Learning, Inference, and Replay of Hidden State Sequences in Recurrent Spiking Neural Networks

Learning to recognize, predict, and generate spatio-temporal patterns and sequences of spikes is a key feature of nervous systems, and essential for solving basic tasks like localization and navigation. How this can be done by a spiking network, however, remains an open question. Here we present a STDP-based framework extending a previous model [1], that can simultaneously learn to abstract hidden states from sensory inputs and learn transition probabilities [2] between these states in recurrent connection weights.

Learning and generation of slow sequences: an application to music composition

Florian François Colombo

Human brains can deal with sequences with temporal dependencies on a broad range of timescales, many of which are several order of magnitude longer than neuronal timescales. Here we introduce an artificial intelligence that learns and produces the complex structure of music, a specific type of slow sequence. Our model employs a separation of fundamental features and multi-layer networks of gated recurrent units. We separate the information contained in monophonic melodies into their rhythm and melody features. The model processes these features in parallel while modelling the relation between them, effectively splitting the joint distribution over note duration and pitch into conditional probabilities. Using such an approach, we were able to automatically learn the temporal dependencies inherent of large corpora of Irish folk songs. We could use the extracted structural rules to generate interesting complete melodies or suggest possible continuations of melody fragments that are coherent with the characteristics of the fragments themselves.

2016

Applications of Approximate Learning and Inference for Probabilistic Models

Young Jun Ko

We develop approximate inference and learning methods for facilitating the use of probabilistic modeling techniques motivated by applications in two different areas. First, we consider the ill-posed inverse problem of recovering an image from an underdetermined system of linear measurements corrupted by noise. Second, we consider the problem of inferring user preferences for items from counts, pairwise comparisons and user activity logs, instances of implicit feedback. Plausible models for images and the noise, incurred when recording them, render posterior inference intractable, while the scale of the inference problem makes sampling based approximations ineffective. Therefore, we develop deterministic approximate inference algorithms for two different augmentations of a typical sparse linear model: first, for the rectified-linear Poisson likelihood, and second, for tree-structured super-Gaussian mixture models. The rectified-linear Poisson likelihood is an alternative noise model, applicable in astronomical and biomedical imaging applications, that operate in intensity regimes in which quantum effects lead to observations that are best described by counts of particles arriving at a sensor, as well as in general Poisson regression problems arising in various fields. In this context we show, that the model-specific computations for Expectation Propagation can be robustly solved by a simple dynamic program. Next, we develop a scalable approximate inference algorithm for structured mixture models, that uses a discrete graphical model to represent dependencies between the latent mixture components of a collection of mixture models. Specifically, we use tree-structured mixtures of super-Gaussians to model the persistence across scales of large coefficients of the Wavelet transform of an image for improved reconstruction. In the second part on models of user preference, we consider two settings: the global static and the contextual dynamic setting. In the global static setting, we represent user-item preferences by a latent low-rank matrix. Instead of using numeric ratings we develop methods to infer this latent representation for two types of implicit feedback: aggregate counts of users interacting with a service and the binary outcomes of pairwise comparisons. We model count data using a latent Gaussian bilinear model with Poisson likelihoods. For this model, we show that the Variational Gaussian approximation can be further relaxed to be available in closed-form by adding additional constraints, leading to an efficient inference algorithm. In the second implicit feedback scenario, we infer the latent preference matrix from pairwise preference statements. We combine a low-rank bilinear model with non-parameteric item- feature regression and develop a novel approximate variational Expectation Maximization algorithm that mitigates the computational challenges due to latent couplings induced by the pairwise comparisons. Finally, in the contextual dynamic setting, we model sequences of user activity at the granularity of single interaction events instead of aggregate counts. Routinely gathered in the background at a large scale in many applications, such sequences can reveal temporal and contextual aspects of user behavior through recurrent patterns. To describe such data, we propose a generic collaborative sequence model based on recurrent neural networks, that combines ideas from collaborative filtering and language modeling.

EPFL2017

Towards goal-driven deep neural network models to elucidate human arm proprioception

Pranav Mamidanna, Alexander Mathis

Proprioceptive signals are a critical component of our ability to perform complex movements, identify our posture and adapt to environmental changes. Our movements are generated by a large number of muscles and are sensed via a myriad of different receptor types. Even the most important ones, muscle spindles, carry highly multiplexed information. For instance, arm movements are sensed via distributed and individually ambiguous activity patterns of muscle spindles, which depend on relative joint configurations rather than the absolute hand position. This high dimensional input (~50 muscles for a human arm) of distributed information poses a challenging decoding problem for the nervous system. Given the diversity in muscle activity, what are the necessary computations that the proprioceptive system needs to perform to sense our movements? Here we studied a proprioceptive variant of the handwritten character recognition task to gain insight into potential computations that the proprioceptive system needs to perform. We focussed on handwritten character classification of muscle length configuration patterns that were required to draw that character. We started from a dataset comprising of pen-tip trajectory data recorded while subjects were writing individual single-stroke characters of the Latin alphabet (Williams et al. ICANN 2006), and employed a musculoskeletal model of the human upper limb to generate muscle length configurations corresponding to drawing the pen-tip trajectories in multiple horizontal and vertical planes. Using this model we created a large, scalable dataset of muscle length configurations corresponding to handwritten characters of varying sizes, locations and orientations (n > 105 samples). To determine the difficulty of this problem, we trained support vector machines (SVM) to solve a binary one-vs-all classification task on the dataset, which achieves an accuracy of 0.89 ± 0.08 (mean ± s.e.m). Contrary to naive expectation, reading out the character at the level of muscles is much easier whereas SVMs trained on the same task using pen-tip coordinates performed relatively poorly: 0.75 ± 0.14. This suggests that the musculoskeletal system itself serves as a non-linear projection to a higher dimensional space, which simplifies character recognition. Next we focussed on goal-driven deep neural network architectures to achieve higher accuracy. Training deep neural networks requires a large, diverse datasets and challenging tasks. We found that the scalable dataset for character recognition we generated is large enough to constrain deep convolutional architectures. We are currently exploring the performance of different deeplearning architectures in solving the handwritten character classification task to investigate which representations are learned and what computations are most efficient. We found that convolutional neural networks factoring out temporal and inter-muscle ('spatial') information achieve almost perfect accuracy for the multi-class problem. These preliminary results suggest that neural networks can learn pose-invariant character recognition from muscle configurations.