nips nips2006 nips2006-165 knowledge-graph by maker-knowledge-mining

165 nips-2006-Real-time adaptive information-theoretic optimization of neurophysiology experiments

Source: pdf

Author: Jeremy Lewi, Robert Butera, Liam Paninski

Abstract: Adaptively optimizing experiments can signiﬁcantly reduce the number of trials needed to characterize neural responses using parametric statistical models. However, the potential for these methods has been limited to date by severe computational challenges: choosing the stimulus which will provide the most information about the (typically high-dimensional) model parameters requires evaluating a high-dimensional integration and optimization in near-real time. Here we present a fast algorithm for choosing the optimal (most informative) stimulus based on a Fisher approximation of the Shannon information and specialized numerical linear algebra techniques. This algorithm requires only low-rank matrix manipulations and a one-dimensional linesearch to choose the stimulus and is therefore efﬁcient even for high-dimensional stimulus and parameter spaces; for example, we require just 15 milliseconds on a desktop computer to optimize a 100-dimensional stimulus. Our algorithm therefore makes real-time adaptive experimental design feasible. Simulation results show that model parameters can be estimated much more efﬁciently using these adaptive techniques than by using random (nonadaptive) stimuli. Finally, we generalize the algorithm to efﬁciently handle both fast adaptation due to spike-history effects and slow, non-systematic drifts in the model parameters. Maximizing the efﬁciency of data collection is important in any experimental setting. In neurophysiology experiments, minimizing the number of trials needed to characterize a neural system is essential for maintaining the viability of a preparation and ensuring robust results. As a result, various approaches have been developed to optimize neurophysiology experiments online in order to choose the “best” stimuli given prior knowledge of the system and the observed history of the cell’s responses. The “best” stimulus can be deﬁned a number of different ways depending on the experimental objectives. One reasonable choice, if we are interested in ﬁnding a neuron’s “preferred stimulus,” is the stimulus which maximizes the ﬁring rate of the neuron [1, 2, 3, 4]. Alternatively, when investigating the coding properties of sensory cells it makes sense to deﬁne the optimal stimulus in terms of the mutual information between the stimulus and response [5]. Here we take a system identiﬁcation approach: we deﬁne the optimal stimulus as the one which tells us the most about how a neural system responds to its inputs [6, 7]. We consider neural systems in † ‡ http://www.prism.gatech.edu/∼gtg120z http://www.stat.columbia.edu/∼liam which the probability p(rt |{xt , xt−1 , ..., xt−tk }, {rt−1 , . . . , rt−ta }) of the neural response rt given the current and past stimuli {xt , xt−1 , ..., xt−tk }, and the observed recent history of the neuron’s activity, {rt−1 , . . . , rt−ta }, can be described by a model p(rt |{xt }, {rt−1 }, θ), speciﬁed by a ﬁnite vector of parameters θ. Since we estimate these parameters from experimental trials, we want to choose our stimuli so as to minimize the number of trials needed to robustly estimate θ. Two inconvenient facts make it difﬁcult to realize this goal in a computationally efﬁcient manner: 1) model complexity — we typically need a large number of parameters to accurately model a system’s response p(rt |{xt }, {rt−1 }, θ); and 2) stimulus complexity — we are typically interested in neural responses to stimuli xt which are themselves very high-dimensional (e.g., spatiotemporal movies if we are dealing with visual neurons). In particular, it is computationally challenging to 1) update our a posteriori beliefs about the model parameters p(θ|{rt }, {xt }) given new stimulus-response data, and 2) ﬁnd the optimal stimulus quickly enough to be useful in an online experimental context. In this work we present methods for solving these problems using generalized linear models (GLM) for the input-output relationship p(rt |{xt }, {rt−1 }, θ) and certain Gaussian approximations of the posterior distribution of the model parameters. Our emphasis is on ﬁnding solutions which scale well in high dimensions. We solve problem (1) by using efﬁcient rank-one update methods to update the Gaussian approximation to the posterior, and problem (2) by a reduction to a highly tractable onedimensional optimization problem. Simulation results show that the resulting algorithm produces a set of stimulus-response pairs which is much more informative than the set produced by random sampling. Moreover, the algorithm is efﬁcient enough that it could feasibly run in real-time. Neural systems are highly adaptive and more generally nonstatic. A robust approach to optimal experimental design must be able to cope with changes in θ. We emphasize that the model framework analyzed here can account for three key types of changes: stimulus adaptation, spike rate adaptation, and random non-systematic changes. Adaptation which is completely stimulus dependent can be accounted for by including enough stimulus history terms in the model p(rt |{xt , ..., xt−tk }, {rt−1 , ..., rt−ta }). Spike-rate adaptation effects, and more generally spike history-dependent effects, are accounted for explicitly in the model (1) below. Finally, we consider slow, non-systematic changes which could potentially be due to changes in the health, arousal, or attentive state of the preparation. Methods We model a neuron as a point process whose conditional intensity function (instantaneous ﬁring rate) is given as the output of a generalized linear model (GLM) [8, 9]. This model class has been discussed extensively elsewhere; brieﬂy, this class is fairly natural from a physiological point of view [10], with close connections to biophysical models such as the integrate-and-ﬁre cell [9], and has been applied in a wide variety of experimental settings [11, 12, 13, 14]. The model is summarized as: tk λt = E(rt ) = f ta aj rt−j ki,t−l xi,t−l + i l=1 (1) j=1 In the above summation the ﬁlter coefﬁcients ki,t−l capture the dependence of the neuron’s instantaneous ﬁring rate λt on the ith component of the vector stimulus at time t − l, xt−l ; the model therefore allows for spatiotemporal receptive ﬁelds. For convenience, we arrange all the stimulus coefﬁcients in a vector, k, which allows for a uniform treatment of the spatial and temporal components of the receptive ﬁeld. The coefﬁcients aj model the dependence on the observed recent activity r at time t − j (these terms may reﬂect e.g. refractory effects, burstiness, ﬁring-rate adaptation, etc., depending on the value of the vector a [9]). For convenience we denote the unknown parameter vector as θ = {k; a}. The experimental objective is the estimation of the unknown ﬁlter coefﬁcients, θ, given knowledge of the stimuli, xt , and the resulting responses rt . We chose the nonlinear stage of the GLM, the link function f (), to be the exponential function for simplicity. This choice ensures that the log likelihood of the observed data is a concave function of θ [9]. Representing and updating the posterior. As emphasized above, our ﬁrst key task is to efﬁciently update the posterior distribution of θ after t trials, p(θt |xt , rt ), as new stimulus-response pairs are trial 100 trial 500 trial 2500 trial 5000 θ true 1 info. max. trial 0 0 random −1 (a) random info. max. 2000 Time(Seconds) Entropy 1500 1000 500 0 −500 0 1000 2000 3000 Iteration (b) 4000 5000 0.1 total time diagonalization posterior update 1d line Search 0.01 0.001 0 200 400 Dimensionality 600 (c) Figure 1: A) Plots of the estimated receptive ﬁeld for a simulated visual neuron. The neuron’s receptive ﬁeld θ has the Gabor structure shown in the last panel (spike history effects were set to zero for simplicity here, a = 0). The estimate of θ is taken as the mean of the posterior, µt . The images compare the accuracy of the estimates using information maximizing stimuli and random stimuli. B) Plots of the posterior entropies for θ in these two cases; note that the information-maximizing stimuli constrain the posterior of θ much more effectively than do random stimuli. C) A plot of the timing of the three steps performed on each iteration as a function of the dimensionality of θ. The timing for each step was well-ﬁt by a polynomial of degree 2 for the diagonalization, posterior update and total time, and degree 1 for the line search. The times are an average over many iterations. The error-bars for the total time indicate ±1 std. observed. (We use xt and rt to abbreviate the sequences {xt , . . . , x0 } and {rt , . . . , r0 }.) To solve this problem, we approximate this posterior as a Gaussian; this approximation may be justiﬁed by the fact that the posterior is the product of two smooth, log-concave terms, the GLM likelihood function and the prior (which we assume to be Gaussian, for simplicity). Furthermore, the main theorem of [7] indicates that a Gaussian approximation of the posterior will be asymptotically accurate. We use a Laplace approximation to construct the Gaussian approximation of the posterior, p(θt |xt , rt ): we set µt to the peak of the posterior (i.e. the maximum a posteriori (MAP) estimate of θ), and the covariance matrix Ct to the negative inverse of the Hessian of the log posterior at µt . In general, computing these terms directly requires O(td2 + d3 ) time (where d = dim(θ); the time-complexity increases with t because to compute the posterior we must form a product of t likelihood terms, and the d3 term is due to the inverse of the Hessian matrix), which is unfortunately too slow when t or d becomes large. Therefore we further approximate p(θt−1 |xt−1 , rt−1 ) as Gaussian; to see how this simpliﬁes matters, we use Bayes to write out the posterior: 1 −1 log p(θ|rt , xt ) = − (θ − µt−1 )T Ct−1 (θ − µt−1 ) + − exp {xt ; rt−1 }T θ 2 + rt {xt ; rt−1 }T θ + const d log p(θ|rt , xt ) −1 = −(θ − µt−1 )T Ct−1 + (2) − exp({xt ; rt−1 }T θ) + rt {xt ; rt−1 }T dθ d2 log p(θ|rt , xt ) −1 = −Ct−1 − exp({xt ; rt−1 }T θ){xt ; rt−1 }{xt ; rt−1 }T dθi dθj (3) Now, to update µt we only need to ﬁnd the peak of a one-dimensional function (as opposed to a d-dimensional function); this follows by noting that that the likelihood only varies along a single direction, {xt ; rt−1 }, as a function of θ. At the peak of the posterior, µt , the ﬁrst term in the gradient must be parallel to {xt ; rt−1 } because the gradient is zero. Since Ct−1 is non-singular, µt − µt−1 must be parallel to Ct−1 {xt ; rt−1 }. Therefore we just need to solve a one dimensional problem now to determine how much the mean changes in the direction Ct−1 {xt ; rt−1 }; this requires only O(d2 ) time. Moreover, from the second derivative term above it is clear that computing Ct requires just a rank-one matrix update of Ct−1 , which can be evaluated in O(d2 ) time via the Woodbury matrix lemma. Thus this Gaussian approximation of p(θt−1 |xt−1 , rt−1 ) provides a large gain in efﬁciency; our simulations (data not shown) showed that, despite this improved efﬁciency, the loss in accuracy due to this approximation was minimal. Deriving the (approximately) optimal stimulus. To simplify the derivation of our maximization strategy, we start by considering models in which the ﬁring rate does not depend on past spiking, so θ = {k}. To choose the optimal stimulus for trial t + 1, we want to maximize the conditional mutual information I(θ; rt+1 |xt+1 , xt , rt ) = H(θ|xt , rt ) − H(θ|xt+1 , rt+1 ) (4) with respect to the stimulus xt+1 . The ﬁrst term does not depend on xt+1 , so maximizing the information requires minimizing the conditional entropy H(θ|xt+1 , rt+1 ) = p(rt+1 |xt+1 ) −p(θ|rt+1 , xt+1 ) log p(θ|rt+1 , xt+1 )dθ = Ert+1 |xt+1 log det[Ct+1 ] + const. rt+1 (5) We do not average the entropy of p(θ|rt+1 , xt+1 ) over xt+1 because we are only interested in the conditional entropy for the particular xt+1 which will be presented next. The equality above is due to our Gaussian approximation of p(θ|xt+1 , rt+1 ). Therefore, we need to minimize Ert+1 |xt+1 log det[Ct+1 ] with respect to xt+1 . Since we set Ct+1 to be the negative inverse Hessian of the log-posterior, we have: −1 Ct+1 = Ct + Jobs (rt+1 , xt+1 ) −1 , (6) Jobs is the observed Fisher information. Jobs (rt+1 , xt+1 ) = −∂ 2 log p(rt+1 |ε = xt θ)/∂ε2 xt+1 xt t+1 t+1 (7) Here we use the fact that for the GLM, the likelihood depends only on the dot product, ε = xt θ. t+1 We can use the Woodbury lemma to evaluate the inverse: Ct+1 = Ct I + D(rt+1 , ε)(1 − D(rt+1 , ε)xt Ct xt+1 )−1 xt+1 xt Ct t+1 t+1 (8) where D(rt+1 , ε) = ∂ 2 log p(rt+1 |ε)/∂ε2 . Using some basic matrix identities, log det[Ct+1 ] = log det[Ct ] − log(1 − D(rt+1 , ε)xt Ct xt+1 ) t+1 = log det[Ct ] + D(rt+1 , ε)xt Ct xt+1 t+1 + o(D(rt+1 , ε)xt Ct xt+1 ) t+1 (9) (10) Ignoring the higher order terms, we need to minimize Ert+1 |xt+1 D(rt+1 , ε)xt Ct xt+1 . In our case, t+1 with f (θt xt+1 ) = exp(θt xt+1 ), we can use the moment-generating function of the multivariate Trial info. max. i.i.d 2 400 −10−4 0 0.05 −10−1 −2 ai 800 2 0 −2 −7 −10 i i.i.d k info. max. 1 1 50 i 1 50 i 1 10 1 i (a) i 100 0 −0.05 10 1 1 (b) i 10 (c) Figure 2: A comparison of parameter estimates using information-maximizing versus random stimuli for a model neuron whose conditional intensity depends on both the stimulus and the spike history. The images in the top row of A and B show the MAP estimate of θ after each trial as a row in the image. Intensity indicates the value of the coefﬁcients. The true value of θ is shown in the second row of images. A) The estimated stimulus coefﬁcients, k. B) The estimated spike history coefﬁcients, a. C) The ﬁnal estimates of the parameters after 800 trials: dashed black line shows true values, dark gray is estimate using information maximizing stimuli, and light gray is estimate using random stimuli. Using our algorithm improved the estimates of k and a. Gaussian p(θ|xt , rt ) to evaluate this expectation. After some algebra, we ﬁnd that to maximize I(θ; rt+1 |xt+1 , xt , rt ), we need to maximize 1 F (xt+1 ) = exp(xT µt ) exp( xT Ct xt+1 )xT Ct xt+1 . t+1 t+1 2 t+1 (11) Computing the optimal stimulus. For the GLM the most informative stimulus is undeﬁned, since increasing the stimulus power ||xt+1 ||2 increases the informativeness of any putatively “optimal” stimulus. To obtain a well-posed problem, we optimize the stimulus under the usual power constraint ||xt+1 ||2 ≤ e < ∞. We maximize Eqn. 11 under this constraint using Lagrange multipliers and an eigendecomposition to reduce our original d-dimensional optimization problem to a onedimensional problem. Expressing Eqn. 11 in terms of the eigenvectors of Ct yields: 1 2 2 F (xt+1 ) = exp( u i yi + ci yi ) ci yi (12) 2 i i i = g( 2 ci yi ) ui yi )h( i (13) i where ui and yi represent the projection of µt and xt+1 onto the ith eigenvector and ci is the corresponding eigenvalue. To simplify notation we also introduce the functions g() and h() which are monotonically strictly increasing functions implicitly deﬁned by Eqn. 12. We maximize F (xt+1 ) by breaking the problem into an inner and outer problem by ﬁxing the value of i ui yi and maximizing h() subject to that constraint. A single line search over all possible values of i ui yi will then ﬁnd the global maximum of F (.). This approach is summarized by the equation: max F (y) = max g(b) · y:||y||2 =e b max y:||y||2 =e,y t u=b 2 ci yi ) h( i Since h() is increasing, to solve the inner problem we only need to solve: 2 ci yi max y:||y||2 =e,y t u=b (14) i This last expression is a quadratic function with quadratic and linear constraints and we can solve it using the Lagrange method for constrained optimization. The result is an explicit system of 1 true θ random info. max. info. max. no diffusion 1 0.8 0.6 trial 0.4 0.2 400 0 −0.2 −0.4 800 1 100 θi 1 θi 100 1 θi 100 1 θ i 100 −0.6 random info. max. θ true θ i 1 0 −1 Entropy θ i 1 0 −1 random info. max. 250 200 i 1 θ Trial 400 Trial 200 Trial 0 (a) 0 −1 20 40 (b) i 60 80 100 150 0 200 400 600 Iteration 800 (c) Figure 3: Estimating the receptive ﬁeld when θ is not constant. A) The posterior means µt and true θt plotted after each trial. θ was 100 dimensional, with its components following a Gabor function. To simulate nonsystematic changes in the response function, the center of the Gabor function was moved according to a random walk in between trials. We modeled the changes in θ as a random walk with a white covariance matrix, Q, with variance .01. In addition to the results for random and information-maximizing stimuli, we also show the µt given stimuli chosen to maximize the information under the (mistaken) assumption that θ was constant. Each row of the images plots θ using intensity to indicate the value of the different components. B) Details of the posterior means µt on selected trials. C) Plots of the posterior entropies as a function of trial number; once again, we see that information-maximizing stimuli constrain the posterior of θt more effectively. equations for the optimal yi as a function of the Lagrange multiplier λ1 . ui e yi (λ1 ) = ||y||2 2(ci − λ1 ) (15) Thus to ﬁnd the global optimum we simply vary λ1 (this is equivalent to performing a search over b), and compute the corresponding y(λ1 ). For each value of λ1 we compute F (y(λ1 )) and choose the stimulus y(λ1 ) which maximizes F (). It is possible to show (details omitted) that the maximum of F () must occur on the interval λ1 ≥ c0 , where c0 is the largest eigenvalue. This restriction on the optimal λ1 makes the implementation of the linesearch signiﬁcantly faster and more stable. To summarize, updating the posterior and ﬁnding the optimal stimulus requires three steps: 1) a rankone matrix update and one-dimensional search to compute µt and Ct ; 2) an eigendecomposition of Ct ; 3) a one-dimensional search over λ1 ≥ c0 to compute the optimal stimulus. The most expensive step here is the eigendecomposition of Ct ; in principle this step is O(d3 ), while the other steps, as discussed above, are O(d2 ). Here our Gaussian approximation of p(θt−1 |xt−1 , rt−1 ) is once again quite useful: recall that in this setting Ct is just a rank-one modiﬁcation of Ct−1 , and there exist efﬁcient algorithms for rank-one eigendecomposition updates [15]. While the worst-case running time of this rank-one modiﬁcation of the eigendecomposition is still O(d3 ), we found the average running time in our case to be O(d2 ) (Fig. 1(c)), due to deﬂation which reduces the cost of matrix multiplications associated with ﬁnding the eigenvectors of repeated eigenvalues. Therefore the total time complexity of our algorithm is empirically O(d2 ) on average. Spike history terms. The preceding derivation ignored the spike-history components of the GLM model; that is, we ﬁxed a = 0 in equation (1). Incorporating spike history terms only affects the optimization step of our algorithm; updating the posterior of θ = {k; a} proceeds exactly as before. The derivation of the optimization strategy proceeds in a similar fashion and leads to an analogous optimization strategy, albeit with a few slight differences in detail which we omit due to space constraints. The main difference is that instead of maximizing the quadratic expression in Eqn. 14 to ﬁnd the maximum of h(), we need to maximize a quadratic expression which includes a linear term due to the correlation between the stimulus coefﬁcients, k, and the spike history coefﬁcients,a. The results of our simulations with spike history terms are shown in Fig. 2. Dynamic θ. In addition to fast changes due to adaptation and spike-history effects, animal preparations often change slowly and nonsystematically over the course of an experiment [16]. We model these effects by letting θ experience diffusion: θt+1 = θt + wt (16) Here wt is a normally distributed random variable with mean zero and known covariance matrix Q. This means that p(θt+1 |xt , rt ) is Gaussian with mean µt and covariance Ct + Q. To update the posterior and choose the optimal stimulus, we use the same procedure as described above1 . Results Our ﬁrst simulation considered the use of our algorithm for learning the receptive ﬁeld of a visually sensitive neuron. We took the neuron’s receptive ﬁeld to be a Gabor function, as a proxy model of a V1 simple cell. We generated synthetic responses by sampling Eqn. 1 with θ set to a 25x33 Gabor function. We used this synthetic data to compare how well θ could be estimated using information maximizing stimuli compared to using random stimuli. The stimuli were 2-d images which were rasterized in order to express x as a vector. The plots of the posterior means µt in Fig. 1 (recall these are equivalent to the MAP estimate of θ) show that the information maximizing strategy converges an order of magnitude more rapidly to the true θ. These results are supported by the conclusion of [7] that the information maximization strategy is asymptotically never worse than using random stimuli and is in general more efﬁcient. The running time for each step of the algorithm as a function of the dimensionality of θ is plotted in Fig. 1(c). These results were obtained on a machine with a dual core Intel 2.80GHz XEON processor running Matlab. The solid lines indicate ﬁtted polynomials of degree 1 for the 1d line search and degree 2 for the remaining curves; the total running time for each trial scaled as O(d2 ), as predicted. When θ was less than 200 dimensions, the total running time was roughly 50 ms (and for dim(θ) ≈ 100, the runtime was close to 15 ms), well within the range of tolerable latencies for many experiments. In Fig. 2 we apply our algorithm to characterize the receptive ﬁeld of a neuron whose response depends on its past spiking. Here, the stimulus coefﬁcients k were chosen to follow a sine-wave; 1 The one difference is that the covariance matrix of p(θt+1 |xt+1 , rt+1 ) is in general no longer just a rankone modiﬁcation of the covariance matrix of p(θt |xt , rt ); thus, we cannot use the rank-one update to compute the eigendecomposition. However, it is often reasonable to take Q to be white, Q = cI; in this case the eigenvectors of Ct + Q are those of Ct and the eigenvalues are ci + c where ci is the ith eigenvalue of Ct ; thus in this case, our methods may be applied without modiﬁcation. the spike history coefﬁcients a were inhibitory and followed an exponential function. When choosing stimuli we updated the posterior for the full θ = {k; a} simultaneously and maximized the information about both the stimulus coefﬁcients and the spike history coefﬁcients. The information maximizing strategy outperformed random sampling for estimating both the spike history and stimulus coefﬁcients. Our ﬁnal set of results, Fig. 3, considers a neuron whose receptive ﬁeld drifts non-systematically with time. We take the receptive ﬁeld to be a Gabor function whose center moves according to a random walk (we have in mind a slow random drift of eye position during a visual experiment). The results demonstrate the feasibility of the information-maximization strategy in the presence of nonstationary response properties θ, and emphasize the superiority of adaptive methods in this context. Conclusion We have developed an efﬁcient implementation of an algorithm for online optimization of neurophysiology experiments based on information-theoretic criterion. Reasonable approximations based on a GLM framework allow the algorithm to run in near-real time even for high dimensional parameter and stimulus spaces, and in the presence of spike-rate adaptation and time-varying neural response properties. Despite these approximations the algorithm consistently provides signiﬁcant improvements over random sampling; indeed, the differences in efﬁciency are large enough that the information-optimization strategy may permit robust system identiﬁcation in cases where it is simply not otherwise feasible to estimate the neuron’s parameters using random stimuli. Thus, in a sense, the proposed stimulus-optimization technique signiﬁcantly extends the reach and power of classical neurophysiology methods. Acknowledgments JL is supported by the Computational Science Graduate Fellowship Program administered by the DOE under contract DE-FG02-97ER25308 and by the NSF IGERT Program in Hybrid Neural Microsystems at Georgia Tech via grant number DGE-0333411. LP is supported by grant EY018003 from the NEI and by a Gatsby Foundation Pilot Grant. We thank P. Latham for helpful conversations. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] I. Nelken, et al., Hearing Research 72, 237 (1994). P. Foldiak, Neurocomputing 38–40, 1217 (2001). K. Zhang, et al., Proceedings (Computational and Systems Neuroscience Meeting, 2004). R. C. deCharms, et al., Science 280, 1439 (1998). C. Machens, et al., Neuron 47, 447 (2005). A. Watson, et al., Perception and Psychophysics 33, 113 (1983). L. Paninski, Neural Computation 17, 1480 (2005). P. McCullagh, et al., Generalized linear models (Chapman and Hall, London, 1989). L. Paninski, Network: Computation in Neural Systems 15, 243 (2004). E. Simoncelli, et al., The Cognitive Neurosciences, M. Gazzaniga, ed. (MIT Press, 2004), third edn. P. Dayan, et al., Theoretical Neuroscience (MIT Press, 2001). E. Chichilnisky, Network: Computation in Neural Systems 12, 199 (2001). F. Theunissen, et al., Network: Computation in Neural Systems 12, 289 (2001). L. Paninski, et al., Journal of Neuroscience 24, 8551 (2004). M. Gu, et al., SIAM Journal on Matrix Analysis and Applications 15, 1266 (1994). N. A. Lesica, et al., IEEE Trans. On Neural Systems And Rehabilitation Engineering 13, 194 (2005).

Reference: text

Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 Real-time adaptive information-theoretic optimization of neurophysiology experiments∗ Jeremy Lewi† School of Bioengineering Georgia Institute of Technology jlewi@gatech. [sent-1, score-0.097]

2 edu Abstract Adaptively optimizing experiments can signiﬁcantly reduce the number of trials needed to characterize neural responses using parametric statistical models. [sent-6, score-0.108]

3 Here we present a fast algorithm for choosing the optimal (most informative) stimulus based on a Fisher approximation of the Shannon information and specialized numerical linear algebra techniques. [sent-8, score-0.285]

4 Simulation results show that model parameters can be estimated much more efﬁciently using these adaptive techniques than by using random (nonadaptive) stimuli. [sent-11, score-0.046]

5 Finally, we generalize the algorithm to efﬁciently handle both fast adaptation due to spike-history effects and slow, non-systematic drifts in the model parameters. [sent-12, score-0.129]

6 In neurophysiology experiments, minimizing the number of trials needed to characterize a neural system is essential for maintaining the viability of a preparation and ensuring robust results. [sent-14, score-0.149]

7 As a result, various approaches have been developed to optimize neurophysiology experiments online in order to choose the “best” stimuli given prior knowledge of the system and the observed history of the cell’s responses. [sent-15, score-0.304]

8 The “best” stimulus can be deﬁned a number of different ways depending on the experimental objectives. [sent-16, score-0.241]

9 One reasonable choice, if we are interested in ﬁnding a neuron’s “preferred stimulus,” is the stimulus which maximizes the ﬁring rate of the neuron [1, 2, 3, 4]. [sent-17, score-0.346]

10 Alternatively, when investigating the coding properties of sensory cells it makes sense to deﬁne the optimal stimulus in terms of the mutual information between the stimulus and response [5]. [sent-18, score-0.507]

11 Here we take a system identiﬁcation approach: we deﬁne the optimal stimulus as the one which tells us the most about how a neural system responds to its inputs [6, 7]. [sent-19, score-0.306]

12 , rt−ta }) of the neural response rt given the current and past stimuli {xt , xt−1 , . [sent-32, score-0.758]

13 , xt−tk }, and the observed recent history of the neuron’s activity, {rt−1 , . [sent-35, score-0.096]

14 Since we estimate these parameters from experimental trials, we want to choose our stimuli so as to minimize the number of trials needed to robustly estimate θ. [sent-39, score-0.21]

15 In particular, it is computationally challenging to 1) update our a posteriori beliefs about the model parameters p(θ|{rt }, {xt }) given new stimulus-response data, and 2) ﬁnd the optimal stimulus quickly enough to be useful in an online experimental context. [sent-43, score-0.321]

16 In this work we present methods for solving these problems using generalized linear models (GLM) for the input-output relationship p(rt |{xt }, {rt−1 }, θ) and certain Gaussian approximations of the posterior distribution of the model parameters. [sent-44, score-0.105]

17 We solve problem (1) by using efﬁcient rank-one update methods to update the Gaussian approximation to the posterior, and problem (2) by a reduction to a highly tractable onedimensional optimization problem. [sent-46, score-0.168]

18 A robust approach to optimal experimental design must be able to cope with changes in θ. [sent-50, score-0.075]

19 We emphasize that the model framework analyzed here can account for three key types of changes: stimulus adaptation, spike rate adaptation, and random non-systematic changes. [sent-51, score-0.364]

20 Adaptation which is completely stimulus dependent can be accounted for by including enough stimulus history terms in the model p(rt |{xt , . [sent-52, score-0.564]

21 Spike-rate adaptation effects, and more generally spike history-dependent effects, are accounted for explicitly in the model (1) below. [sent-59, score-0.191]

22 Finally, we consider slow, non-systematic changes which could potentially be due to changes in the health, arousal, or attentive state of the preparation. [sent-60, score-0.064]

23 Methods We model a neuron as a point process whose conditional intensity function (instantaneous ﬁring rate) is given as the output of a generalized linear model (GLM) [8, 9]. [sent-61, score-0.153]

24 For convenience, we arrange all the stimulus coefﬁcients in a vector, k, which allows for a uniform treatment of the spatial and temporal components of the receptive ﬁeld. [sent-64, score-0.324]

25 The experimental objective is the estimation of the unknown ﬁlter coefﬁcients, θ, given knowledge of the stimuli, xt , and the resulting responses rt . [sent-70, score-1.156]

26 As emphasized above, our ﬁrst key task is to efﬁciently update the posterior distribution of θ after t trials, p(θt |xt , rt ), as new stimulus-response pairs are trial 100 trial 500 trial 2500 trial 5000 θ true 1 info. [sent-74, score-1.111]

27 1 total time diagonalization posterior update 1d line Search 0. [sent-79, score-0.166]

28 001 0 200 400 Dimensionality 600 (c) Figure 1: A) Plots of the estimated receptive ﬁeld for a simulated visual neuron. [sent-81, score-0.101]

29 The neuron’s receptive ﬁeld θ has the Gabor structure shown in the last panel (spike history effects were set to zero for simplicity here, a = 0). [sent-82, score-0.233]

30 The images compare the accuracy of the estimates using information maximizing stimuli and random stimuli. [sent-84, score-0.182]

31 B) Plots of the posterior entropies for θ in these two cases; note that the information-maximizing stimuli constrain the posterior of θ much more effectively than do random stimuli. [sent-85, score-0.375]

32 The timing for each step was well-ﬁt by a polynomial of degree 2 for the diagonalization, posterior update and total time, and degree 1 for the line search. [sent-87, score-0.198]

33 (We use xt and rt to abbreviate the sequences {xt , . [sent-91, score-1.111]

34 ) To solve this problem, we approximate this posterior as a Gaussian; this approximation may be justiﬁed by the fact that the posterior is the product of two smooth, log-concave terms, the GLM likelihood function and the prior (which we assume to be Gaussian, for simplicity). [sent-98, score-0.25]

35 Furthermore, the main theorem of [7] indicates that a Gaussian approximation of the posterior will be asymptotically accurate. [sent-99, score-0.124]

36 We use a Laplace approximation to construct the Gaussian approximation of the posterior, p(θt |xt , rt ): we set µt to the peak of the posterior (i. [sent-100, score-0.729]

37 the maximum a posteriori (MAP) estimate of θ), and the covariance matrix Ct to the negative inverse of the Hessian of the log posterior at µt . [sent-102, score-0.209]

38 Therefore we just need to solve a one dimensional problem now to determine how much the mean changes in the direction Ct−1 {xt ; rt−1 }; this requires only O(d2 ) time. [sent-107, score-0.053]

39 Moreover, from the second derivative term above it is clear that computing Ct requires just a rank-one matrix update of Ct−1 , which can be evaluated in O(d2 ) time via the Woodbury matrix lemma. [sent-108, score-0.071]

40 To choose the optimal stimulus for trial t + 1, we want to maximize the conditional mutual information I(θ; rt+1 |xt+1 , xt , rt ) = H(θ|xt , rt ) − H(θ|xt+1 , rt+1 ) (4) with respect to the stimulus xt+1 . [sent-112, score-2.297]

41 The ﬁrst term does not depend on xt+1 , so maximizing the information requires minimizing the conditional entropy H(θ|xt+1 , rt+1 ) = p(rt+1 |xt+1 ) −p(θ|rt+1 , xt+1 ) log p(θ|rt+1 , xt+1 )dθ = Ert+1 |xt+1 log det[Ct+1 ] + const. [sent-113, score-0.142]

42 rt+1 (5) We do not average the entropy of p(θ|rt+1 , xt+1 ) over xt+1 because we are only interested in the conditional entropy for the particular xt+1 which will be presented next. [sent-114, score-0.093]

43 Jobs (rt+1 , xt+1 ) = −∂ 2 log p(rt+1 |ε = xt θ)/∂ε2 xt+1 xt t+1 t+1 (7) Here we use the fact that for the GLM, the likelihood depends only on the dot product, ε = xt θ. [sent-118, score-1.682]

44 t+1 We can use the Woodbury lemma to evaluate the inverse: Ct+1 = Ct I + D(rt+1 , ε)(1 − D(rt+1 , ε)xt Ct xt+1 )−1 xt+1 xt Ct t+1 t+1 (8) where D(rt+1 , ε) = ∂ 2 log p(rt+1 |ε)/∂ε2 . [sent-119, score-0.578]

45 Using some basic matrix identities, log det[Ct+1 ] = log det[Ct ] − log(1 − D(rt+1 , ε)xt Ct xt+1 ) t+1 = log det[Ct ] + D(rt+1 , ε)xt Ct xt+1 t+1 + o(D(rt+1 , ε)xt Ct xt+1 ) t+1 (9) (10) Ignoring the higher order terms, we need to minimize Ert+1 |xt+1 D(rt+1 , ε)xt Ct xt+1 . [sent-120, score-0.096]

46 05 10 1 1 (b) i 10 (c) Figure 2: A comparison of parameter estimates using information-maximizing versus random stimuli for a model neuron whose conditional intensity depends on both the stimulus and the spike history. [sent-131, score-0.62]

47 The images in the top row of A and B show the MAP estimate of θ after each trial as a row in the image. [sent-132, score-0.12]

48 C) The ﬁnal estimates of the parameters after 800 trials: dashed black line shows true values, dark gray is estimate using information maximizing stimuli, and light gray is estimate using random stimuli. [sent-137, score-0.096]

49 Gaussian p(θ|xt , rt ) to evaluate this expectation. [sent-139, score-0.559]

50 After some algebra, we ﬁnd that to maximize I(θ; rt+1 |xt+1 , xt , rt ), we need to maximize 1 F (xt+1 ) = exp(xT µt ) exp( xT Ct xt+1 )xT Ct xt+1 . [sent-140, score-1.179]

51 For the GLM the most informative stimulus is undeﬁned, since increasing the stimulus power ||xt+1 ||2 increases the informativeness of any putatively “optimal” stimulus. [sent-142, score-0.467]

52 To obtain a well-posed problem, we optimize the stimulus under the usual power constraint ||xt+1 ||2 ≤ e < ∞. [sent-143, score-0.223]

53 11 under this constraint using Lagrange multipliers and an eigendecomposition to reduce our original d-dimensional optimization problem to a onedimensional problem. [sent-145, score-0.116]

54 11 in terms of the eigenvectors of Ct yields: 1 2 2 F (xt+1 ) = exp( u i yi + ci yi ) ci yi (12) 2 i i i = g( 2 ci yi ) ui yi )h( i (13) i where ui and yi represent the projection of µt and xt+1 onto the ith eigenvector and ci is the corresponding eigenvalue. [sent-147, score-0.68]

55 We maximize F (xt+1 ) by breaking the problem into an inner and outer problem by ﬁxing the value of i ui yi and maximizing h() subject to that constraint. [sent-150, score-0.173]

56 A single line search over all possible values of i ui yi will then ﬁnd the global maximum of F (. [sent-151, score-0.117]

57 250 200 i 1 θ Trial 400 Trial 200 Trial 0 (a) 0 −1 20 40 (b) i 60 80 100 150 0 200 400 600 Iteration 800 (c) Figure 3: Estimating the receptive ﬁeld when θ is not constant. [sent-169, score-0.101]

58 A) The posterior means µt and true θt plotted after each trial. [sent-170, score-0.105]

59 To simulate nonsystematic changes in the response function, the center of the Gabor function was moved according to a random walk in between trials. [sent-172, score-0.113]

60 We modeled the changes in θ as a random walk with a white covariance matrix, Q, with variance . [sent-173, score-0.101]

61 In addition to the results for random and information-maximizing stimuli, we also show the µt given stimuli chosen to maximize the information under the (mistaken) assumption that θ was constant. [sent-175, score-0.173]

62 Each row of the images plots θ using intensity to indicate the value of the different components. [sent-176, score-0.057]

63 B) Details of the posterior means µt on selected trials. [sent-177, score-0.105]

64 C) Plots of the posterior entropies as a function of trial number; once again, we see that information-maximizing stimuli constrain the posterior of θt more effectively. [sent-178, score-0.459]

65 equations for the optimal yi as a function of the Lagrange multiplier λ1 . [sent-179, score-0.079]

66 ui e yi (λ1 ) = ||y||2 2(ci − λ1 ) (15) Thus to ﬁnd the global optimum we simply vary λ1 (this is equivalent to performing a search over b), and compute the corresponding y(λ1 ). [sent-180, score-0.117]

67 For each value of λ1 we compute F (y(λ1 )) and choose the stimulus y(λ1 ) which maximizes F (). [sent-181, score-0.223]

68 This restriction on the optimal λ1 makes the implementation of the linesearch signiﬁcantly faster and more stable. [sent-183, score-0.062]

69 To summarize, updating the posterior and ﬁnding the optimal stimulus requires three steps: 1) a rankone matrix update and one-dimensional search to compute µt and Ct ; 2) an eigendecomposition of Ct ; 3) a one-dimensional search over λ1 ≥ c0 to compute the optimal stimulus. [sent-184, score-0.591]

70 The most expensive step here is the eigendecomposition of Ct ; in principle this step is O(d3 ), while the other steps, as discussed above, are O(d2 ). [sent-185, score-0.058]

71 Here our Gaussian approximation of p(θt−1 |xt−1 , rt−1 ) is once again quite useful: recall that in this setting Ct is just a rank-one modiﬁcation of Ct−1 , and there exist efﬁcient algorithms for rank-one eigendecomposition updates [15]. [sent-186, score-0.077]

72 While the worst-case running time of this rank-one modiﬁcation of the eigendecomposition is still O(d3 ), we found the average running time in our case to be O(d2 ) (Fig. [sent-187, score-0.114]

73 Incorporating spike history terms only affects the optimization step of our algorithm; updating the posterior of θ = {k; a} proceeds exactly as before. [sent-192, score-0.35]

74 The derivation of the optimization strategy proceeds in a similar fashion and leads to an analogous optimization strategy, albeit with a few slight differences in detail which we omit due to space constraints. [sent-193, score-0.096]

75 The main difference is that instead of maximizing the quadratic expression in Eqn. [sent-194, score-0.061]

76 14 to ﬁnd the maximum of h(), we need to maximize a quadratic expression which includes a linear term due to the correlation between the stimulus coefﬁcients, k, and the spike history coefﬁcients,a. [sent-195, score-0.476]

77 The results of our simulations with spike history terms are shown in Fig. [sent-196, score-0.201]

78 In addition to fast changes due to adaptation and spike-history effects, animal preparations often change slowly and nonsystematically over the course of an experiment [16]. [sent-199, score-0.096]

79 We model these effects by letting θ experience diffusion: θt+1 = θt + wt (16) Here wt is a normally distributed random variable with mean zero and known covariance matrix Q. [sent-200, score-0.097]

80 This means that p(θt+1 |xt , rt ) is Gaussian with mean µt and covariance Ct + Q. [sent-201, score-0.583]

81 To update the posterior and choose the optimal stimulus, we use the same procedure as described above1 . [sent-202, score-0.165]

82 Results Our ﬁrst simulation considered the use of our algorithm for learning the receptive ﬁeld of a visually sensitive neuron. [sent-203, score-0.12]

83 We took the neuron’s receptive ﬁeld to be a Gabor function, as a proxy model of a V1 simple cell. [sent-204, score-0.101]

84 We used this synthetic data to compare how well θ could be estimated using information maximizing stimuli compared to using random stimuli. [sent-207, score-0.182]

85 The stimuli were 2-d images which were rasterized in order to express x as a vector. [sent-208, score-0.12]

86 1 (recall these are equivalent to the MAP estimate of θ) show that the information maximizing strategy converges an order of magnitude more rapidly to the true θ. [sent-210, score-0.094]

87 These results are supported by the conclusion of [7] that the information maximization strategy is asymptotically never worse than using random stimuli and is in general more efﬁcient. [sent-211, score-0.173]

88 The solid lines indicate ﬁtted polynomials of degree 1 for the 1d line search and degree 2 for the remaining curves; the total running time for each trial scaled as O(d2 ), as predicted. [sent-216, score-0.192]

89 When θ was less than 200 dimensions, the total running time was roughly 50 ms (and for dim(θ) ≈ 100, the runtime was close to 15 ms), well within the range of tolerable latencies for many experiments. [sent-217, score-0.046]

90 2 we apply our algorithm to characterize the receptive ﬁeld of a neuron whose response depends on its past spiking. [sent-219, score-0.288]

91 However, it is often reasonable to take Q to be white, Q = cI; in this case the eigenvectors of Ct + Q are those of Ct and the eigenvalues are ci + c where ci is the ith eigenvalue of Ct ; thus in this case, our methods may be applied without modiﬁcation. [sent-221, score-0.146]

92 the spike history coefﬁcients a were inhibitory and followed an exponential function. [sent-222, score-0.201]

93 When choosing stimuli we updated the posterior for the full θ = {k; a} simultaneously and maximized the information about both the stimulus coefﬁcients and the spike history coefﬁcients. [sent-223, score-0.649]

94 The information maximizing strategy outperformed random sampling for estimating both the spike history and stimulus coefﬁcients. [sent-224, score-0.52]

95 3, considers a neuron whose receptive ﬁeld drifts non-systematically with time. [sent-226, score-0.235]

96 We take the receptive ﬁeld to be a Gabor function whose center moves according to a random walk (we have in mind a slow random drift of eye position during a visual experiment). [sent-227, score-0.19]

97 The results demonstrate the feasibility of the information-maximization strategy in the presence of nonstationary response properties θ, and emphasize the superiority of adaptive methods in this context. [sent-228, score-0.114]

98 Conclusion We have developed an efﬁcient implementation of an algorithm for online optimization of neurophysiology experiments based on information-theoretic criterion. [sent-229, score-0.09]

99 Reasonable approximations based on a GLM framework allow the algorithm to run in near-real time even for high dimensional parameter and stimulus spaces, and in the presence of spike-rate adaptation and time-varying neural response properties. [sent-230, score-0.343]

100 Thus, in a sense, the proposed stimulus-optimization technique signiﬁcantly extends the reach and power of classical neurophysiology methods. [sent-232, score-0.049]

similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('rt', 0.559), ('xt', 0.552), ('ct', 0.309), ('stimulus', 0.223), ('glm', 0.147), ('stimuli', 0.12), ('spike', 0.105), ('neuron', 0.105), ('posterior', 0.105), ('trial', 0.103), ('receptive', 0.101), ('history', 0.096), ('adaptation', 0.064), ('ci', 0.063), ('gabor', 0.062), ('coef', 0.06), ('paninski', 0.058), ('eigendecomposition', 0.058), ('det', 0.058), ('jobs', 0.055), ('liam', 0.055), ('ta', 0.054), ('yi', 0.054), ('neurophysiology', 0.049), ('cients', 0.048), ('georgia', 0.048), ('ert', 0.048), ('maximizing', 0.043), ('ui', 0.042), ('tk', 0.042), ('trials', 0.038), ('linesearch', 0.037), ('onedimensional', 0.037), ('rankone', 0.037), ('woodbury', 0.037), ('effects', 0.036), ('response', 0.036), ('update', 0.035), ('ring', 0.035), ('eld', 0.035), ('strategy', 0.034), ('maximize', 0.034), ('et', 0.034), ('changes', 0.032), ('lagrange', 0.03), ('intensity', 0.029), ('drifts', 0.029), ('running', 0.028), ('entropy', 0.028), ('plots', 0.028), ('peak', 0.027), ('responses', 0.027), ('adaptive', 0.027), ('log', 0.026), ('walk', 0.026), ('entropies', 0.026), ('spatiotemporal', 0.026), ('diagonalization', 0.026), ('slow', 0.025), ('optimal', 0.025), ('ciency', 0.024), ('covariance', 0.024), ('hessian', 0.023), ('characterize', 0.023), ('dim', 0.023), ('updating', 0.023), ('ef', 0.023), ('past', 0.023), ('accounted', 0.022), ('aj', 0.021), ('solve', 0.021), ('informative', 0.021), ('exp', 0.021), ('optimization', 0.021), ('search', 0.021), ('eigenvectors', 0.02), ('derivation', 0.02), ('degree', 0.02), ('diffusion', 0.02), ('neural', 0.02), ('online', 0.02), ('instantaneous', 0.02), ('simulation', 0.019), ('random', 0.019), ('system', 0.019), ('inverse', 0.019), ('approximation', 0.019), ('conditional', 0.019), ('gaussian', 0.018), ('ms', 0.018), ('fisher', 0.018), ('timing', 0.018), ('neuroscience', 0.018), ('experimental', 0.018), ('matrix', 0.018), ('interested', 0.018), ('algebra', 0.018), ('quadratic', 0.018), ('estimate', 0.017), ('emphasize', 0.017)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.99999994 165 nips-2006-Real-time adaptive information-theoretic optimization of neurophysiology experiments

Author: Jeremy Lewi, Robert Butera, Liam Paninski

2 0.27553096 154 nips-2006-Optimal Change-Detection and Spiking Neurons

Author: Angela J. Yu

Abstract: Survival in a non-stationary, potentially adversarial environment requires animals to detect sensory changes rapidly yet accurately, two oft competing desiderata. Neurons subserving such detections are faced with the corresponding challenge to discern “real” changes in inputs as quickly as possible, while ignoring noisy ﬂuctuations. Mathematically, this is an example of a change-detection problem that is actively researched in the controlled stochastic processes community. In this paper, we utilize sophisticated tools developed in that community to formalize an instantiation of the problem faced by the nervous system, and characterize the Bayes-optimal decision policy under certain assumptions. We will derive from this optimal strategy an information accumulation and decision process that remarkably resembles the dynamics of a leaky integrate-and-ﬁre neuron. This correspondence suggests that neurons are optimized for tracking input changes, and sheds new light on the computational import of intracellular properties such as resting membrane potential, voltage-dependent conductance, and post-spike reset voltage. We also explore the inﬂuence that factors such as timing, uncertainty, neuromodulation, and reward should and do have on neuronal dynamics and sensitivity, as the optimal decision strategy depends critically on these factors. 1

3 0.24567227 164 nips-2006-Randomized PCA Algorithms with Regret Bounds that are Logarithmic in the Dimension

Author: Manfred K. Warmuth, Dima Kuzmin

Abstract: We design an on-line algorithm for Principal Component Analysis. In each trial the current instance is projected onto a probabilistically chosen low dimensional subspace. The total expected quadratic approximation error equals the total quadratic approximation error of the best subspace chosen in hindsight plus some additional term that grows linearly in dimension of the subspace but logarithmically in the dimension of the instances. 1

4 0.22067302 203 nips-2006-implicit Online Learning with Kernels

Author: Li Cheng, Dale Schuurmans, Shaojun Wang, Terry Caelli, S.v.n. Vishwanathan

Abstract: We present two new algorithms for online learning in reproducing kernel Hilbert spaces. Our ﬁrst algorithm, ILK (implicit online learning with kernels), employs a new, implicit update technique that can be applied to a wide variety of convex loss functions. We then introduce a bounded memory version, SILK (sparse ILK), that maintains a compact representation of the predictor without compromising solution quality, even in non-stationary environments. We prove loss bounds and analyze the convergence rate of both. Experimental evidence shows that our proposed algorithms outperform current methods on synthetic and real data. 1

5 0.21995172 152 nips-2006-Online Classification for Complex Problems Using Simultaneous Projections

Author: Yonatan Amit, Shai Shalev-shwartz, Yoram Singer

Abstract: We describe and analyze an algorithmic framework for online classiﬁcation where each online trial consists of multiple prediction tasks that are tied together. We tackle the problem of updating the online hypothesis by deﬁning a projection problem in which each prediction task corresponds to a single linear constraint. These constraints are tied together through a single slack parameter. We then introduce a general method for approximately solving the problem by projecting simultaneously and independently on each constraint which corresponds to a prediction sub-problem, and then averaging the individual solutions. We show that this approach constitutes a feasible, albeit not necessarily optimal, solution for the original projection problem. We derive concrete simultaneous projection schemes and analyze them in the mistake bound model. We demonstrate the power of the proposed algorithm in experiments with online multiclass text categorization. Our experiments indicate that a combination of class-dependent features with the simultaneous projection method outperforms previously studied algorithms. 1

6 0.18644601 184 nips-2006-Stratification Learning: Detecting Mixed Density and Dimensionality in High Dimensional Point Clouds

7 0.12463009 125 nips-2006-Logarithmic Online Regret Bounds for Undiscounted Reinforcement Learning

8 0.12304231 60 nips-2006-Convergence of Laplacian Eigenmaps

9 0.11848892 99 nips-2006-Information Bottleneck Optimization and Independent Component Extraction with Spiking Neurons

10 0.11356084 98 nips-2006-Inferring Network Structure from Co-Occurrences

11 0.10872178 162 nips-2006-Predicting spike times from subthreshold dynamics of a neuron

12 0.10457277 187 nips-2006-Temporal Coding using the Response Properties of Spiking Neurons

13 0.099826857 61 nips-2006-Convex Repeated Games and Fenchel Duality

14 0.09041921 189 nips-2006-Temporal dynamics of information content carried by neurons in the primary visual cortex

15 0.08891385 1 nips-2006-A Bayesian Approach to Diffusion Models of Decision-Making and Response Time

16 0.083746009 76 nips-2006-Emergence of conjunctive visual features by quadratic independent component analysis

17 0.075996071 75 nips-2006-Efficient sparse coding algorithms

18 0.074848905 44 nips-2006-Bayesian Policy Gradient Algorithms

19 0.073430821 202 nips-2006-iLSTD: Eligibility Traces and Convergence Analysis

20 0.072736457 59 nips-2006-Context dependent amplification of both rate and event-correlation in a VLSI network of spiking neurons

similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, -0.223), (1, -0.191), (2, -0.243), (3, 0.111), (4, -0.161), (5, -0.096), (6, 0.052), (7, -0.296), (8, -0.109), (9, -0.1), (10, 0.039), (11, 0.041), (12, 0.09), (13, -0.021), (14, -0.124), (15, 0.041), (16, -0.002), (17, -0.031), (18, 0.042), (19, 0.007), (20, 0.02), (21, -0.009), (22, 0.062), (23, -0.05), (24, 0.002), (25, 0.05), (26, 0.075), (27, -0.131), (28, -0.052), (29, -0.01), (30, -0.091), (31, 0.011), (32, 0.075), (33, -0.028), (34, 0.04), (35, 0.157), (36, 0.048), (37, 0.085), (38, -0.014), (39, 0.017), (40, -0.062), (41, 0.049), (42, 0.105), (43, -0.004), (44, 0.136), (45, -0.012), (46, -0.125), (47, -0.006), (48, 0.131), (49, -0.087)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.98630351 165 nips-2006-Real-time adaptive information-theoretic optimization of neurophysiology experiments

Author: Jeremy Lewi, Robert Butera, Liam Paninski

2 0.6808145 154 nips-2006-Optimal Change-Detection and Spiking Neurons

Author: Angela J. Yu

3 0.64335728 203 nips-2006-implicit Online Learning with Kernels

Author: Li Cheng, Dale Schuurmans, Shaojun Wang, Terry Caelli, S.v.n. Vishwanathan

4 0.62680238 164 nips-2006-Randomized PCA Algorithms with Regret Bounds that are Logarithmic in the Dimension

Author: Manfred K. Warmuth, Dima Kuzmin

5 0.55308312 152 nips-2006-Online Classification for Complex Problems Using Simultaneous Projections

Author: Yonatan Amit, Shai Shalev-shwartz, Yoram Singer

6 0.53906131 184 nips-2006-Stratification Learning: Detecting Mixed Density and Dimensionality in High Dimensional Point Clouds

7 0.49484807 202 nips-2006-iLSTD: Eligibility Traces and Convergence Analysis

8 0.46117601 189 nips-2006-Temporal dynamics of information content carried by neurons in the primary visual cortex

9 0.44655082 98 nips-2006-Inferring Network Structure from Co-Occurrences

10 0.40224037 29 nips-2006-An Information Theoretic Framework for Eukaryotic Gradient Sensing

11 0.39269868 44 nips-2006-Bayesian Policy Gradient Algorithms

12 0.38677096 6 nips-2006-A Kernel Subspace Method by Stochastic Realization for Learning Nonlinear Dynamical Systems

13 0.37057599 162 nips-2006-Predicting spike times from subthreshold dynamics of a neuron

14 0.34716007 61 nips-2006-Convex Repeated Games and Fenchel Duality

15 0.33785129 1 nips-2006-A Bayesian Approach to Diffusion Models of Decision-Making and Response Time

16 0.33080325 192 nips-2006-Theory and Dynamics of Perceptual Bistability

17 0.30899781 79 nips-2006-Fast Iterative Kernel PCA

18 0.30865923 146 nips-2006-No-regret Algorithms for Online Convex Programs

19 0.30640864 125 nips-2006-Logarithmic Online Regret Bounds for Undiscounted Reinforcement Learning

20 0.30150613 187 nips-2006-Temporal Coding using the Response Properties of Spiking Neurons

similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(1, 0.11), (3, 0.031), (6, 0.176), (7, 0.057), (9, 0.048), (12, 0.013), (20, 0.036), (22, 0.135), (44, 0.075), (57, 0.045), (65, 0.061), (69, 0.02), (71, 0.086)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.88653833 165 nips-2006-Real-time adaptive information-theoretic optimization of neurophysiology experiments

Author: Jeremy Lewi, Robert Butera, Liam Paninski

2 0.86707425 114 nips-2006-Learning Time-Intensity Profiles of Human Activity using Non-Parametric Bayesian Models

Author: Alexander T. Ihler, Padhraic Smyth

Abstract: Data sets that characterize human activity over time through collections of timestamped events or counts are of increasing interest in application areas as humancomputer interaction, video surveillance, and Web data analysis. We propose a non-parametric Bayesian framework for modeling collections of such data. In particular, we use a Dirichlet process framework for learning a set of intensity functions corresponding to different categories, which form a basis set for representing individual time-periods (e.g., several days) depending on which categories the time-periods are assigned to. This allows the model to learn in a data-driven fashion what “factors” are generating the observations on a particular day, including (for example) weekday versus weekend effects or day-speciﬁc effects corresponding to unique (single-day) occurrences of unusual behavior, sharing information where appropriate to obtain improved estimates of the behavior associated with each category. Applications to real–world data sets of count data involving both vehicles and people are used to illustrate the technique. 1

3 0.84430993 5 nips-2006-A Kernel Method for the Two-Sample-Problem

Author: Arthur Gretton, Karsten M. Borgwardt, Malte Rasch, Bernhard Schölkopf, Alex J. Smola

Abstract: We propose two statistical tests to determine if two samples are from different distributions. Our test statistic is in both cases the distance between the means of the two samples mapped into a reproducing kernel Hilbert space (RKHS). The ﬁrst test is based on a large deviation bound for the test statistic, while the second is based on the asymptotic distribution of this statistic. The test statistic can be computed in O(m2 ) time. We apply our approach to a variety of problems, including attribute matching for databases using the Hungarian marriage method, where our test performs strongly. We also demonstrate excellent performance when comparing distributions over graphs, for which no alternative tests currently exist.

4 0.82506478 95 nips-2006-Implicit Surfaces with Globally Regularised and Compactly Supported Basis Functions

Author: Christian Walder, Olivier Chapelle, Bernhard Schölkopf

Abstract: We consider the problem of constructing a function whose zero set is to represent a surface, given sample points with surface normal vectors. The contributions include a novel means of regularising multi-scale compactly supported basis functions that leads to the desirable properties previously only associated with fully supported bases, and show equivalence to a Gaussian process with modiﬁed covariance function. We also provide a regularisation framework for simpler and more direct treatment of surface normals, along with a corresponding generalisation of the representer theorem. We demonstrate the techniques on 3D problems of up to 14 million data points, as well as 4D time series data. 1

5 0.72218961 191 nips-2006-The Robustness-Performance Tradeoff in Markov Decision Processes

Author: Huan Xu, Shie Mannor

Abstract: Computation of a satisfactory control policy for a Markov decision process when the parameters of the model are not exactly known is a problem encountered in many practical applications. The traditional robust approach is based on a worstcase analysis and may lead to an overly conservative policy. In this paper we consider the tradeoff between nominal performance and the worst case performance over all possible models. Based on parametric linear programming, we propose a method that computes the whole set of Pareto efﬁcient policies in the performancerobustness plane when only the reward parameters are subject to uncertainty. In the more general case when the transition probabilities are also subject to error, we show that the strategy with the “optimal” tradeoff might be non-Markovian and hence is in general not tractable. 1

6 0.71376532 62 nips-2006-Correcting Sample Selection Bias by Unlabeled Data

7 0.71067631 61 nips-2006-Convex Repeated Games and Fenchel Duality

8 0.70878029 162 nips-2006-Predicting spike times from subthreshold dynamics of a neuron

9 0.70816463 187 nips-2006-Temporal Coding using the Response Properties of Spiking Neurons

10 0.70311922 83 nips-2006-Generalized Maximum Margin Clustering and Unsupervised Kernel Learning

11 0.70092666 203 nips-2006-implicit Online Learning with Kernels

12 0.69781071 65 nips-2006-Denoising and Dimension Reduction in Feature Space

13 0.69475591 20 nips-2006-Active learning for misspecified generalized linear models

14 0.69324183 195 nips-2006-Training Conditional Random Fields for Maximum Labelwise Accuracy

15 0.69274336 135 nips-2006-Modelling transcriptional regulation using Gaussian Processes

16 0.691441 76 nips-2006-Emergence of conjunctive visual features by quadratic independent component analysis

17 0.69137865 6 nips-2006-A Kernel Subspace Method by Stochastic Realization for Learning Nonlinear Dynamical Systems

18 0.69072407 175 nips-2006-Simplifying Mixture Models through Function Approximation

19 0.68868762 126 nips-2006-Logistic Regression for Single Trial EEG Classification

20 0.68807596 32 nips-2006-Analysis of Empirical Bayesian Methods for Neuroelectromagnetic Source Localization