Computer Science

The backpropagation of error algorithm (backprop) has been instrumental in the recent success of deep learning. However, a key question remains as to whether backprop can be formulated in a manner suitable for implementation in neural circuitry. The primary challenge is to ensure that any candidate formulation uses only local information, rather than relying on global signals as in standard backprop. Recently several algorithms for approximating backprop using only local signals have been proposed. However, these algorithms typically impose other requirements which challenge biological plausibility: for example, requiring complex and precise connectivity schemes, or multiple sequential backwards phases with information being stored across phases. Here, we propose a novel algorithm, Activation Relaxation (AR), which is motivated by constructing the backpropagation gradient as the equilibrium point of a dynamical system. Our algorithm converges rapidly and robustly to the correct backpropagation gradients, requires only a single type of computational unit, utilises only a single parallel backwards relaxation phase, and can operate on arbitrary computation graphs. We illustrate these properties by training deep neural networks on visual classification tasks, and describe simplifications to the algorithm which remove further obstacles to neurobiological implementation (for example, the weight-transport problem, and the use of nonlinear derivatives), while preserving performance.

This paper introduces AdaSwarm, a novel gradient-free optimizer which has similar or even better performance than the Adam optimizer adopted in neural networks. In order to support our proposed AdaSwarm, a novel Exponentially weighted Momentum Particle Swarm Optimizer (EMPSO), is proposed. The ability of AdaSwarm to tackle optimization problems is attributed to its capability to perform good gradient approximations. We show that, the gradient of any function, differentiable or not, can be approximated by using the parameters of EMPSO. This is a novel technique to simulate GD which lies at the boundary between numerical methods and swarm intelligence. Mathematical proofs of the gradient approximation produced are also provided. AdaSwarm competes closely with several state-of-the-art (SOTA) optimizers. We also show that AdaSwarm is able to handle a variety of loss functions during backpropagation, including the maximum absolute error (MAE).

Control was from its very beginning an important concept in Cybernetics. Later on, with the works of W. Ross Ashby, for example, biological concepts such as adaptation were interpreted in the light of cybernetic systems theory. Adaptation is the process by which a system is capable of regulating or controlling itself in order to adapt to changes of its inner and outer environment in order to maintain a homeostatic state. In earlier works we have developed a metamodel that on the one hand refers to cybernetic concepts such as structure, operation, and system, and on the other to the philosophy of individuation of Gilbert Simondon. The result is the so-called allagmatic method that is capable of creating concrete models of systems such as artificial neural networks and cellular automata starting from abstract building blocks. In this paper, we add to our already existing method the cybernetic concepts of control and especially adaptation. In regard to the metamodel, we rely again on philosophical theories, this time the philosophy of organism of Alfred N. Whitehead. We show how these new meta-theoretical concepts are described formally and how they are implemented in program code. We also show what role they play in simple experiments. We conclude that philosophical abstract concepts help to better understand the process of creating computer models and their control and adaptation. In the outlook we discuss how the allagmatic method needs to be extended in order to cover the field of complex systems and Norbert Wiener's ideas on control.

In evolutionary computation, different reproduction operators have various search dynamics. To strike a well balance between exploration and exploitation, it is attractive to have an adaptive operator selection (AOS) mechanism that automatically chooses the most appropriate operator on the fly according to the current status. This paper proposes a new AOS mechanism for multi-objective evolutionary algorithm based on decomposition (MOEA/D). More specifically, the AOS is formulated as a multi-armed bandit problem where the dynamic Thompson sampling (DYTS) is applied to adapt the bandit learning model, originally proposed with an assumption of a fixed award distribution, to a non-stationary setup. In particular, each arm of our bandit learning model represents a reproduction operator and is assigned with a prior reward distribution. The parameters of these reward distributions will be progressively updated according to the performance of its performance collected from the evolutionary process. When generating an offspring, an operator is chosen by sampling from those reward distribution according to the DYTS. Experimental results fully demonstrate the effectiveness and competitiveness of our proposed AOS mechanism compared with other four state-of-the-art MOEA/D variants.

The adaptive learning capabilities seen in biological neural networks are largely a product of the self-modifying behavior emerging from online plastic changes in synaptic connectivity. Current methods in Reinforcement Learning (RL) only adjust to new interactions after reflection over a specified time interval, preventing the emergence of online adaptivity. Recent work addressing this by endowing artificial neural networks with neuromodulated plasticity have been shown to improve performance on simple RL tasks trained using backpropagation, but have yet to scale up to larger problems. Here we study the problem of meta-learning in a challenging quadruped domain, where each leg of the quadruped has a chance of becoming unusable, requiring the agent to adapt by continuing locomotion with the remaining limbs. Results demonstrate that agents evolved using self-modifying plastic networks are more capable of adapting to complex meta-learning learning tasks, even outperforming the same network updated using gradient-based algorithms while taking less time to train.

Reservoir computers are a type of neuromorphic computer that may be built a an analog system, potentially creating powerful computers that are small, light and consume little power. Typically a reservoir computer is build by connecting together a set of nonlinear nodes into a network; connecting the nonlinear nodes may be difficult or expensive, however. This work shows how a reservoir computer may be expanded by adding functions to its output. The particular functions described here are linear filters, but other functions are possible. The design and construction of linear filters is well known, and such filters may be easily implemented in hardware such as field programmable gate arrays (FPGA's). The effect of adding filters on the reservoir computer performance is simulated for a signal fitting problem, a prediction problem and a signal classification problem.

In the evolutionary multi-objective optimization (EMO) community, it is usually assumed that the final population is presented to the decision maker as the result of the execution of an EMO algorithm. Recently, an unbounded external archive was used to evaluate the performance of EMO algorithms in some studies where a pre-specified number of solutions are selected from all the examined non-dominated solutions. In this framework, which is referred to as the solution selection framework, the final population does not have to be a good solution set. Thus, the solution selection framework offers higher flexibility to the design of EMO algorithms than the final population framework. In this paper, we examine the design of MOEA/D under these two frameworks. First, we show that the performance of MOEA/D is improved by linearly changing the reference point specification during its execution through computational experiments with various combinations of initial and final specifications. Robust and high performance of the solution selection framework is observed. Then, we examine the use of a genetic algorithm-based offline hyper-heuristic method to find the best configuration of MOEA/D in each framework. Finally, we further discuss solution selection after the execution of an EMO algorithm in the solution selection framework.

Always-on artificial intelligent (AI) functions such as keyword spotting (KWS) and visual wake-up tend to dominate total power consumption in ultra-low power devices. A key observation is that the signals to an always-on function are sparse in time, which a spiking neural network (SNN) classifier can leverage for power savings, because the switching activity and power consumption of SNNs tend to scale with spike rate. Toward this goal, we present a novel SNN classifier architecture for always-on functions, demonstrating sub-300nW power consumption at the competitive inference accuracy for a KWS and other always-on classification workloads.

Although quality indicators play a crucial role in benchmarking evolutionary multi-objective optimization algorithms, their properties are still unclear. One promising approach for understanding quality indicators is the use of the optimal distribution of objective vectors that optimizes each quality indicator. However, it is difficult to obtain the optimal distribution for each quality indicator, especially when its theoretical property is unknown. Thus, optimal distributions for most quality indicators have not been well investigated. To address these issues, first, we propose a problem formulation of finding the optimal distribution for each quality indicator on an arbitrary Pareto front. Then, we approximate the optimal distributions for nine quality indicators using the proposed problem formulation. We analyze the nine quality indicators using their approximated optimal distributions on eight types of Pareto fronts of three-objective problems. Our analysis demonstrates that uniformly-distributed objective vectors over the entire Pareto front are not optimal in many cases. Each quality indicator has its own optimal distribution for each Pareto front. We also examine the consistency among the nine quality indicators.

Locomotion is a crucial challenge for legged robots that is addressed "effortlessly" by biological networks abundant in nature, named central pattern generators (CPG). The multitude of CPG network models that have so far become biomimetic robotic controllers is not applicable to the emerging neuromorphic hardware, depriving mobile robots of a robust walking mechanism that would result in inherently energy-efficient systems. Here, we propose a brain-morphic CPG controler based on a comprehensive spiking neural-astrocytic network that generates two gait patterns for a hexapod robot. Building on the recently identified astrocytic mechanisms for neuromodulation, our proposed CPG architecture is seamlessly integrated into Intel's Loihi neuromorphic chip by leveraging a real-time interaction framework between the chip and the robotic operating system (ROS) environment, that we also propose. Here, we demonstrate that a Loihi-run CPG can be used to control a walking robot with robustness to sensory noise and varying speed profiles. Our results pave the way for scaling this and other approaches towards Loihi-controlled locomotion in autonomous mobile robots.