MLLS 2010      Second German-Korean Workshop on Machine Learning in Life Sciences

Computational Neuroscience

Abstracts

Note that for all abstracts, only the presenting author is named here.

Tuesday, Feb 2

09:40	Higher-Order Markov Models for Learning Sequential Recall Memory Byoung-Tak Zhang, Seoul National University
Recall memory underlies many, if not all, human cognitive information processing. In particular, associative recall of temporal sequences is known to build the foundation of intelligent behavior of animals and humans, such as generating novel sequential patterns related with actions, gestures, songs, and sentences. While existing machine learning models are good at recognition memory tasks, they are weak at modeling sequential recall memory. In this talk we discuss our on-going research effort to build a probabilistic associative memory architecture that is suitable for life-long learning situations in which the learner is exposed to a stream of sequential data and the observed data can be kept for only limited time, and thus incremental learning is imperative. We present a higher-order Markov model (HOMM) that represents the probabilistic sequential relationships of the data variables by explicitly employing higher-order terms. The crux of the method is to use the hypergraph structure that is evolved by a cross-entropy-based Monte Carlo sampling algorithm to learn the structure and parameters of the variable-order Markov model based on a stream of sequence data. The higher-order terms facilitate fast learning of specific or long-distance temporal patterns while the lower-order terms allow for robust learning of general or short-distance patterns. The performance of the method is demonstrated on cognitive modeling of learning language and music from a dialogue corpus of TV dramas and a collection of MP3 music, respectively.
10:20	The variable discharge of cortical neurons revisited Martin Nawrot, Freie Universität Berlin
The output of individual cortical neurons as recorded in the living brain shows high response variability across experimental repetitions (e.g. Shadlen & Newsome, 1998). Yet, the cortex is able to process sensory information with an intriguing temporal fidelity and behavioral responses are timed with high accuracy in each instance. To solve this apparent contradiction we chose various experimental approaches to investigate different neuron-intrinsic and neuron-extrinsic sources of cortical variability. (1) Using dynamic photo-stimulation in acute neocortical brain slices (Nawrot et al., 2009) we studied the contribution of synaptic variability. (2) Somatic noise current injection in vitro (Nawrot et al., 2008) allowed us to quantify single neuron output variability for balanced excitatory and inhibitory input under stationary input conditions. (3) Analysis of single unit recordings from the motor cortex of behaving monkeys allowed us to measure the additional excess variability in vivo. (4) Observation of large scale brain signals such as human epicortical field potentials (Mehring et al., 2004) were used to monitor network activity on global spatial and temporal scales. I will argue that neuron-intrinsic sources of variability are negligible. About half of the observed single neuron variability in vivo can be explained by the stochastic nature of the balanced input and of synaptic transmission. About 50% of the in vivo variability may be attributed to global ongoing activity dynamics in the cortical network. Our results imply that the Poisson point process is a deficient model for the description of spike train statistics of real cortical neurons.   Mehring C, Nawrot MP, Cardoso de Oliveira S, Vaadia E, Schulze-Bonhage A, Aertsen A, Ball T (2004) Comparing information about arm movement direction in single channels of local and epicortical field potentials from monkey and human motor cortex. J Physiol Paris 98: 498 Nawrot MP, Boucsein C, Rodriguez-Molina V, Riehle A, Aertsen A, Rotter S (2008) Measurement of variability dynamics in cortical spike trains. J Neurosci Meth 169: 374-390 Nawrot MP, Schnepel P, Aertsen A and Boucsein C (2009) Precisely timed signal transmission in neocortical networks with reliable intermediate-range projections. Frontiers in Neural Circuits 3:1 Shadlen & Newsome (1998) The Variable Discharge of Cortical Neurons: Implications for Connectivity, Computation, and Information Coding. J Neurosci 18:3870
11:30	Emergent functional structures and neural correlates of consciousness in the brain Seunghwan Kim, POSTECH
The brain is considered to be the most complex system, a fertile ground for understanding the complexity of its functions through dynamical modeling. In this talk, we present some mathematical models that help to reveal the emergence of complexity in various functions of the brain through functional self-organization processes. Though nonlinear analysis, we also present some recent results on how the functional connectivity arises and changes in the brain, which underlies the consciousness dynamics of the nervous system, in particular, in general anesthesia studies. The implications of our work to the various brain function are discussed.
12:10	Approximate Inference for continuous time Markov processes Manfred Opper, Technische Universität Berlin
Continuous time Markov processes (such as jump processes and diffusions) play an important role in the modelling of dynamical systems in many scientific areas ranging from physics to systems biology. In a variety of applications, the stochastic state of the system as a function of time is not directly observed. One has only access to a set of noisy observations. The problem is then to infer the unknown state path as best as possible. In addition, model parameters (like diffusion constants or transition rates) may also be unknown and have to be estimated from the data. While it is fairly straightforward to present a theoretical solution to optimal estimation, the practical solution can be very time consuming. In this talk I will discuss efficient approximation techniques for this problem. I will then illustrate the method for a toy model of estimating a stochastic stimulus sequence from an observed spike train.
14:00	Trial-to-trial variability of spike response of V1 and saccadic response time Choongkil Lee, Seoul National University
Single neurons in the primary visual cortex (V1) show variability in spike activity in response to an identical visual stimulus. We have examined the behavioral consequences of the variability in spike activity of V1 neurons for visually-guided saccades. We recorded single cell activities from V1 of the monkeys trained to detect and make a saccade toward a visual target of varying contrast, and analyzed trial-to-trial covariation between the onset time or firing rate of neural response and saccadic response time (RT). For the workshop, I will present recent data that contrary to the traditional view, trial-to-trial variability of neural response was correlated with RT variability. Individual V1 neurons accounted for approximately 5% of RT variability on average. When the activity of V1 neurons was pooled, the explanatory power increased with the number of pooled neurons up to about 20 V1 neurons which accounted for at least 18% of RT variability. Neural latency was found to be a better predictor of RT than firing rate was. Interestingly, results indicated that the independent contribution of firing rate for RT was virtually none, suggesting that neural latency and firing rate carried different signals in the visually-guided saccade task.
14:40	How subthreshold membrane-potential resonances can shape spike-train patterns Susanne Schreiber, Humboldt Universität Berlin
Many neurons exhibit subthreshold membrane-potential resonances, which lead to maximal subthreshold voltage responses at preferred non-zero stimulation frequencies. Because subthreshold resonances are known to influence the rhythmic activity at the network level, it is vital to understand how they affect spike generation on the single-cell level. We therefore investigated both resonant and nonresonant neurons of rat entorhinal cortex. A minimal resonate-and-fire type model based on measured physiological parameters captures fundamental properties of experimentally recorded neuronal firing statistics surprisingly well and helps to shed light on the mechanisms that shape spike patterns: 1) subthreshold resonance together with a spike-induced reset of subthreshold oscillations leads to spike clustering and 2) spike-induced dynamics influence the fine structure of interspike interval (ISI) distributions and are responsible for ISI correlations appearing at higher firing rates (3 Hz). Both mechanisms are likely to account for the specific discharge characteristics of various cell types.
15:50	Zero-Crossing-Based Sound Source Localization and Segregation Rhee Man Kil, KAIST
In the human auditory system, sound source localization relies on the comparison of auditory input obtained from two separate ears. The main cues are inter-aural time differences (ITDs) and inter-aural intensity differences (IIDs). For computing ITDs, Jeffress (1948) suggested a simple and intuitive hypothesis to measure ITDs in the auditory system. Motivated by Jeffress's model, ITDs are usually estimated using the cross-correlation (CC) of firing rates of auditory signals coming from the channels of left and right ears. However, this approach requires high computational complexity involved in the computation of CC, and they suffer from inaccuracies in estimating the ITDs, especially in noisy multi-source environments. In this context, a method of estimating ITDs using the zero-crossing time differences (ZCTDs) (Kim and Kil, 2007) detected from the filter-bank outputs of the left and right sensors, was suggested previously. This approach is in accordance with Jeffress's hypothesis in which the time difference is actually measured using delay components and coincidence detectors. In this approach, one of the notable properties in the statistics of ITD estimates is that their variances are closely related to the signal-to-noise ratios (SNRs) of filtered signals, enabling us to perform noise-robust estimation of ITDs using the estimated SNRs. As a result, the ZCTD method with SNR estimation is able to provide an accurate estimate of sound source directions in noisy environments while offering significantly less computational complexity compared to the CC-based methods. On the other hand, the sound source localization plays an important role for segregating the selected sound. The human auditory system is known to be able to select a specific sound source among multiple sound sources even in noisy environments using various cues including these spatial cues, for instance, the capability of handling the cocktail party problem. The concept of sound source localization can be applied to sound segregation using two sensors. In this application, we presume that multiple sound sources are present in noisy environments. Here, we consider a sound segregation method based on the sound source localization using spatial cues such as ITDs and IIDs. In our approach, a target sound source mixed with multiple sound sources is segregated using a mask in the time-frequency domain. Here, for each sound segment in the time-frequency domain, the sound source direction is investigated. Then, the sound segment originating from a target sound source are selected and other sound segments originating from interfering sound sources are blocked. This direction of research is based on Bregman's theory of auditory scene analysis (ASA) (Bregman, 1990), in which the notion of auditory grouping; that is, elements that are likely to have arisen from the same environment event are combined, was used to segregate the target sound from an acoustic mixture. Motivated by Bregman's work, computational auditory scene analysis (CASA) systems are developed to mimic the sound separation ability of human listeners. Some of these works showed the impressive performances of speech segregation using the training procedure for every spatial configuration; that is, the number of sound sources and/or the directions of sound sources. This is not favorable for implementing this method in real applications. In this context, the zero-crossing-based method for a possible solution to these problems is suggested.
16:30	Machine Learning for Critical Feature Extraction Felix Wichmann , Technische Universität Berlin
Understanding perception and the underlying cognitive processes on a behavioral level requires a solution to the feature identification problem: Which are the features on which sensory systems base their computations and what techniques can we use to extract them? Thus one of the central challenges in psychophysics is to try and infer the critical features, or cues, human observers make use of when they see or hear: for real-world, complex stimuli, what aspect of the visual or auditory stimulus actually influences behaviour? Over the last years in my laboratory we have developed exploratory, data-driven non-linear system identification techniques based on modern machine learning methods to infer the critical features from human behavioural judgments. I will present these methods and show what their benefits are over the traditional "classification image" and "bubbles technique" approaches.

Wednesday, Feb 3

09:30	Modelling Cortical Representations Klaus Obermayer, Technische Universität Berlin
In my talk I will first present results from a map model of primary visual cortex, where we analysed how much evidence recent single unit recordings from cat area 17 provide for a particular cortical "operating point". Using a Bayesian analysis we find, that the experimental data most strongly support a regime where the local cortical network provides dominant excitatory and inhibitory recurrent inputs (compared to the feedforward drive). Most interestingly, the data supports an operating regime which is close to the border to instability. Hence it is conceivable, that modulatory effects like visual attention may briefly shift the operating point into these regimes, leading to an increased sensitivity of cortical responses to visual inputs. Secondly, I will talk about new ways to quantify spike count correlations among populations of neurons. I will use copulas to construct discrete multivariate distributions that are appropriate to model dependent spike counts of several neurons. With copulas it is possible to use arbitrary marginal distributions such as Poisson or negative binomial that are better suited for modeling single neuron noise distributions than the most often applied normal approximation. Copulas place a wide range of dependence structures at the disposal and can thus be used to quantify higher order interactions. I will apply this framework to multi-tetrode data recorded from macaque prefrontal cortex, where standard noise models fail to accurately describe the measured spike-count distribution. Finally, I will discuss results of developmental perturbations imposed on the visual system of adolescent cats through retinal lesions. Using a computational model of visual cortical responses, I will show that the lesion induced changes of neuronal response properties are consistent with spike timing-dependent plasticity (STDP) learning rules. STDP causes visual cortical receptive fields to converge by creating a competition between neurons for the control of spike timing within the network. The spatial scale of this competition appears to depend on the balance of excitation and inhibition and can in principle be controlled by synaptic scaling type mechanisms. This reveals a novel way by which the capacity of cortical learning rules to transfer response properties between neurons can be effectively switched on and off.
10:10	Combining unsupervised learning and linear discriminant analysis for efficient feature extraction based on subtle differences Soo-Young Lee, KAIST
Standard unsupervised feature extraction algorithms such as PCA, ICA (Independent Component Analysis), and NMF (Non-negative Matrix Factorization) are optimized for minimizing reconstruction error, and usually extract the primary information. In many classification applications one needs features optimized for the discrimination capability of the specific task. Here we report two supervised learning algorithms, i.e., discriminant ICA (dICA) and discriminant NMF (dNMF), for feature extraction based on the subtle differences by maximizing discriminative performance. These algorithms are obtained by combining Linear Discriminant Analysis (LDA) with ICA and NMF, respectively. The algorithms are applied to several classification tasks such as emotion recognition from speeches. Since the primary information in speeches is linguistic content and emotional information is at most secondary, the extraction of discriminative features from the subtle differences is essential.
11:20	Waiting times and correlated spiking of neurons Lutz Schimansky-Geier, Humboldt Universität Berlin
We discuss simple models for neuronal activity and calculate waiting time densities. Dynamical features of single neurons and of ensembles are presented. We underline the special role of correlations between spiking. They achieve importance during signal detection in the electroreceptor of the paddlefish. There, coherent epithelial oscillations induce extended serial correlations in the neuronal firing. We show that the coherent epithelial oscillations act to reduce the variability of neuronal firing, and to enhance the discriminability of weak signals.
12:00	Hypernetwork Models for Sequential Context Learning: Application to Music Generation Byoung-Hee Kim, Seoul National University
Modeling sequential patterns has been challenging problem in various fields, e.g., motif-finding from biological sequences in bioinformatics, part-of-speech problem in language processing. Key issue to the sequential modeling is how to extract and learn underlying context. Hidden Markov models or dynamic Bayesian networks are usual choices for dealing with sequential data. We present hypernetwork models for sequential context modeling. Elements of these models are evolving hypergraph structure and simple operations for learning and inference, which makes the models being suited for hardware implementation. We focus on the application to music generation in this talk and show methods and results of automatic generation of MIDI melodies with given musical styles. The implication of our work to neural computation is discussed.
14:00	An efficient feature selection wrapper with paired t-test for naive Bayes classifiers: an application to microarray data Kyu-Baek Hwang, Soongsil University
Feature selection is one of the most crucial steps in pattern recognition. It influences the whole subsequent classification procedures. There are two kinds of approaches for this task. According to the given situation, they are divided into filter-based and wrapper-based methods. In general, filter-based feature selection requires less computational effort than wrapper-based approaches, and yields less optimal classification performance. In this paper, we propose an Efficient wRapper based on Paired t-Test (ERPT), balancing the above trade-off for naive Bayes classifiers. The ERPT first finds strongly-relevant features which are statistically significant. Then, some of weakly-relevant features are added. The ERPT consumes much smaller computational resources than usual wrapper-based methods and produces better classification performance than usual filter-based methods, especially when the number of features are relatively larger than sample size. Thus, our method is readily applicable to a gigantic data set in the viewpoint of wrapper-based feature selection. We show classification performances and time complexity of the proposed method via an extensive set of experiments.
14:40	Decoding conscious and unconscious mental states from brain activity: From basic science to applied neurotechnology John-Dylan Haynes, Charité Universitätsmedizin Berlin
Recent advances in neuroimaging have shown that it is possible to decode a person's mental states with high precision. Such "brain reading" is achieved by training pattern classifiers to recognize the specific activity patterns that encode individual thoughts in the human brain. Many types of mental states have been decoded, including percepts, memories, intentions, and even unconscious decisions. This research has already provided new insights into a number of fundamental questions regarding the human mind, such as the link between brain and conscious experience, the processes underlying decisions and free will. Now this field is also beginning to explore the first promising applications. The list of potential real-world applications is long and includes brain-computer-interfaces, detection of residual mental activity in coma patients, lie detection, reading out concealed mental states and neuromarketing. This talk will give an overview of this emerging field with a particular emphasis on our work on conscious and unconscious decision making and on potential applications.
15:50	Machine learning techniques in functional neuroimaging of mental disorders Philipp Sterzer, Charité Universitätsmedizin Berlin
In the last few years there has been growing interest in the use of machine learning classifiers for analyzing neuroimaging data. Machine learning algorithms have been used to train classifiers to decode stimuli, mental states, behaviours and other variables of interest from human brain imaging data and have shown that data can contain pattern information that often cannot be revealed by classical univariate statistical methods. One obvious application of machine learning algorithms is their use in clinical populations, e.g., patients with psychiatric disorders, with a number of possible motivations. These range from the pathophysiological understanding of disease processes to the development of novel tools for diagnostic classification or the prediction of treatment responses. I will review recent advances in the psychiatric-clinical application of machine learning and will focus on methodological issues that represent major challenges for the further development and application of this promising approach in clinical neuroimaging.