A Trainable Sequence Learner that Learns and Recognizes Two-Input Sequence Patterns Jan Hohenheim

2025-04-30 0 0 1.76MB 5 页 10玖币

侵权投诉

A Trainable Sequence Learner that Learns and

Recognizes Two-Input Sequence Patterns

Jan Hohenheim

University of Zurich

Zurich, Switzerland

jan@hohenheim.ch

Tommaso Stecconi

IBM Research GmbH

Zurich Research Laboratory

Work carried out at University of Zurich

Zurich, Switzerland

tec@zurich.ibm.com

Zhaoyu Devon Liu

CUHK BME

Chinese University of Hong Kong

Hong Kong, China

zhyliu@link.cuhk.edu.hk

Pietro Palopoli

D-ITET

ETH Zurich

Zurich, Switzerland

ppalopoli@student.ethz.ch

Abstract—We present two designs for an analog circuit that can

learn to detect a temporal sequence of two inputs. The training

phase is done by feeding the circuit with the desired sequence and,

after the training is completed, each time the trained sequence

is encountered again the circuit will emit a signal of correct

recognition. Sequences are in the order of tens of nanoseconds.

The ﬁrst design can reset the trained sequence on runtime but

assumes very strict timing of the inputs. The second design can

only be trained once but is lenient in the input’s timing.

Index Terms—sequence, learning, analog, design, circuit, coin-

cidence detector

I. INTRODUCTION

Sequence pattern recognition is both central to how our

brain works and important for many modern AI applications

such as speech recognition, speaker identiﬁcation, automatic

medical diagnosis or general classiﬁcation. Since the human

brain performs pattern learning and recognition with extreme

energy efﬁciency, parallelism, and relatively good speed, the

need to replicate these advantages in silica becomes apparent.

Some similar works already exist. Liu et al. [1] designed

a multi-terminal transistor that can behave as a sequence

detector by tuning the time delay between pulses fed into the

transistor terminals. However, this system cannot train itself

using certain input sequences like our brain.

As a starting point, we present an idea of how a simple

sequence pattern learning and recognition chip could look like

and provide two possible implementation schematics.

II. LEARNING ALGORITHM

Inputs are expected to be pulses coming from two different

sources that arrive in some regular interval, determined by

the design used. We will call them Signal A (or just A) and

Signal B (or just B). They are instantly delayed by the same

default amount of time. The goal is to make them overlap in

time (learning phase) so that later the correct recognition can

occur. This is done by treating A as a reference and shifting

B’s delay up or down to make it coincide with A inside the

circuit.

This report was produced in collaboration with the Institute of Neuroinfor-

matics (University of Zurich and ETH Zurich)

Fig. 1. Overview of the shared logic behind the circuits

To do so, It is necessary to analyze the extent of the delay

between the inputs, which is done by using a coincidence

detector inspired by the Jeffress model of sound localization

[2]. This model works by letting the inputs move through a

series of delaying nodes from opposite sides. They will meet

at a node determined by their temporal offset. If they meet in

the middle it means that they came in simultaneously. Since

our inputs are assumed to always come with some offset, a

meeting in the middle must mean that we successfully tweaked

B’s delay so that the training is ﬁnished.

Before this is the case, the inputs will meet at another node,

indicating how off from our goal we are. E.g. if we wish to

detect the sequence “AB”, B will come in after A, so the

signals will meet on one of the nodes on the right in ﬁgure 1.

meaning that we must aim to decrease B’s delay. This task is

carried out by a delay adjustment unit speciﬁc to the design

in question. In any case, the delay modiﬁed is a row of 8

tunable delay [3] units. Their delay is inversely proportional

to a shared bias voltage.

III. SEQUENCE LEARNER DESIGNS

A. Design A

1) Usage:

Main inputs Vin1 and Vin2

Auxiliary inputs Vreset to reset the learned sequence

Constant inputs Vdelay at 1.18 V

arXiv:2210.12193v1 [cs.NE] 21 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ATrainableSequenceLearnerthatLearnsandRecognizesTwo-InputSequencePatternsJanHohenheimUniversityofZurichZurich,Switzerlandjan@hohenheim.chTommasoStecconiIBMResearchGmbHZurichResearchLaboratoryWorkcarriedoutatUniversityofZurichZurich,Switzerlandtec@zurich.ibm.comZhaoyuDevonLiuCUHKBMEChineseUniversityo...

展开>> 收起<<

A Trainable Sequence Learner that Learns and Recognizes Two-Input Sequence Patterns Jan Hohenheim.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

A Trainable Sequence Learner that Learns and Recognizes Two-Input Sequence Patterns Jan Hohenheim

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: