An Image Memory Logic Unit Inspired by Human Retina

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

1810 IEEE ELECTRON DEVICE LETTERS, VOL. 42, NO.

12, DECEMBER 2021

An Image Memory Logic Unit Inspired


by Human Retina
Wen Du , Caihong Li, Ziyang Long , Yixuan Huang, Lingzhi Luo, Jihua Zou,
and Jiang Wu , Senior Member, IEEE

Abstract — Human-inspired neuromorphic computing high computing efficiency and low power consumption [1].
system breaks the limit of “von Neumann bottleneck” and However, the physical separation of memory and process-
improves the computing efficiency. However, the processing ing units, which is imposed by conventional von Neumann
of a large amount of non-structural visual data results in
a considerable waste of computing resources. In human computing architecture, results in the communication bottle-
visual system, visual information is preprocessed by the neck between the logic and memory [2]. Human-inspired
retina before transmitted to the visual cortex, which effec- neuromorphic computing can perform complex tasks due
tively reduces data redundancy. Therefore, multifunctional to simulating brain at the physical level [3]. This kind of
intelligent devices which integrate sensing-memory- in-memory computing breaks the limit of the well-known
processing functions are desired for the artificial visual
system. Herein, an image memory logic unit (IMLU) inspired “memory wall” or “von Neumann bottleneck” [4]. A variety
by human retina is proposed. The IMLU is formed by a of in-memory computing devices have been proposed [5]–[7],
simple two-terminal photodetector and comparator module. and the perception of sight, olfaction, taste and touch has been
The photodetector is based on the Mox W1−x S2 alloy with achieved [8]–[10]. Among them, visual perception is particu-
persistent photo conductivity effect. In addition, the IMLU larly important [11], but the processing of non-structural visual
exhibits capabilities of in situ sensing, memory and logical
operations (“AND”, “OR”, and “XOR”). Furthermore, data requires a vast data storage and computation to obtain
the IMLU array enables image processing functions such meaningful information [12]–[14]. Therefore, multifunctional
as dilation, erosion and difference recognition, which devices which integrate sensing, memory and processing func-
shows great potential to provide simplified information tions are desired. It is also highly desired for neuromorphic
processing for subsequent tasks in machine vision. computing to perform logical operations with optical signals
Index Terms — Human-inspired, logical operation, image because of the low crosstalk and high bandwidth of light [15].
processing, sensing-memory-processing system. In human retina, ganglion cells carry out specific calcula-
tions on the visual scene to extract visual features of interest,
I. I NTRODUCTION and then transmit these processed sets to the brain [16].

T HE development of artificial intelligence (AI) technology


and Internet-of-Things (IoT) makes it essential to achieve
Inspired by this, an integrated sensing-memory-processing
system based on the image memory logical unit (IMLU) is
presented in this work. Logical operations (AND, OR, and
Manuscript received September 15, 2021; revised October 3, 2021;
accepted October 3, 2021. Date of publication October 5, 2021; date
XOR) on two sequenced optical singles can be performed by
of current version November 24, 2021. This work was supported in changing the reference voltage of a comparator module. The
part by the National Natural Science Foundation of China under Grant output state of the IMLU is stored in the photodetector in
61974014, in part by the Fundamental Research Funds for the Central
Universities under Grant ZYGX2019Z018, in part by the Innovation
the form of conductance. Furthermore, the image processing
Group Project of Sichuan Province under Grant 20CXTD0090, in part by functions of the IMLU array are conceptually demonstrated
the University of Electronic Science and Technology of China (UESTC) including dilation, erosion and difference detection.
Shared Research Facilities of Electromagnetic Wave and Matter Interac-
tion under Grant Y0301901290100201, and in part by the Open Project
of Key Laboratory of Infrared Imaging Materials and Detectors, Chinese II. D EVICE FABRICATION
Academy of Sciences. The review of this letter was arranged by Editor
B. Govoreanu. (Corresponding author: Jiang Wu.) A. Preparation of the Material
Wen Du, Caihong Li, Yixuan Huang, Lingzhi Luo, and Jihua Zou are The as-grown Mo1−x Wx S2 (0≤x≤1) alloy was prepared
with the Institute of Fundamental and Frontier Sciences, University of
Electronic Science and Technology of China, Chengdu 610054, China. using CVD method at atmosphere pressure. The Si/SiO2
Ziyang Long is with the Glasgow College, University of Electronic substrates were cleaned using acetone and isopropanol, respec-
Science and Technology of China, Chengdu 610054, China. tively. First, a substrate was placed in a quartz boat to hold
Jiang Wu is with the Institute of Fundamental and Frontier Sciences,
University of Electronic Science and Technology of China, Chengdu the precursor powder. The WO3 powder and NaCl powder
610054, China, also with the State Key Laboratory of Electronic Thin with the mixture ratio of 10:1 was placed on this precursor
Films and Integrated Devices, University of Electronic Science and substrate, and the MoO3 powder (half weight of WO3 ) and
Technology of China, Chengdu 610065, China, and also with the Key
Laboratory of Infrared Imaging Materials and Detectors, Shanghai Insti- a small amount of NaOH powder was placed on the other
tute of Technical Physics, Chinese Academy of Sciences, Shanghai side. Second, a growth substrate was put over the previous
200083, China (e-mail: jiangwu@uestc.edu.cn). substrate. Third, the sulfur powder (0.5 g) was placed in
Color versions of one or more figures in this letter are available at
https://doi.org/10.1109/LED.2021.3118053. another quartz boat in the upstream of the tube. The tube center
Digital Object Identifier 10.1109/LED.2021.3118053 was ramped up to 750 ◦ C in 15 min and kept for 5 min with

0741-3106 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Vardhaman College of Engineering. Downloaded on December 09,2021 at 08:51:39 UTC from IEEE Xplore. Restrictions apply.
DU et al.: IMLU INSPIRED BY HUMAN RETINA 1811

Fig. 1. Schematic diagram of artificial optical synapse and material Fig. 2. Optical characteristics of the photodetector. (a) Currents of the
measurements. (a) Schematic diagram of human visual perception. device triggered by an optical pulse with different light intensity. The unit
(b) Schematic diagram of device. (c) Schematic illustration of the photo- of the light intensity is mW/cm2 . The pulse duration is 0.2 s. (b) The
carriers trapping and de-trapping process in Mox W1−x S2 alloy. dependence of light intensity on the current under different illumination
time. (c) Paired-pulse facilitation (PPF) of the device triggered by two
the Ar of 45 sccm. The sulfur powder was heated to 220 ◦ C optical pulses. The unit of the light intensity is mW/cm2 . The pulse
duration is 0.2 s and the interval is 2 s. (d) PPF of the device triggered
when the tube center was 670 ◦ C. After that, the tube center by two optical pulses with different interval time.
reached 875 ◦ C in 5 min and maintained for 5 min with Ar/H2
of 40/5 sccm. Finally, the tube was removed from the furnace triggered by a short optical pulse of 0.2 s, decays to initial
and was cooled down to room temperature naturally. state slowly. This is due to the abundant defects (mainly sulfur
vacancies) in the Mox W1−x S2 that prevent the recombination
B. Preparation and Measurement of the Photodetector of photocarriers. A higher current is induced by an increased
The electrodes (Ti/Au) were prepared by photolithography light intensity due to the generation of more photocarriers.
and electron beam evaporation. The optoelectronic charac- The current recover process is divided into rapid process and
teristics were measured in a probe station connected to a slow process. The rapid process occurs because of the rapid
semiconductor parameter analyzer (KEITHLEY 4200). The de-trapping process of the trapped carriers by the defects and
405 nm laser is used as the light source. the fast relaxation of the untrapped carriers, whereas the slow
process is due to the slow de-trapping process [14], [15], [22].
The fast process is attribute to the de-trapping of carriers near
III. R ESULTS AND D ISCUSSION
the interface and in shallow level defects, and the slow process
A. Device Structure and Optoelectronic Characteristics
originate from the de-trapping of carriers far away from
The retina of human visual system can store and preprocess the interface and in deep level defects. Similarly, the spike
visual information while perceiving light (Fig. 1(a)). The response currents are further increased with the raised opti-
bionic image sensor inspired by this can reduce the redundant cal pulse duration (Fig. 2(b)). In addition, the paired-pulse
information and improve the computational efficiency. facilitation (PPF) of the device is examined by applying two
Transition metal dichalcogenides (TMDs) are effective sequential optical pulses. As shown in Fig. 2(c), the effect of
candidates in forming the charge-trapping layer of bionic light intensity on the peak current is similar to that triggered by
photoelectric devices [17]–[20]. In addition, defect engineering single optical spike. When the interval time is smaller than the
has been utilized to create the synaptic behavior [14], [21]. photocarriers recombination time, the current after the second
Since alloying is a common method to introduce defects in illumination will be higher than the first. As the interval time
TMDs, a large area alloyed Mox W1−x S2 film is prepared as increases, the number of recombined photocarriers increases
the photosensitive layer of a photodetector. The structure of correspondingly, so that the second photocurrent decreases
the device is depicted in Fig. 1(b). As shown in Fig. 1(c), the (Fig. 2(d)).
defects of sulfur vacancies trap some photocarriers, and the
photocarriers are de-trapped gradually after the illumination is
stopped. The slowly de-trapping process enables the persistent B. Sensing-Memory-Processing Procedure of IMLU
photoconductivity effect of the Mox W1−x S2 alloy. The devices with persistent photoconductivity have many
Fig. 2 shows the optoelectronic characteristics of the pho- bionic applications because they can sense and store
todetector based on Mox W1−x S2 . The curves of the current information in-situ [23], [24]. Here, the proposed IMLU
over time after illumination with different light intensity are performs in-situ sensing, memory and preprocess of two
shown in Fig. 2(a). The current of the device, after being sequenced optical signals. The light-off state is defined as the

Authorized licensed use limited to: Vardhaman College of Engineering. Downloaded on December 09,2021 at 08:51:39 UTC from IEEE Xplore. Restrictions apply.
1812 IEEE ELECTRON DEVICE LETTERS, VOL. 42, NO. 12, DECEMBER 2021

Fig. 4. Simulation results of the circuits and logical operations of the


IMLU. (a) Schematic diagram of single pixel and (left) the block circuit
diagram of the comparator module (right) for “AND”, “OR” and “XOR”
operations. (b)-(c) The input voltages (dash lines) and the output voltages
(pink solid lines) after “AND” / ”OR” operation (b), and “XOR” operation
(c) to the input images in Figure 3. The corresponding inset images are
the output states.

Fig. 3. Sensing-memory-logical operations processes of the four pixels.


(a) Current curves of the four pixels under four different light inputs. The
duration is 0.35s and the interval is 2 s. (b) Forgetting curves of the four
pixels. The dash lines are the reference currents. (c).

“0” state whereas the light-on state is defined as the “1” state.
For the four pixels (P1, P2, P3, P4), the input signal states are
(0, 0), (0, 1), (1, 0) and (1, 1), respectively.
Fig. 5. Logical operations of images of actual scene. The white pixels
Fig. 3 depicts the sensing and memory characteristics of the represent the “1” state, and the black pixels represent the “0” state.
alloy-based photodetector. Fig. 3(a) shows the current changes
of the four pixels under two different input states. The light the output state is “0”. The reference voltage for the “AND”
duration of 0.35 s can be interpreted as the time of information operation and “OR” operation is converted from Ref-1 and
reading. The forgetting (relaxation) curves of the four pixels Ref-2 in Fig. 3(b), respectively. For the “XOR” operation,
after the second information reading are shown in Fig. 3(b). when Vref−2 <Vin <Vref−1 , the output state is “1”; otherwise,
The current can be divided into three ranges based on two the output state is “0”. Fig. 4(b)-(c) describe the corresponding
reference current curves. The reference current 1 (2) is the converted voltages and output voltages of four pixels, and the
average of the current of P1 (P2) and P3 (P4). It is worth high output voltage means the output state is “1”.
noting that the indistinguishable P2 and P3 can be resolved In order to demonstrate the application potential of the
by adding input sequences [25]. However, this situation does IMLU, the functions of large array based on IMLU are simu-
not affect the implementation of subsequent functions. In order lated to deal with actual scene images (Fig. 5). The simulations
to visually demonstrate the sensing and memory processes of are implemented using Python 3.7 and Open Source Com-
the four pixels, the evolution of the current map is illustrated puter Vision Library. As shown in Fig. 5, the inputted binary
in Fig. 3(c). The four pixels in the initial state read the matrix images have small difference that human vision can hardly
of (0, 0, 1, 1), which is the sensing process, and store it. Then, distinguish. The “AND” (“OR”) operation increases (reduces)
the second matrix of (0, 1, 0, 1) is inputted after an interval the different pixel information, whereas the “XOR” operation
of 1.8 s. highlights the changing information, such as moving vehicles
The currents of the four pixels will be different because and pedestrians.
of the influence of the first input, and the logical operations
of the two input states can be accomplished through a sim- IV. C ONCLUSION
ple circuit design of each pixel (Fig. 4(a)). The currents of In summary, an IMLU based on large area Mox W1−x S2
the photodetector are converted to voltages for comparison alloys is proposed. The IMLU, which performs functions of
with reference voltages. A LED is used to show the digital sensing, memory and logical operations, provides great poten-
output state visually. If the output state is “1”, the LED tial in the intelligent image sensors by outputting meaningful
is on, otherwise the LED is off. For the “AND” / “OR” information to improve the processing efficiency of subsequent
operations, when Vin >Vref , the output state is “1”; otherwise, tasks in machine vision.

Authorized licensed use limited to: Vardhaman College of Engineering. Downloaded on December 09,2021 at 08:51:39 UTC from IEEE Xplore. Restrictions apply.
DU et al.: IMLU INSPIRED BY HUMAN RETINA 1813

R EFERENCES [14] T. Ahmed, M. Tahir, M. X. Low, Y. Ren, S. A. Tawfik, E. L. H. Mayes,


S. Kuriakose, S. Nawaz, M. J. S. Spencer, H. Chen, M. Bhaskaran,
[1] X. Feng, S. Li, S. L. Wong, S. Tong, L. Chen, P. Zhang, L. Wang,
S. Sriram, and S. Walia, “Fully light-controlled memory and neu-
X. Fong, D. Chi, and K.-W. Ang, “Self-selective multi-terminal mem-
romorphic computation in layered black phosphorus,” Adv. Mater.,
transistor crossbar array for in-memory computing,” ACS Nano, vol. 15,
vol. 33, no. 10, Nov. 2020, Art. no. 2004207, doi: 10.1002/
no. 1, pp. 1764–1774, Jan. 2021, doi: 10.1021/acsnano.0c09441.
adma.202004207.
[2] X. Xiong, J. Kang, Q. Hu, C. Gu, T. Gao, X. Li, and Y. Wu,
[15] Y.-X. Hou, Y. Li, Z.-C. Zhang, J.-Q. Li, D.-H. Qi, X.-D. Chen,
“Reconfigurable logic-in-memory and multilingual artificial synapses
J.-J. Wang, B.-W. Yao, M.-X. Yu, T.-B. Lu, and J. Zhang, “Large-
based on 2D heterostructures,” Adv. Funct. Mater., vol. 30, no. 11,
scale and flexible optical synapses for neuromorphic computing
Mar. 2020, Art. no. 1909645, doi: 10.1002/adfm.201909645.
and integrated visible information sensing memory processing,” ACS
[3] L. Hu, J. Yang, J. Wang, P. Cheng, L. O. Chua, and F. Zhuge, “All- Nano, vol. 15, no. 1, pp. 1497–1508, Jan. 2021, doi: 10.1021/
optically controlled memristor for optoelectronic neuromorphic comput- acsnano.0c08921.
ing,” Adv. Funct. Mater., vol. 31, no. 4, Nov. 2020, Art. no. 2005582, [16] T. Gollisch and M. Meister, “Eye smarter than scientists
doi: 10.1002/adfm.202005582. believed: Neural computations in circuits of the retina,” Neuron,
[4] Z. Sun, J. Li, C. Liu, S. Yang, and F. Yan, “Trap-assisted charge vol. 65, no. 2, pp. 150–164, Jan. 2010, doi: 10.1016/j.neuron.
storage in titania nanocrystals toward optoelectronic nonvolatile mem- 2009.12.009.
ory,” Nano Lett., vol. 21, no. 1, pp. 723–730, Jan. 2021, doi: [17] X. Pan, T. Jin, J. Gao, C. Han, Y. Shi, and W. Chen, “Stimuli-
10.1021/acs.nanolett.0c04370. enabled artificial synapses for neuromorphic perception: Progress and
[5] M. Prezioso, F. Merrikh-Bayat, B. D. Hoskins, G. C. Adam, perspectives,” Small, vol. 16, no. 34, Aug. 2020, Art. no. 2001504, doi:
K. K. Likharev, and D. B. Strukov, “Training and operation of an inte- 10.1002/smll.202001504.
grated neuromorphic network based on metal-oxide memristors,” Nature, [18] T. Wang, J. Meng, Z. He, L. Chen, H. Zhu, Q. Sun, S. Ding,
vol. 521, no. 7550, pp. 61–64, May 2015, doi: 10.1038/nature14441. P. Zhou, and D. W. Zhang, “Ultralow power wearable heterosynapse with
[6] M.-K. Kim and J.-S. Lee, “Ferroelectric analog synaptic transis- photoelectric synergistic modulation,” Adv. Sci., vol. 7, no. 8, Apr. 2020,
tors,” Nano Lett., vol. 19, no. 3, pp. 2044–2050, Mar. 2019, doi: Art. no. 1903480, doi: 10.1002/advs.201903480.
10.1021/acs.nanolett.9b00180. [19] Y. Hu, M. Dai, W. Feng, X. Zhang, S. Zhang, B. Tan, H. Shang,
[7] S. R. Nandakumar, M. Le Gallo, C. Piveteau, V. Joshi, Y. Q. Fu, and P. Hu, “Monolayer hydrophilic MoS2 with strong
G. Mariani, I. Boybat, G. Karunaratne, R. Khaddam-Aljameh, charge trapping for atomically thin neuromorphic vision systems,” Mater.
U. Egger, A. Petropoulos, T. Antonakopoulos, B. Rajendran, Horizons, vol. 7, no. 12, pp. 3316–3324, Dec. 2020, doi: 10.1039/
A. Sebastian, and E. Eleftheriou, “Mixed-precision deep learning D0MH01472A.
based on computational memory,” Frontiers Neurosci., vol. 14, p. 406, [20] B. Yao, J. Li, X. Chen, M. Yu, Z. Zhang, Y. Li, T. Lu, and J. Zhang,
May 2020, doi: 10.3389/fnins.2020.00406. “Non-volatile electrolyte-gated transistors based on graphdiyne/MoS2
[8] S. Zhang, K. Guo, L. Sun, Y. Ni, L. Liu, W. Xu, L. Yang, and with robust stability for low-power neuromorphic computing and
W. Xu, “Selective release of different neurotransmitters emulated by logic-in-memory,” Adv. Funct. Mater., vol. 31, no. 25, Apr. 2021,
a p–i–n junction synaptic transistor for environment-responsive action Art. no. 2100069, doi: 10.1002/adfm.202100069.
control,” Adv. Mater., vol. 33, no. 10, Mar. 2021, Art. no. 2007350, doi: [21] T. Ahmed, S. Kuriakose, S. Abbas, M. J. S. Spencer, M. A. Rahman,
10.1002/adma.202007350. M. Tahir, Y. Lu, P. Sonar, V. Bansal, M. Bhaskaran, S. Sriram, and
[9] J. Yu, G. Gao, J. Huang, X. Yang, J. Han, H. Zhang, Y. Chen, C. Zhao, S. Walia, “Multifunctional optoelectronics via harnessing defects in lay-
Q. Sun, and Z. L. Wang, “Contact-electrification-activated artificial ered black phosphorus,” Adv. Funct. Mater., vol. 29, no. 39, Sep. 2019,
afferents at femtojoule energy,” Nature Commun., vol. 12, no. 1, p. 1581, Art. no. 1901991, doi: 10.1002/adfm.201901991.
Mar. 2021, doi: 10.1038/s41467-021-21890-1. [22] S. Seo, S.-H. Jo, S. Kim, J. Shim, S. Oh, J.-H. Kim, K. Heo, J.-W. Choi,
[10] T. Wang, H. Huang, X. Wang, and X. Guo, “An artificial olfactory C. Choi, S. Oh, D. Kuzum, H.-S.-P. Wong, and J.-H. Park, “Artificial
inference system based on memristive devices,” InfoMat, vol. 3, no. 7, optic-neural synapse for colored and color-mixed pattern recognition,”
pp. 804–813, Jul. 2021, doi: 10.1002/inf2.12196. Nature Commun., vol. 9, no. 1, p. 5106, Nov. 2018, doi: 10.1038/s41467-
[11] H. Li, X. Jiang, W. Ye, H. Zhang, L. Zhou, F. Zhang, D. She, Y. Zhou, 018-07572-5.
and S.-T. Han, “Fully photon modulated heterostructure for neuromor- [23] H. Jang, C. Liu, H. Hinton, M. Lee, H. Kim, M. Seol, H. Shin,
phic computing,” Nano Energy, vol. 65, Nov. 2019, Art. no. 104000, S. Park, and D. Ham, “An atomically thin optoelectronic machine vision
doi: 10.1016/j.nanoen.2019.104000. processor,” Adv. Mater., vol. 32, no. 36, Jul. 2020, Art. no. 2002431, doi:
[12] S. Wang, C.-Y. Wang, P. Wang, C. Wang, Z.-A. Li, C. Pan, 10.1002/adma.202002431.
Y. Dai, A. Gao, C. Liu, J. Liu, H. Yang, X. Liu, B. Cheng, K. Chen, [24] Z.-C. Zhang, Y. Li, J.-J. Wang, D.-H. Qi, B.-W. Yao,
Z. Wang, K. Watanabe, T. Taniguchi, S.-J. Liang, and F. Miao, M.-X. Yu, X.-D. Chen, and T.-B. Lu, “Synthesis of wafer-scale
“Networking retinomorphic sensor with memristive crossbar for brain- graphdiyne/graphene heterostructure for scalable neuromorphic
inspired visual perception,” Nat. Sci. Rev., vol. 8, no. 2, Feb. 2021, computing and artificial visual systems,” Nano Res., early access,
Art. no. nwaa172, doi: 10.1093/nsr/nwaa172. pp. 1–10, Mar. 2021, doi: 10.1007/s12274-021-3381-4.
[13] F. Zhou, Z. Zhou, J. Chen, T. H. Choy, J. Wang, N. Zhang, Z. Lin, [25] L. Sun, Z. Wang, J. Jiang, Y. Kim, B. Joo, S. Zheng, S. Lee,
S. Yu, J. Kang, H.-S.-P. Wong, and Y. Chai, “Optoelectronic resistive W. J. Yu, B.-S. Kong, and H. Yang, “In-sensor reservoir comput-
random access memory for neuromorphic vision sensors,” Nature Nan- ing for language learning via two-dimensional memristors,” Sci. Adv.,
otechnol., vol. 14, no. 8, pp. 776–782, Aug. 2019, doi: 10.1038/s41565- vol. 7, no. 20, May 2021, Art. no. eabg1455, doi: 10.1126/sciadv.
019-0501-3. abg1455.

Authorized licensed use limited to: Vardhaman College of Engineering. Downloaded on December 09,2021 at 08:51:39 UTC from IEEE Xplore. Restrictions apply.

You might also like