Ultralow-power optoelectronic synaptic transistors based on polyzwitterion dielectrics for in-sensor reservoir computing

Abstract

Bio-inspired transistor synapses use solid electrolytes to achieve low-power operation and rich synaptic behaviors via ion diffusion and trapping. While these neuromorphic devices hold great promise, they still suffer from challenges such as high leakage currents and power consumption, electrolysis risk, and irreversible conductance changes due to long-range ion migrations and permanent ion trapping. In addition, their response to light is generally limited because of “exciton-polaron quenching”, which restricts their potential in in-sensor neuromorphic visions. To address these issues, we propose replacing solid electrolytes with polyzwitterions, where the cation and anion are covalently concatenated via a flexible alkyl chain, thus preventing long-range ion migrations while inducing good photoresponses to the transistors via interfacial charge trapping. Our detailed studies reveal that polyzwitterion-based transistors exhibit optoelectronic synaptic behavior with ultralow-power consumption (~250 aJ per spike) and enable high-performance in-sensor reservoir computing, achieving 95.56% accuracy in perceiving the trajectory of moving basketballs.

Replacing electrolyte dielectric with polyzwitterions enables ultralow-power optoelectronic synapse for in-sensor computing.

INTRODUCTION

Driven by Moore’s law, digital artificial intelligence (AI) hardware has undergone rapid development, featuring nanometer-sized transistors and ultrahigh integration density (1–6). However, digital AI hardware is still far from parallel the efficiency and intelligence of the brain, mainly because of two reasons. First, the huge amount of data shuttling between sensing, processing, and memory units leads to large energy and time overheads. Second, transistor size is close to its physical limit, and technology node scaling is getting increasingly cost ineffective. This is further intensified by the growing complexity of AI applications such as autonomous driving and smart chatbots (1–11). On the contrary, the human brain features extremely low power consumption (around 20 W), with the energy consumed by a single excitatory spike measuring only 1 fJ (10⁻¹⁵ J) (5, 7–17), while being able to handle complex problems such as real-time visual recognition, making it an attractive model for developing energy-efficient neuromorphic AI hardware.

To emulate the operation of the human brain, bio-inspired ionic transistor-based artificial synapses have been developed (7–9, 12–21). To date, solid electrolyte dielectrics with mobile free ions, such as LiClO₄-doped polyethylene glycol (9, 12, 14, 16), lithium phosphorous oxynitride electrolyte (17), chitosan (15, 16), and others, have been used in transistors to achieve high dielectric capacitances and replicate different synaptic behaviors (8, 16–17, 19). Regulating presynaptic input signals, such as magnitude, pulse width, and pulse number, generates diverse synaptic behaviors, including short-term plasticity (STP), long-term plasticity (LTP), spike time-dependent plasticity, and others (22–24). In addition, such ionic transistor-based artificial synapses feature colocation of neural signal processing and synaptic weight storage, obviating the von Neumann bottleneck. Despite their promise, these solid electrolyte transistors still face several challenges: (i) The mobile free ions would cause a large leakage current between the gate and source/drain electrodes due to long-range ion migrations, which notably increases the static power consumption (25–30); (ii) the high electric fields pose a potential risk of electrolysis (28); (iii) the irreversible synaptic potentiation may result from permanent ion trapping or semiconductor doping (9, 12, 17, 31); (iv) transistors based on these electrolyte dielectrics generally poorly respond to light irradiation due to “exciton-polaron quenching”, disqualifying them as in-sensor neuromorphic vision devices (Fig. 1) (32).

Fig. 1. The advantages of polyzwitterion dielectric in transistor-based artificial synaptic devices.

The ions in conventional electrolyte dielectric can move freely and contribute to high leakage. The ions in polyzwitterion dielectric are linked by alkyl chain and can well suppress the leakage current.

To address the challenges outlined above, we propose a previously unreported electrolyte design strategy that covalently connects the cation and anion via a flexible alkyl chain, specifically poly (sulfobetaine methacrylate) (PSBMA). This design eliminates the long-range migration of free ions during polarization while maintaining well-defined polarization-relaxation behaviors under different applied presynaptic inputs (Fig. 1). In addition, polyzwitterions exhibit high capacitance (up to 2000 nF cm⁻² @20 Hz) and permittivity (k ~ 165), enabling low operation voltage and thus low power consumption (~250 aJ per spike) in synaptic transistors. Furthermore, PSBMA induces pronounced photoresponse in the transistor via an “interfacial trapping” strategy. The well-defined polarization-relaxation behaviors of polyzwitterions render the resulting transistors unprecedently rich optoelectrical synaptic behaviors, including excitatory postsynaptic current (EPSC), paired-pulse facilitation (PPF), STP, LTP, short-term memory (STM), and long-term memory (LTM), making it suitable for image sensing, storage, and real-time preprocessing (33–40). Leveraging these advantages, this artificial synapse enables high-performance in-sensor reservoir computing (RC), achieving 95.56% accuracy in perceiving the trajectory of a moving basketball. Furthermore, the excellent water solubility of polyzwitterion-based flexible devices ensures their degradation upon demand (41), reducing the possibility of e-waste. This work represents an example of opto-neuromorphic computing based on electrolyte dielectrics–based transistors and provides a design for flexible, eco-friendly, and ultralow-power artificial synapses, paving the way for highly efficient edge neuromorphic computing.

RESULTS

Characterization of PSBMA dielectrics

The chemical structure of PSBMA and its capacitor architecture are shown in Fig. 2A. PSBMA has a large dipole moment (~16 D) (41), resulting in a large capacitance of up to 2000 nF cm⁻² at 20 Hz and a dielectric permittivity of 165. Its capacitance remains high even at a high frequency of 10,000 Hz (Fig. 2B), outperforming conventional solid electrolyte dielectrics that drastically decrease their capacitances over frequency and lose over 95% capacitance under identical conditions. Moreover, the lack of mobile free ions in PSBMA chains brings down the conductivity and leakage current to a much lower value than that of solid electrolytes (Fig. 2, C and E), which is critical to reducing static power consumption in synaptic devices. The low leakage current and higher capacitance make PSBMA dielectrics highly suitable for constructing low-power devices. Thermogravimetric analysis demonstrates that PSBMA film has good thermal stability, with only a 5% weight loss at 278°C (fig. S2A). Differential scanning calorimetry confirms the noncrystalline nature of the film (fig. S2B). These features make PSBMA an attractive candidate as a dielectric material in synaptic devices. In addition, the excellent water solubility and flexibility of the PSBMA capacitor enable it to adhere to curved surfaces, such as green leaves, and degrade on demand (Fig. 2D), underscoring its eco-friendly nature. The PSBMA film remains intact in common solvents for device fabrication (fig. S2C), rendering excellent compatibility with the patterning of subsequent semiconducting layers.

Fig. 2. Dielectric properties of PSBMA.

(A) Schematic illustration of the chemical structure of PBSMA and the architecture of the capacitor. (B) Capacitance and the dielectric constant of PSBMA films (thickness, 280 nm) at different frequencies. (C) Leakage current density of PSBMA films at different voltages. (D) The demonstration of degradation of the PSBMA capacitor array. (E) Conductivity comparison of solid electrolyte-based dielectric layers and the polyzwitterion dielectric layer (47, 48). All statistical data are presented as a mean ± SD (n = 10).

Characterization of organic field-effect transistors with PSBMA dielectric layer

The three-terminal organic field-effect transistor (OFET) containing a 2-decyl[1]benzothieno[3,2-b][1]benzothiophene (BTBT-C10) semiconducting layer and a PSBMA dielectric layer is shown in Fig. 3A. The source-drain current (I_d) reaches 10⁻⁴ A at a low operation voltage of −4 V due to the very high capacitance of PSBMA (Fig. 3C and fig. S3). Figure 3B shows a typical output curve of the OFET. The subthreshold swing is only 95 mV/Dec, much lower than the previous reports and very close to the theoretical value of 60 mV/Dec (42), implying that the device could easily turn on. A dual sweep of the transfer curve (V_g sweeps from 0 to −4 V and then back to 0 V) shows an obvious hysteresis loop (fig. S5), manifesting the potential of the OFET as an artificial synapse (43–45). The devices show typical hole mobility (μ), threshold voltage (V_th), and on/off ratio of 1.2 cm² V⁻¹ s⁻¹, −2.1 V, and 10⁶ at −4 V, respectively (Fig. 3C and fig. S4). OFETs bearing other semiconductors [such as Pentacene and N,N′-dioctyl-naphthalene tetracarboxylic diimide (C8-NDI)] also exhibit good electrical performance (fig. S8), indicating the excellent generality of the PSBMA dielectric layer. The above results demonstrate the capability of PSBMA as the dielectric layer for high-performance OFETs. In addition, PSBMA renders the OFET remarkable flexibility. The I_d of the flexible OFETs shows negligible decay even after 400 bending cycles with a bending radius of 0.62 cm (Fig. 3D).

Fig. 3. Characterization of PSBMA-based transistors.

(A) Schematic illustration of the flexible OFETs with PSBMA dielectrics. (B) The output curve and (C) typical p-type transfer curve of the OFETs with BTBT-C10 as the semiconductor and PSBMA as the dielectric layer (V_d = −2 V). (D) The I_d change over bending cycles with a bending radius of 0.62 cm. The statistical data are presented as a mean ± SD (n = 10). (E) The low-power ESPC behavior of the OFETs triggered by a presynaptic spike (V_g = −0.3 V, 28 ms). (F) Comparison of energy consumption of transistors fabricated based on various ionic dielectric layers (49–61).

Next, we evaluate the capability of PSBMA-based OFETs as artificial synapses. As shown in Fig. 4, the OFET exhibits a typical EPSC. Upon being triggered by a presynaptic spike (V_g = −2 V, pulse width = 0.2 s), the EPSC showed a peak current value of 13.5 nA and then gradually returned to the baseline (the EPSC was recorded at a constant V_d of −0.05 V; fig. S9). When a pair of identical presynaptic spikes with an interval (Δt) of 0.2 s was applied to the gate electrode, the EPSC triggered by the second presynaptic spike is twice as large as the EPSC triggered by the first spike, which is similar to the phenomenon of PPF observed in biological synaptic systems. According to the equation of energy consumption (E_c)

$$E_{\mathrm{c}}=V_{\text{read}}\times I_{\text{peak}}\times t$$

where V_read represents the drain voltage, I_peak represents the spike peak current, and t is the spike duration time. The E_c could be effectively reduced by using a small V_read and short presynaptic spike. In Fig. 3E, the V_read is set to −5 μV with a spike duration time of 28 ms, and the EPSC peak current value is −1.7 nA. On the basis of the equation described above, the E_c per spike is approximately 250 aJ, much lower than that of most reported artificial synapses based on electrolyte-gated OFET (EGOFET) (Fig. 3F). This result indicates the potential for low-energy optoelectronic neuromorphic computing devices.

Fig. 4. Photoresponsive of the flexible OFETs.

(A) The net photocurrent of OFETs over V_g after 5- and 10-s light irradiation with an intensity of 4.2 mW cm⁻², respectively (n = 10). (B) The transfer curves of transistor under dark conditions, after 5-s light exposure, and after 10-s light exposure, respectively (light intensity of 4.2 mW cm⁻², V_d = −2 V). The photocurrent response of the transistor (C) under different pulse widths (light intensity of 0.7 mW cm⁻², V_g = V_d = −2 V) and (D) under different light intensities (pulse widths of 4 s, V_g = V_d = −2 V). (E) I_d as a function of the light pulse number. The phototransistors are working in a real-time “sampling” mode with a constant V_g and V_d of −2 V. (F) Illustrations of dynamic learning and forgetting of the letter “F” (light intensity of 0.7 mW cm⁻²). (G) The ESPC behavior of the transistors triggered by a light pulse (intensity: 4.2 mW cm⁻², pulse width: 0.3 s, V_d = −1 μV, V_g = −2 V).

Photoresponsive behaviors of the flexible OFETs

Conventional EGOFETs barely respond to light irradiation due to the “exciton-polaron quenching” effect and are therefore unable to serve as neuromorphic retinas, as reported previously (32). However, PSBMA-based OFETs exhibit a notable increase in I_d when exposed to light irradiation (more than 50 μA at V_g of −4 V), as illustrated in Fig. 4A (fig. S10). The transfer characteristics of the OFETs after exposure to a light intensity of 4.2 mW cm⁻² for 0, 5, and 10 s, respectively, are depicted in Fig. 4B. The I_d increases continuously without rapid saturation over the light intensities or pulse widths (Fig. 4, C and D), which is akin to the STP and LTP of the human brain. Furthermore, the detectivity (D*), photo/dark current ratio (P), and photoresponsivity (R) of the OFET are elaborated in fig. S11. The larger the pulse width, the higher the I_d of the OFETs, demonstrating the potential for use in neuromorphic retinas. Figure 4E demonstrates that as the presynaptic spike numbers increase, the I_d also increases and takes longer to decay, which is a typical process of learning and memory in the human brain where STM is transformed into LTM (fig. S15). These devices could also be reset by clearing the residual photocurrent. As shown in fig. S14, after the light is removed, the photocurrent undergoes a slow decay, but it rapidly recovers upon receiving an electrical pulse (V_g = −2 V, 0.5 s). The device, after reset, is prepared to receive the next sequence of optical pulses and can repeatedly respond to new optical pulses with similar initial photocurrents. These results highlight the robust learning capability of PSBMA-based OFETs. To further demonstrate this capability, we applied a light pulse with an intensity of 4.2 mW cm⁻² to a transistor array through a photomask with an “F” pattern. Each transistor corresponds to one image pixel. As shown in Fig. 4F, the irradiated pixels gain a higher I_d than the masked pixels, and the longer the irradiation time, the higher the I_d of the irradiated pixels. As a result, the letter F is sensed and memorized in the array after removing the light pulse, demonstrating the retina-like signal preprocessing capability of the transistor array. These properties enable the transistor to behave as a photosynapse with a fading memory. In addition, thanks to the high capacitance of the PSBMA layer, the transistor-based photosynapse can also work in a low-power mode. When the transistor is working in a “sampling” mode with a constant V_d of −1 μV under a light irradiation of 0.3 s (4.2 mW cm⁻²), the E_c is approximately 330 aJ per spike (Fig. 4G). It is important to note that in the calculation of power consumption, we have not accounted for the energy consumption attributed to light irradiation. This approach is consistent with the methodology reported in the literature, ensuring good comparability with previous studies (18, 22, 36).

Mechanism study for the photoresponses and fading memory of OFETs

To gain a deeper understanding of the unique photoresponses and fading memory of the PSBMA-based OFETs, we conducted a morphological study of the dielectric layer and semiconductor film using atomic force microscopy (AFM) and grazing-incidence x-ray diffraction (GIXD). Figure 5 (A and B) displays the respective surface morphology of PSBMA and BTBT-C10 films. The PSBMA film is amorphous and featureless, with a root-mean-square (RMS) value of 0.299 nm. In contrast, the BTBT-C10 film shows a high crystalline feature, along with a higher RMS of 1.39 nm. GIXD characterizations provide additional evidence for the high crystallinity of BTBT-C10. The sharp diffraction spots assigned to the {0 0 l} planes in the out-of-plane direction and the {1 k l} planes along the in-plane direction imply that the BTBT-C10 film has a good long-range order in three-dimensional space. The BTBT-C10 molecules adopt a head-to-head and tail-to-tail packing mode, with a periodic spacing of 43.4 and 3.97 Å along the “Z” and “Y” axis, respectively (Fig. 5F; the Cambridge Crystallographic Data Centre deposition number of BTBT-C10 is 1857063). The excellent molecular ordering in the semiconducting layer indicates that the PSBMA layer does not impede the growth of the BTBT-C10 film, which is crucial for achieving high charge carrier mobilities and on/off ratios, as well as low operation voltages. However, highly ordered semiconductors typically result in poor photoresponses of the transistor, as the photogenerated holes and electrons recombine rapidly in the crystal and thereby quench the photocurrent (46). We observed pronounced photoresponses and fading memory of the transistor in this study (Fig. 4). This could be attributed to the interfacial trapping of the charge carrier between the semiconductor and PSBMA dielectrics. The surface charge trap density of PSBMA is around 10 times higher than that of poly(methyl methacrylate) (PMMA) (see supporting information for details). As shown in Fig. 5 (G and H), upon applying a negative V_g, band bending occurs and the holes are injected into the interface between BTBT-C10 and the PSBMA layer from the source electrode. Concurrently, the polyzwitterions align themselves with anions pointing toward the BTBT-C10 layer and cations away from it. Driven by the V_d, these holes become mobile and form I_d. Light irradiation generates more holes at the highest occupied molecular orbital and electrons at the lowest unoccupied molecular orbital of the BTBT-C10, respectively. Part of the holes are trapped by the anions of the polyzwitterions, which slows down the recombination of the electrons and holes, giving rise to a pronounced photocurrent and a fading memory. As a control, OFETs with PMMA dielectrics without ionic traps do not exhibit any fading memory (fig. S13), because the electrons and holes can undergo rapid recombination.

Fig. 5. Characterization of films and mechanism study.

AFM characterizations of (A) PSBMA and (B) BTBT-C10 films, respectively. (C) 2D GIXD of BTBT-C10 films, with the out-of-plane diffraction peaks shown in (D), and in-plane diffraction peaks shown in (E). (F) The schematical illustration of molecular packing, ordering, and orientation of BTBT-C10. (G and H) Schematic illustration of the mechanism of photoresponses and fading memory.

Nonlinear mapping of 5-bit input of the OFETs

The above fading memory, or the nonlinear dynamic evolution of channel current upon the light stimulus of PSBMA-based OFETs, is key to in-sensor RC. This approach leverages the inherent photoresponses dynamics of OFETs for feature extraction, followed by a trainable lightweight readout map, suitable for sequence data analysis at a low training cost. To characterize their capability for sequential input embedding, we applied optical pulse trains to the transistor (Fig. 6B and fig. S16). Each optical pulse train contains five pulses, with “1” and “0” representing an optical pulse with light intensity of 70 and 0 μW cm⁻², respectively, and a pulse width of 2.0 s. Thanks to the efficient exciton dissociation and hole trapping at the BTBT-C10/PSBMA interface, the I_ds increases continuously over consecutive light pulses, rather than rapidly saturating. After the light is removed, the trapped holes at the BTBT-C10/PSBMA interface slowly recombine with the free electrons, resulting in a nonlinear decay of the output I_d and thus a STM. Therefore, the combined effect of photogating and fading memory enables synaptic facilitation/depression and dynamic transistor states with well-separated outputs, mimicking the synapses of visual signal pathway. As shown in Fig. 6C, 32 optical pulse trains ranging from (00000) to (11111) generate 32 clearly distinguishable states, highlighting the robust ability to map complicated spatiotemporal signals into reservoir states.

Fig. 6. Nonlinear mapping of 5-bit input of the OFETs.

(A) Schematical illustration of biological synapse and neural network. (B) Illustration of optical pulse trains applied to the OFETs and the process of RC. (C) Experimental outputs of 32 different 5-bit optical pulse trains, ranging from (00000) to (11111). Each data point represents the average of 10 repeated tests (n = 10).

In-sensor RC for motion recognition

The experimental results about photoresponses to 5-bit sequential optical spikes, presented in Fig. 6C, demonstrate the potential of OFETs photosynapses for in-sensor RC. Grouping OFETs into a pixel array, we propose using the in-sensor reservoir for motion recognition, as illustrated in Fig. 7A. The reservoir maps complex inputs to reservoir states for feature extraction and the linear output layer uses these features for classification. RC has the advantage of minimal training; specifically, the input and reservoir layers remain static, while only the weight of the linear output layer needs to be trained. This renders it well suited for edge learning applications.

Fig. 7. In-sensor RC system for basketball motion recognition.

(A) Schematic illustration of the RC system for classifying the basketball motion. (B) The photonic synapse array for mapping spatiotemporal vision information. Each shooting event is divided into five sequential frames (F_t4, F_t3, F_t2, F_t1, and F_t0), representing five distinct moments during ball movement. (C) Schematic illustration of the trajectory of basketball in five different states. (D) Confusion matrix for categorizing various basketball shooting actions. (E) Performance of the RC network over epochs. The blue and pink dots correspond to the classification accuracy of the software and our device-based RC system, respectively. (F) Dimensionality reduction of the reservoir outputs using linear discriminant analysis (LDA). (G) The number of training operations for RC, artificial neural network (ANN) without hidden layers and with one hidden layer, respectively, showing RC notably reduces the training complexity. (H) Initial/final conductance distributions before/after training.

The in-sensor reservoir captured basketball shooting trajectories using a customized dataset comprising five representative classes: up, down, left, right, and middle, shown in Fig. 7C. To integrate spatial and temporal information from successive frames into compact representations, we divided each shooting into five sequential frames (F_t4, F_t3, F_t2, F_t1, and F_t0), representing five distinct moments during ball movement, as shown in Fig. 7B. For each pixel, its temporal evolution is reduced to a 5-bit sequence, s(t), consisting of 32 possible binary vectors encoded as light intensities. The photosynapses of the pixel array, where individual transistors produce features (conductance) corresponding to one of the 32 unique illumination patterns (shown in Fig. 6C). By doing so, temporal pixel evolutions are embedded into the conductance of individual phototransistors, which are concatenated to represent the temporal feature of the video.

The readout map is a trainable classification head for the extracted features by the photosynapse, which is a fully connected layer here. As shown in Fig. 7D, the confusion matrix for categorizing various basketball shootings exhibited dominant diagonal elements, indicating high accuracy within each category. Our training outcomes are detailed in Fig. 7E, with an overall simulated accuracy of 95.56% based on electrically characterized device behaviors as depicted by red dots, matching the performance of the software baseline using the one-layer fully connected neural network (blue dots). Noise experiments further confirm that the 32 states are clearly distinguishable (fig. S17). Figure 7H displays the evolution of both initial and final weight distributions throughout the readout layer. In Fig. 7F, we present the two-dimensional clustering of OFET photosynapse extracted feature vectors using linear discriminant analysis (LDA) for dimensionality reduction. In this representation, distinct shooting states are denoted by spheres of varying colors, while the “x” and “y” axes correspond to the new two-dimensional feature vector after dimensionality reduction. The figure demonstrates clear differentiation among the five shooting states. These features undergo effective nonlinear transformations owing to the STMs of the PSBMA-based OFETs. Last, Fig. 7G compares the number of training parameters required for a RC system, a single-layer fully connected neural network, and a two-layer fully connected neural network. We observe that the number of parameters demanded for a deep fully connected network grows rapidly, while this quantity remains fixed for the RC system. This highlights the advantages of RC for cost-effective real-time edge learning, given its capacity for maintaining manageable complexity even when depth increases.

DISCUSSION

To address the challenges associated with conventional solid electrolyte dielectrics–based artificial synapses, such as high-power consumption and poor responses to light, we proposed a previously unreported electrolyte design strategy. This strategy involves covalently connecting the cation and anion with a flexible alkyl chain (i.e., PSBMA) to eliminate the long-range migration of free ions during polarization while maintaining well-defined polarization-relaxation behaviors under different applied presynaptic inputs. The resulting polyzwitterions exhibit a high capacitance (up to 2000 nF cm⁻² @20 Hz) and permittivity (k ~ 165), enabling low operation voltage and low power consumption (~250 aJ per spike) for the synaptic transistors. In addition, PSBMA induces a pronounced photoresponse of the transistor via an interfacial trapping strategy. Furthermore, the transistor exhibits unprecedently rich optoelectrical synaptic behaviors, including EPSC, PPF, STP, LTP, STM, and LTM, making it capable for image sensing, memorization, and real-time preprocessing. Leveraging these behaviors, this artificial synapse serves as an in-sensor reservoir, which is capable of perceiving the trajectory of a moving basketball with an accuracy of 95.56%. Furthermore, because of its excellent water solubility, the polyzwitterion-based flexible devices can degrade upon demand, eliminating the possibility of e-waste. This work represents an example of opto-neuromorphic computing based on electrolyte dielectrics–based transistor and provides a design for flexible, eco-friendly, and ultralow-power artificial synapses, paving the way for highly efficient neuromorphic computing.

MATERIALS AND METHODS

Simulation of the in-sensor RC system for basketball motion recognition

Initially, 25 shots were taken for each class, resulting in a total of 125 records captured via video footage. We selected five consecutive frames from each recording to create a dataset, and each frame was resized to 150 × 150 pixels before being binarized. To form the dataset for the simulated in-sensor RC system, we merged the five frames of each shot in chronological order to create a 22,500 × 5 matrix, where the pulse sequence of each row corresponds to one of the 32 illumination patterns (“00000” to “11111”). The in-sensor RC system consists of 22,500 optically operated memory devices serving as the reservoir nodes, with the conductance programming of these memory devices tuned on the basis of experimental measurements obtained from individual devices. We used a fully connected readout layer with 22,500 inputs (corresponding to the reservoir nodes) and five output neurons representing the different basketball classes (“up”, “down”, “left”, “right”, and “middle”). The readout layer was implemented via software, and we trained the model by minimizing categorical cross-entropy loss using gradient descent optimized with the Adam algorithm (with an initial learning rate set at 0.0001). The total number of training epochs was set to 150. Last, we divided the data into training and test sets, with 80 groups reserved for training and 45 for evaluating the model’s performance.

Supplementary Materials

This PDF file includes:

Supplementary Text

Figs. S1 to S21

References