Search

Understanding how the brain solves demanding computational tasks is one of the most exciting scientific challenges of our times. So far, recurrently connected artificial neural network models (RANNs) were primarily used to reverse-engineer brain computations. We show that it is now also feasible to reverse-engineer computations of detailed data-based large-scale models of cortical microcircuits. Furthermore, results of these analyses produce hypotheses that can readily be tested in biological experiments since they clarify from which neurons one needs to record and what type of information can be expected at specific time points during a trial. We apply this approach to a demanding visual processing task that has often been used in mouse experiments. Both the cortical microcircuit model and RANNs can solve this task as well as the mouse. But the resulting network dynamics matches only for the cortical microcircuit model experimental data on the sparseness of network activity and the impact of individual neurons on the network decision. Reverse-engineering of the computation in the cortical microcircuit model suggests that a particular subset of neurons causes a bifurcation of the network dynamics that triggers the network decision. Altogether, our results introduce a new type of neural network model for brain computations.Teaser: Large-scale modeling on supercomputers provides new tools for understanding how the brain computes

The predictive coding theory, although attractive, is far from being proven. Supporters of this theory agree that bottom-up sensory inputs and top-down predictions of these inputs must be compared in certain types of neurons called error neurons. Excitatory neurons in layer 2/3 (E2/3) of the primary visual cortex (V1) are ideal candidates to act as error neurons, although how these error neurons are generated is poorly understood. In this study, we aimed to gain insight into how the genetically encoded structure of canonical microcircuits in the neocortex implements the emergence of error neurons. To this end, we used a biologically realistic computational model of V1, developed by the Allen Institute, to study the effect that sudden changes in bottom-up input had on the dynamics of E2/3 neurons. We found that the responses of these neurons can be divided into two main classes: one that depolarized (reporting positive errors; dVf neurons) and one that hyperpolarized (reporting negative errors; hVf neurons). Detailed analysis of both network and effective connectivity allowed us to uncover the mechanism that led to the dynamic segregation of these neurons. This mechanism was found to be the competition between the external visual input, originating in the thalamus, and the recurrent inhibition, originating mainly in layers 2/3 and 4. In contrast, we found no evidence of similar division and responses in excitatory infragranular neurons of layers 5 and 6. Our results are in agreement with recent experimental findings and shed light on the mechanisms responsible for the emergence of error neurons.

We analyze visual processing capabilities of a large-scale model for area V1 that arguably provides the most comprehensive accumulation of anatomical and neurophysiological data to date. We find that this brain-like neural network model can reproduce a number of characteristic visual processing capabilities of the brain, in particular the capability to solve diverse visual processing tasks, also on temporally dispersed visual information, with remarkable robustness to noise. This V1 model, whose architecture and neurons markedly differ from those of deep neural networks used in current artificial intelligence (AI), such as convolutional neural networks (CNNs), also reproduces a number of characteristic neural coding properties of the brain, which provides explanations for its superior noise robustness. Because visual processing is substantially more energy efficient in the brain compared with CNNs in AI, such brain-like neural networks are likely to have an impact on future technology: as blueprints for visual processing in more energy-efficient neuromorphic hardware.

The neocortex can be viewed as a tapestry consisting of variations of rather stereotypical local cortical microcircuits. Hence understanding how these microcircuits compute holds the key to understanding brain function. Intense research efforts over several decades have culminated in a detailed model of a generic cortical microcircuit in the primary visual cortex from the Allen Institute. We are presenting here methods and first results for understanding computational properties of this largescale data-based model. We show that it can solve a standard image-change-detection task almost as well as the living brain. Furthermore, we unravel the computational strategy of the model and elucidate the computational role of diverse subtypes of neurons. Altogether this work demonstrates the feasibility and scientific potential of a methodology based on close interaction of detailed data and large-scale computer modelling for understanding brain function.

Recent research resolves the challenging problem of building biophysically plausible spiking neural models that are also capable of complex information processing. This advance creates new opportunities in neuroscience and neuromorphic engineering, which we discussed at an online focus meeting. Recent research resolves the challenging problem of building biophysically plausible spiking neural models that are also capable of complex information processing. Zenke et al. report on a focus meeting discussing the opportunities created by this advance for neuroscience and neuromorphic engineering.

The common procedure in reservoir computing is to take a “found” reservoir, such as a recurrent neural network with randomly chosen synaptic weights or a complex physical device, and to adapt the weights of linear readouts from this reservoir for a particular computing task. We address the question of whether the performance of reservoir computing can be significantly enhanced if one instead optimizes some (hyper)parameters of the reservoir, not for a single task but for the range of all possible tasks in which one is potentially interested, before the weights of linear readouts are optimized for a particular computing task. After all, networks of neurons in the brain are also known to be not randomly connected. Rather, their structure and parameters emerge from complex evolutionary and developmental processes, arguably in a way that enhances the speed and accuracy of subsequent learning of any concrete task that is likely to be essential for the survival of the organism. We apply the Learning-to-Learn (L2L) paradigm to mimic this two-tier process, where a set of (hyper)parameters of the reservoir are optimized for a whole family of learning tasks. We found that this substantially enhances the performance of reservoir computing for the families of tasks that we considered. Furthermore, L2L enables a new form of reservoir learning that tends to enable even faster learning, where not even the weights of readouts need to be adjusted for learning a concrete task. We present demos and performance results of these new forms of reservoir computing for reservoirs that consist of networks of spiking neurons and are hence of particular interest from the perspective of neuroscience and implementations in spike-based neuromorphic hardware. We leave it as an open question of what performance advantage the new methods that we propose provide for other types of reservoirs.

Understanding how one-shot learning can be accomplished through synaptic plasticity in neural networks of the brain is a major open problem. We propose that approximations toBPTTin recurrent networks of spiking neurons (RSNNs) such ase-propcannot achieve this because their local synaptic plasticity is gated by learning signals that are rather ad hoc from a biological perspective: Random projections of instantaneously arising losses at the network outputs, analogously as in Broadcast Alignment for feedforward networks. In contrast, synaptic plasticity is gated in the brain by learning signals such as dopamine, which are emitted by specialized brain areas, e.g. VTA. These brain areas have arguably been optimized by evolution to gate synaptic plasticity in such a way that fast learning of survival-relevant tasks is enabled. We found that a corresponding model architecture, where learning signals are emitted by a separate RSNN that is optimized to facilitate fast learning, enables one-shot learning via local synaptic plasticity in RSNNs for large families of learning tasks. The same learning approach also supports fast spike-based learning of posterior probabilities of potential input sources, thereby providing a new basis for probabilistic reasoning in RSNNs. Our new learning approach also solves an open problem in neuromorphic engineering, where on-chip one-shot learning capability is highly desirable for spike-based neuromorphic devices, but could so far not be achieved. Our method can easily be mapped into neuromorphic hardware, and thereby solves this problem.

Recurrently connected networks of spiking neurons underlie the astounding information processing capabilities of the brain. Yet in spite of extensive research, how they can learn through synaptic plasticity to carry out complex network computations remains unclear. We argue that two pieces of this puzzle were provided by experimental data from neuroscience. A mathematical result tells us how these pieces need to be combined to enable biologically plausible online network learning through gradient descent, in particular deep reinforcement learning. This learning method–called e-prop–approaches the performance of backpropagation through time (BPTT), the best-known method for training recurrent neural networks in machine learning. In addition, it suggests a method for powerful on-chip learning in energy-efficient spike-based hardware for artificial intelligence., Bellec et al. present a mathematically founded approximation for gradient descent training of recurrent neural networks without backwards propagation in time. This enables biologically plausible training of spike-based neural network models with working memory and supports on-chip training of neuromorphic hardware.

Hyperparameters and learning algorithms for neuromorphic hardware are usually chosen by hand. In contrast, the hyperparameters and learning algorithms of networks of neurons in the brain, which they aim to emulate, have been optimized through extensive evolutionary and developmental processes for specific ranges of computing and learning tasks. Occasionally this process has been emulated through genetic algorithms, but these require themselves hand-design of their details and tend to provide a limited range of improvements. We employ instead other powerful gradient-free optimization tools, such as cross-entropy methods and evolutionary strategies, in order to port the function of biological optimization processes to neuromorphic hardware. As an example, we show that this method produces neuromorphic agents that learn very efficiently from rewards. In particular, meta-plasticity, i.e., the optimization of the learning rule which they use, substantially enhances reward-based learning capability of the hardware. In addition, we demonstrate for the first time Learning-to-Learn benefits from such hardware, in particular, the capability to extract abstract knowledge from prior learning experiences that speeds up the learning of new but related tasks. Learning-to-Learn is especially suited for accelerated neuromorphic hardware, since it makes it feasible to carry out the required very large number of network computations.

The membrane-bound, solubilized, and trypsin-treated forms of Mg, Ca-ATPase from E. coli are inhibited by ruthenium red[R R ]. The inhibition is noncompetitive and is reduced at higher sub strate concentrations. n-Butanol-extracted ATPase is not inhibited by ruthenium red and is not activated by KCl.