 we conclude that reiteracy is significant at level α = 0.05.
COST CRITERION PERFORMANCES
In this paragraph we analyze the relation between Entropy ( H ) and our cost criterion evaluation. We test the cost criterion on two different sources. We then compute c min (k max ) as a function of the message length n tot . A source is composed of a regular and a random part. The regular part is constituted by few sequences that are repeated a number of times. The random part is constituted by random sequences generated by N independent extractions from a homogeneous distribution: namely, all integers ranging from 0 to k max have probability 1/(k max +1) of occurrence.
Each message is composed of r n tot and (1−r) n tot sequences belonging, respectively, to the regular and to the random part.
We define n r the number of distinct sequences of the regular part. Each of them occurs rn tot /n r times in a message.
If N and k max are not too small For an infinitely long message of n tot sequences, H represents the smallest possible average description length of a sequence. The asymptotic evaluation of our cost criterion, c min ( k max ), in Document S1 , provides a significantly less parsimonious description. This holds true also for finite length messages, as can be observed in Fig. 2A,B,G,H , where the average values of H/(Nlog(k max +1)C(k)) are plotted as a function of n tot . The difference between entropy and c min ( k max ) descreases as n tot increases. Conversely, it strongly increases as a function of r .
Upper row (plots A,B,C,D,E,F) related to the first source. k max = 10, N = 5. A) The function H/(C(k max )Nlog(k max +1)) is estimated, as a function of n tot , at different r values ranging from 0 to 1. B) The same as in A). Here the cost criterion is modified to start evaluation from Δc 2,1 instead of Δc 1,0 . C,D) M values estimated with the original (C) and the modified (D) criterion. E,F) D max estimated with the original (C) and the modified (D) criterion. Lower row (plots G,H,I,J,K,L) related to the second source ( k max = 10, N = 5).
The more regular the message, the worse c min ( k max ) approximates H . The differential database cost Δc 1,M,0 = 2Nlog(k max +1) , accounting for the lowest H/(Nlog(k+1)C(k)) values at low n tot , fades away for increasing values of this parameter (see Methods section). The differential flag cost Δc 3,M,0 is independent of message length and accounts for the non-optimality at large n tot values. Δc 3,M,0 plays a key role in the choice of the EC number, typically leading to a conservative selection that prevents from overfitting. The contribution of Δc 3,M,0 could be reduced by modifying and expanding the flag and the database structures, in order to use a single flag to encode several sequences. This possibility, under current experimental check, can effectively reduce Δc 3,M,0 and the number of flags, and will be presented in a forthcoming paper. The differential flag cost Δc 3,M,0 is sometimes too large and may impair the detection of regular patterns that are significantly present in the message. To avoid the problem, the cost criterion can be evaluated starting from M = 1 instead than from M = 0. Results obtained with this modification are drawn in Fig. 2B,D,K,H,J,L while those relative to the original algorithm are presented in Fig. 2A,C,E,G,I,K . The values of M and D max are reported, respectively, in Fig. 2C,D,I,J and in Fig. 2E,F,K,L .
The regular part of the first source ( Fig. 2A–F ) is composed of the following sequences
1 10 5 1
3 5 5 2
10 0 5 3
10 2 5 10
2 9 6 10
having set k max = 10. For short r or n tot values no ECs are detected from our algorithm (white region in Fig. 2E,F ). Instead, increasing these parameters, the algorithm always extracts four distinct ECs with D max = 1(blue region in Fig. 2E,F ). The sharp transition between 0 and 4 ECs can be expected because the regular sequences we introduced are really different from each other.
The following sequences
1 0 1 0 2 0 0 2
3 4 4 4 3 3 4 4
10 10 9 10 10 10 10 10
10 9 10 10 10 10 10 9
2 3 2 2 1 3 2 1
are used to generate the regular part of the second source ( Fig. 2G–L ). No ECs are detected for low r or n tot values (white region in Fig. 2K,L ). With slight increments of these parameters, a single EC is detected containing 7 or 8 distinct sequences of the regular part (violet region in Fig. 2I,J and red region in Fig. 2K,L ). When the EC contains all 8 sequences D max = 72. In comparison with those of the first source, these sequences are very similar to each other. Accordingly, when the number of their occurrence is not too high, the algorithm detects them as different noisy versions of a unique essential pattern. For further increments of n tot or r , our algorithm returns, as ECs, 7 of the 8 sequences composing the regular part of the message (red region in Fig. 2I,J and blue region in Fig. 2K,L ). As expected, when the repetition of single sequences becomes significant, the regular sequences cannot be seen as a single noisy pattern. In fact, if it is true, the occurrence of the distinct 72 sequences represented in the EC should roughly follow a homogeneous distribution. This is not the case because only 8 out of 72 sequences occur with high probability ( r/8 ) while the remaining ones have probability (1−r)/11 5 .
A number of other sources have been tested, by varying k max and N , with matching results. In general, the non optimality of the cost criterion leads to conservative choices about the number of the ECs. Moreover, by slightly modifying the cost criterion, the selection of ECs becomes more inclusive. The algorithm typically achieves good performances in the separation of regular components from noise. These include well-tuned generalization capabilities to avoid noise-induced multiplication of the ECs for messages that are too short.
RESULTS
We used four different simulation groups to validate the algorithm.
We generated independent geometric processes with the parameter p a = [p a1 … p aN ] for a wide range of N sources and spiking frequencies f a = p a \Δt .
We generated M independent geometric processes with the parameter p a , given M≤N . We called these events activation processes ( Fig. 2A ). At each event of activation, processes were assigned to a random subset of N\M sources ( Fig. 2B ). Then, for a period of T time-bins (the length of the activation period), the activated sources were driven by independent geometric processes with parameter p b = f b Δt .
We used the same strategy as simulation group 2, but we assigned the M activation processes to M predefined subsets of N\M sources.
All predefined M subsets were driven by a common trigger following a geometric process with parameter p a . When an event took place all subsets were activated at successive lags of T time-bins.
Activation processes embody the strong non-stationarity of neuronal activity typical both of conscious and unconscious states. In particular, they closely mimic the transition between ‘up’ and ‘down’ states recorded both intracellularly and extracellularly during recordings of ongoing activity in sleep or anesthesia [26] . The term ‘up’ states commonly indicates a depolarized state of intracellular potential typically constituting the substrate for high frequency discharges. Conversely, the term ‘down’ states indicates a hyperpolarized state of intracellular potential leading the cell firing activity to very low frequency regimes.
Homogeneous poisson-like activity is simulated by group 1 ( Fig. 3A ). Multisite activation processes can randomly distribute over the spiking sources (like in group 2, Fig. 3B ), repeat on stereotyped subsets of sources (group 3, Fig. 3C ) and even exhibit precise serial patterns among the different subsets (group 4, Fig. 3D ). Such serial patterns were effectively observed in hippocampal place cells of rats [19] in recordings performed both during spatially constrained tasks and, soon after that, during sleep. Task-related serial patterns expressed during asleep could be time-compressed up to a factor of 20.
A) Structure of a single simulated source. Spikes are displayed as blue points. (*) Represent an inter-spike interval during an active state, having set at f b the mean frequency during the active periods. (**) Represent a period between two successive activation given f a the mean frequency of activation occurrence. B) Given N = 4 channels, we can observe three successive activations of overlapping subsets of N/M = 2 channels.
Constant time bin sampling was obtained by dividing each simulation into constant periods containing, on average, k spikes. The x axes in plots H–J represent these progressively augmenting periods called, for brevity,