• Infrared and Laser Engineering
  • Vol. 54, Issue 3, 20250073 (2025)
Jinshuai DU, Yin DENG, Shiyang CAO, Zeying LU..., Jie LI, Zhuo HAN, Jinmin ZHOU, Ke WANG, Lili GUI and Kun XU|Show fewer author(s)
Author Affiliations
  • National Key Laboratory of Information Photonics and Optical Communications, Beijing University of Posts and Telecommunications, Beijing 100876, China
  • show less
    DOI: 10.3788/IRLA20250073 Cite this Article
    Jinshuai DU, Yin DENG, Shiyang CAO, Zeying LU, Jie LI, Zhuo HAN, Jinmin ZHOU, Ke WANG, Lili GUI, Kun XU. Biological intelligent computing based on in vitro neural networks: Key technologies and research status (invited)[J]. Infrared and Laser Engineering, 2025, 54(3): 20250073 Copy Citation Text show less
    Overview of in vitro neuronal models. (a) Rodent neuronal culture; (b) hiPSC-derived neuronal culture[25]
    Fig. 1. Overview of in vitro neuronal models. (a) Rodent neuronal culture; (b) hiPSC-derived neuronal culture[25]
    Fabrication of multi-compartment microfluidic structures and demonstration of neuron culture. (a) Flow chart of the mold and physical fabrication of multi-compartment microfluidic structures[28]; (b) Schematic diagram of the connection between neuronal culture and axon growth in the dual-compartmental neuron culture[29]; (c)-(d) Physical and fluorescence diagrams of the neuronal culture structure in the four-compartmental neuron culture[30]
    Fig. 2. Fabrication of multi-compartment microfluidic structures and demonstration of neuron culture. (a) Flow chart of the mold and physical fabrication of multi-compartment microfluidic structures[28]; (b) Schematic diagram of the connection between neuronal culture and axon growth in the dual-compartmental neuron culture[29]; (c)-(d) Physical and fluorescence diagrams of the neuronal culture structure in the four-compartmental neuron culture[30]
    Experimental methods and effects of guiding the unidirectional growth of neuronal axons. (a) Flow chart of mold and physical production of multi-compartment microfluidic structure[31]; (b)-(e) Schematic diagrams of four different shapes of structures guiding axonal unidirectional growth[32-35]; (f) Fluorescence effect diagrams of guided neuronal unidirectional growth[35]; (g) Comparison of the scales of the direction of neuronal axonal growth under unidirectional guided structure and straight pathway structure[36]
    Fig. 3. Experimental methods and effects of guiding the unidirectional growth of neuronal axons. (a) Flow chart of mold and physical production of multi-compartment microfluidic structure[31]; (b)-(e) Schematic diagrams of four different shapes of structures guiding axonal unidirectional growth[32-35]; (f) Fluorescence effect diagrams of guided neuronal unidirectional growth[35]; (g) Comparison of the scales of the direction of neuronal axonal growth under unidirectional guided structure and straight pathway structure[36]
    Ablation of biological neural networks using laser ablation systems. (a) Biological neural network under fluorescence imaging before ablation; (b) Biological neural network under fluorescence imaging after ablation;(c) Curve of characteristic indicators of biological neural network changing over time; (d) Biological neuronal synapses after ablation; (e) Biological neuronal synapses 10 days after ablation
    Fig. 4. Ablation of biological neural networks using laser ablation systems. (a) Biological neural network under fluorescence imaging before ablation; (b) Biological neural network under fluorescence imaging after ablation;(c) Curve of characteristic indicators of biological neural network changing over time; (d) Biological neuronal synapses after ablation; (e) Biological neuronal synapses 10 days after ablation
    BNN information input and output methods[75]
    Fig. 5. BNN information input and output methods[75]
    Overall schematic diagram of biological intelligent computing system based on BNN
    Fig. 6. Overall schematic diagram of biological intelligent computing system based on BNN
    Brainoware for speech recognition tasks[9]. (a) Schematic diagram of speech recognition process; (b) Confusion matrix before and after training; (c) The left figure shows the change in accuracy during Brainoware training, and the three dotted lines represent the accuracy of random conditions, logistic regression, and the standard ESN algorithm; the right figure shows the quantitative change in synaptic connections during Brainoware training; (d) Changes in functional connectivity before and after training
    Fig. 7. Brainoware for speech recognition tasks[9]. (a) Schematic diagram of speech recognition process; (b) Confusion matrix before and after training; (c) The left figure shows the change in accuracy during Brainoware training, and the three dotted lines represent the accuracy of random conditions, logistic regression, and the standard ESN algorithm; the right figure shows the quantitative change in synaptic connections during Brainoware training; (d) Changes in functional connectivity before and after training
    mBNN acts as a generalization filter to improve recognition performance. (a) Schematic diagram of speech recognition, introducing the switch operation to verify the generalization ability of mBNN for different speakers; (b) The accuracy of introducing switch and no switch, the data is directly decoded by the linear decoder; (c) The accuracy of introducing switch, no switch and random conditions; (d)-(f) Verify the generalization ability of mBNN for different numbers, the rest is the same as (a)-(c)
    Fig. 8. mBNN acts as a generalization filter to improve recognition performance. (a) Schematic diagram of speech recognition, introducing the switch operation to verify the generalization ability of mBNN for different speakers; (b) The accuracy of introducing switch and no switch, the data is directly decoded by the linear decoder; (c) The accuracy of introducing switch, no switch and random conditions; (d)-(f) Verify the generalization ability of mBNN for different numbers, the rest is the same as (a)-(c)
    (a) DishBrain overview; (b) Schematic diagram of the experimental setup (left) and the corresponding inference model (right); (c) Neural signals are used to control the movements of an animated mouse, whose "brain" is exposed to microscopic imaging; feedback from the environment determines the subsequent electrical stimulation of the network of living neurons in the MEA
    Fig. 9. (a) DishBrain overview; (b) Schematic diagram of the experimental setup (left) and the corresponding inference model (right); (c) Neural signals are used to control the movements of an animated mouse, whose "brain" is exposed to microscopic imaging; feedback from the environment determines the subsequent electrical stimulation of the network of living neurons in the MEA
    Schematic diagram of the experimental setup for physical reservoir computing (PRC) based on living neuronal cultures
    Fig. 10. Schematic diagram of the experimental setup for physical reservoir computing (PRC) based on living neuronal cultures
    (a) Schematic diagram of the experimental setup of reservoir computing based on in vitro neuronal culture. Calcium imaging and light stimulation for optogenetic control were performed using green light and blue light, respectively. DM: dichroic mirror; DMD: digital micromirror device; LED: light-emitting diode; (b) Experimental changes in neuronal weights during the experiment. The gray and white backgrounds represent the learning process and the cessation of learning; (c) FORCE-learning output and weight adjustment derivatives of the left computational execution area. The neural network weights of the computational execution area were continuously trained during the first 400 seconds to adjust the output of FORCE-learning (red solid line) to follow the target signal (blue dotted line). After training, the output signal is almost consistent with the target signal. The green line represents the derivative of the weight adjustment of all neurons
    Fig. 11. (a) Schematic diagram of the experimental setup of reservoir computing based on in vitro neuronal culture. Calcium imaging and light stimulation for optogenetic control were performed using green light and blue light, respectively. DM: dichroic mirror; DMD: digital micromirror device; LED: light-emitting diode; (b) Experimental changes in neuronal weights during the experiment. The gray and white backgrounds represent the learning process and the cessation of learning; (c) FORCE-learning output and weight adjustment derivatives of the left computational execution area. The neural network weights of the computational execution area were continuously trained during the first 400 seconds to adjust the output of FORCE-learning (red solid line) to follow the target signal (blue dotted line). After training, the output signal is almost consistent with the target signal. The green line represents the derivative of the weight adjustment of all neurons
    Jinshuai DU, Yin DENG, Shiyang CAO, Zeying LU, Jie LI, Zhuo HAN, Jinmin ZHOU, Ke WANG, Lili GUI, Kun XU. Biological intelligent computing based on in vitro neural networks: Key technologies and research status (invited)[J]. Infrared and Laser Engineering, 2025, 54(3): 20250073
    Download Citation