We focus on the task of approximating the optimal value function in deep reinforcement learning. This iterative process is comprised of solving a sequence of optimization problems where the loss function changes per iteration. The common approach to solving this sequence of problems is to employ modern variants of the stochastic gradient descent algorithm such as Adam. These optimizers maintain their own internal parameters such as estimates of the first-order and the second-order moments of the gradient, and update them over time. Therefore, information obtained in previous iterations is used to solve the optimization problem in the current iteration. We demonstrate that this can contaminate the moment estimates because the optimization landscape can change arbitrarily from one iteration to the next one. To hedge against this negative effect, a simple idea is to reset the internal parameters of the optimizer when starting a new iteration. We empirically investigate this resetting idea by employing various optimizers in conjunction with the Rainbow algorithm. We demonstrate that this simple modification significantly improves the performance of deep RL on the Atari benchmark.

This paper proposes an adaptation mechanism for heterogeneous multi-agent systems to align the agents' internal parameters, based on enforced consensus through strong couplings. Unlike homogeneous systems, where exact consensus is attainable, the heterogeneity in node dynamics precludes perfect synchronization. Nonetheless, previous work has demonstrated that strong coupling can induce approximate consensus, whereby the agents exhibit emergent collective behavior governed by the so-called blended dynamics. Building on this observation, we introduce an adaptation law that gradually aligns the internal parameters of agents without requiring direct parameter communication. The proposed method reuses the same coupling signal employed for state synchronization, which may result in a biologically or sociologically plausible adaptation process. Under a persistent excitation condition, we prove that the linearly parametrized vector fields of the agents converge to each other, thereby making the dynamics asymptotically homogeneous, and leading to exact consensus of the state variables.

The computationally cheap machine learning architecture of random feature maps can be viewed as a single-layer feedforward network in which the weights of the hidden layer are random but fixed and only the outer weights are learned via linear regression. The internal weights are typically chosen from a prescribed distribution. The choice of the internal weights significantly impacts the accuracy of random feature maps. We address here the task of how to best select the internal weights. In particular, we consider the forecasting problem whereby random feature maps are used to learn a one-step propagator map for a dynamical system. We provide a computationally cheap hit-and-run algorithm to select good internal weights which lead to good forecasting skill. We show that the number of good features is the main factor controlling the forecasting skill of random feature maps and acts as an effective feature dimension. Lastly, we compare random feature maps with single-layer feedforward neural networks in which the internal weights are now learned using gradient descent. We find that random feature maps have superior forecasting capabilities whilst having several orders of magnitude lower computational cost.

Artificial neural networks have achieved impressive results in very disparate tasks, both in science and in everyday life. The bottleneck in the optimization of artificial neural networks is the learning procedure, i.e., the process through which the internal parameters of the model are optimized to accomplish a desired task. The learning procedure used in the best networks today is gradient descent, in which the internal parameters are incrementally changed in order to improve performance, as measured by a cost function. In feed forward networks this procedure can be implemented efficiently, using error backpropagation. In more complex networks it is implemented by backpropagation through time. Biological systems that learn do not seem to use error backpropagation as the latter cannot be naturally performed by the internal dynamics of the system. Better understanding of biological learning systems could pass through developing learning algorithms in which the two phases of the model (the neuronal and the learning dynamics) can be implemented using similar procedures (or the same circuitry). Such approaches may also be particularly interesting for implementation in analog physical systems, which may lead to improvements in speed or energy consumption. Quantum versions of neural networks and more generally machine learning have attracted much attention recently, as they could offer improved performance over classical algorithms, see e.g.

We focus on the task of approximating the optimal value function in deep reinforcement learning. This iterative process is comprised of solving a sequence of optimization problems where the loss function changes per iteration. The common approach to solving this sequence of problems is to employ modern variants of the stochastic gradient descent algorithm such as Adam. These optimizers maintain their own internal parameters such as estimates of the first-order and the second-order moments of the gradient, and update them over time. Therefore, information obtained in previous iterations is used to solve the optimization problem in the current iteration. We demonstrate that this can contaminate the moment estimates because the optimization landscape can change arbitrarily from one iteration to the next one. To hedge against this negative effect, a simple idea is to reset the internal parameters of the optimizer when starting a new iteration. We empirically investigate this resetting idea by employing various optimizers in conjunction with the Rainbow algorithm. We demonstrate that this simple modification significantly improves the performance of deep RL on the Atari benchmark.

In this paper, we lay out a novel model of neuroplasticity in the form of a horizontal-vertical integration model of neural processing. The horizontal plane consists of a network of neurons connected by adaptive transmission links. This fits with standard computational neuroscience approaches. Each individual neuron also has a vertical dimension with internal parameters steering the external membrane-expressed parameters. These determine neural transmission. The vertical system consists of (a) external parameters at the membrane layer, divided into compartments (spines, boutons) (b) internal parameters in the sub-membrane zone and the cytoplasm with its protein signaling network and (c) core parameters in the nucleus for genetic and epigenetic information. In such models, each node (=neuron) in the horizontal network has its own internal memory. Neural transmission and information storage are systematically separated. This is an important conceptual advance over synaptic weight models. We discuss the membrane-based (external) filtering and selection of outside signals for processing. Not every transmission event leaves a trace. We also illustrate the neuron-internal computing strategies from intracellular protein signaling to the nucleus as the core system. We want to show that the individual neuron has an important role in the computation of signals. Many assumptions derived from the synaptic weight adjustment hypothesis of memory may not hold in a real brain. We present the neuron as a self-programming device, rather than passively determined by ongoing input. We believe a new approach to neural modeling will benefit the third wave of AI. Ultimately we strive to build a flexible memory system that processes facts and events automatically.

Physics-Informed Neural Networks (PINNs) have recently been proposed to solve scientific and engineering problems, where physical laws are introduced into neural networks as prior knowledge. With the embedded physical laws, PINNs enable the estimation of critical parameters, which are unobservable via physical tools, through observable variables. For example, Power Electronic Converters (PECs) are essential building blocks for the green energy transition. PINNs have been applied to estimate the capacitance, which is unobservable during PEC operations, using current and voltage, which can be observed easily during operations. The estimated capacitance facilitates self-diagnostics of PECs. Existing PINNs are often manually designed, which is time-consuming and may lead to suboptimal performance due to a large number of design choices for neural network architectures and hyperparameters. In addition, PINNs are often deployed on different physical devices, e.g., PECs, with limited and varying resources. Therefore, it requires designing different PINN models under different resource constraints, making it an even more challenging task for manual design. To contend with the challenges, we propose Automated Physics-Informed Neural Networks (AutoPINN), a framework that enables the automated design of PINNs by combining AutoML and PINNs. Specifically, we first tailor a search space that allows finding high-accuracy PINNs for PEC internal parameter estimation. We then propose a resource-aware search strategy to explore the search space to find the best PINN model under different resource constraints. We experimentally demonstrate that AutoPINN is able to find more accurate PINN models than human-designed, state-of-the-art PINN models using fewer resources.

First-order methods for solving quadratic programs (QPs) are widely used for rapid, multiple-problem solving and embedded optimal control in large-scale machine learning. The problem is, these approaches typically require thousands of iterations, which makes them unsuitable for real-time control applications that have tight latency constraints. To address this issue, a research team from the University of California, Princeton University and ETH Zurich has proposed RLQP, an accelerated QP solver based on operator-splitting QP (OSQP) that uses deep reinforcement learning (RL) to compute a policy that adapts the internal parameters of a first-order quadratic program (QP) solver to speed up the solver's convergence rate. The team performed their speed-up on the OSQP solver, which solves QPs using a first-order alternating direction method of multipliers (ADMM), an efficient first-order optimization algorithm. The RLQP strives to learn a policy to adapt the internal parameters of the ADMM algorithm between iterations in order to minimize solve times.

This year's virtual ICML conference hosted 10800 attendees from 75 countries. Apparently, the virtual format makes big research conferences such as ICML more accessible to the AI community all over the world. With almost 5000 research papers submitted to ICML 2020 and an acceptance rate of 21.8%, a total of 1088 papers were presented at the conference. As usual, the Outstanding Papers awards were given to exemplary papers at this year's ICML. To help you stay aware of the most prominent AI research breakthroughs, we've summarized the key ideas of these papers.