Multiscale neuroplasticity modulation mechanism-inspired brain-like continuous learning

Artificial intelligence urgently needs to learn from the microscopic, mesoscopic, macroscopic and other multi-scale neuroplasticity fusion computer system in biological systems in order to inspire the realization of more efficient brain-like continuous learning algorithms and eliminate the widespread catastrophic forgetting caused by artificial learning methods such as backpropagation (BP) in artificial neural networks. Neuromodulators such as dopamine, serotonin, serotonin, and norepinephrine commonly found in biological systems are often released through specific glands (Figure 1A), and remotely diffuse and project to a certain range of target neuron populations, and according to the different levels of modulation concentration, complex modulation effects on local neurons, synapses and other microscopic plasticities (Figure 1B).

Figure 1: Neural modulation in the brain. (A) Four neuromodulators and their pathways. (B) Nonlinear neural modulation. (C) Diverse local plasticity of neural modulation

Inspired by the biological neural modulation mechanism, the team of Xu Bo researchers from the Institute of Automation of the Chinese Academy of Sciences, together with Academician Pu Muming of the Center for Excellence in Brain Science and Intelligent Technology of the Chinese Academy of Sciences, and researcher Li Chengyu of Lingang Laboratory, etc., modeled global neuromodulation plasticity such as dopamine and acetylcholine, and local timing-dependent plasticity (Spike Timing-Dependent Plasticity. STDP) and other multi-scale neuroplasticity mechanisms were integrated to obtain a new brain-like learning method based on neural modulation-dependent plasticity (NACA). This method refers to the complex neural modulation pathway structure in the brain and constructs a mathematical model of the neural modulation pathway in the form of an expected matrix encoded (Figure 2A), generating dopamine supervision signals at different concentrations after receiving the stimulus signal and further affecting local synapses and neuronal plasticity types (Figure 2B). NACA supports the training of pulse (Spiking Neural Network (SNN) and artificial neural network (ANN) neural network (Figure 2D, E) using pure feedforward streaming learning methods, synchronizes with the input signal through global dopamine diffusion support, and even precedes the forward information propagation of the input signal, coupled with selective adjustment of STDP. This allows NACA to exhibit significant advantages in rapid convergence and mitigation of catastrophic forgetting.

Figure 2: NACA calculation model. (A) Neural modulation pathway modeling and population expectation coding. (B) Multiple types of local plasticity. (C-E) THE OVERALL FLOW OF THE NACA ALGORITHM AND ITS ROLE IN SNN AND ANN

In two typical picture and speech pattern recognition tasks, the research team evaluated the NACA algorithm from the aspects of accuracy and computational cost, and selected two global learning algorithms E-prop and BRP as comparisons in SNN, and TP (Target Propagation) and BP algorithms as comparison objects in ANN. On both the image classification (MNIST) and speech recognition (TIDigits) standard datasets, NACA achieves higher classification accuracy (about 1.92%) and lower learning energy consumption (about 98%). After verifying the fitting ability of the static classification task, the research team focused on testing the continuous learning ability of the NACA algorithm on Class-CL and expanding the neural modulation to the neuronal plasticity range (Figure 3A, B). In five categories of continuous learning tasks (including continuous MNIST handwritten digits, continuous Alphabet handwritten letters, continuous MathGreek handwritten mathematical symbols, continuous Cifar-10 natural pictures, and continuous DvsGesture dynamic gestures), NACA algorithm has lower energy consumption than BP and EWC algorithms and is found to greatly alleviate the catastrophic forgetting problem (Figure 3C-G).

Figure 3: The performance of the NACA algorithm in the Class-CL task. (A,B) Neural modulation affects both local neurons and synaptic plasticity. (C-G) NACA performance comparison with EWC, BP, etc

Researchers believe that NACA is a class of biologically reasonable global optimization algorithms, which use macroscopic plasticity to further “modulate” local plasticity, which can be regarded as a “Plasticity of Plasticity” method, which has intuitive functional consistency with “learn to learn” and “learn to learn”. The algorithm also obtains performance and computing cost advantages in the optimization problem of SNN and ANN, and plays an important role in continuous learning, which is more suitable for the dynamic task paradigm of biological living environment and practical application scenarios, and these comprehensive features such as pure feedforward learning, low training energy consumption, and support for dynamic continuous learning are also expected to further guide the design of new brain-like chips.

The relevant research work has been supported by the Strategic Pilot Project A of the Chinese Academy of Sciences, the Shanghai Science and Technology Major Project, the Lingang Laboratory Project, and the Youth Promotion Association of the Chinese Academy of Sciences. The research paper, titled A brain-inspired algorithm that mitigates catastrophic forgetting of artificial and spiking neural networks with low computational cost, has been published online in Science Advances, a journal of Science. (Source: Institute of Automation, Chinese Academy of Sciences)

Related paper information:

Special statement: This article is reproduced only for the need to disseminate information, and does not mean to represent the views of this website or confirm the authenticity of its content; If other media, websites or individuals reprint and use from this website, they must retain the “source” indicated on this website and bear their own legal responsibilities such as copyright; If the author does not wish to be reprinted or contact the reprint fee, please contact us.

Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button