Page 43 - EETimes Europe June 2021

EE|Times EUROPE 43
EE|Times EUROPE
EU Project Looks to Mimic Bio Neural-System Processing

solutions and allows synapses to remember
their past activities. Both are crucial for
enabling the implementation of next-
generation on-chip learning.

APPLICATIONS
The technology developed in the MeM-Scales
project will enable new solutions for the
internet of things. “Today, the only solutions
available for automatic learning and complex
data interpretation are based on a cloud-
computing paradigm where locally extracted
sensor data is transmitted by edge devices to
Figure 2: Multi PCM-trace concept 3 remote servers,” said Vianello. In the future
IoT, however, much computing volume will
be offloaded from central servers and dele-
seconds (spoken sentences) and at much said Vianello. This simple technique allows gated to small controllers and smart sensors
longer intervals (motor learning). precision to be recovered. directly where their services are needed.
“The most complicated part is that there Another possibility is to exploit resis- One major application domain where this
are still many unknowns about how brains tive memory variability to build stochastic is valid is in autonomous navigation and
exactly work,” said Vianello. “There is a great synapses that keep track of and store two movement of vehicles such as robots, drones,
deal of understanding, but we still have a long variables: the mean and the variance (i.e., and even cars. In this case, one could take
way to go to understand the exact encoding, the error bar) of the associated probabil- advantage of a heterogeneous collection of
processing, and decoding in the brain. One ity distribution, said Vianello. “Stochastic video cameras, radar sensors, and potentially
of the understandings we have is that the synapses will enable the design of Bayesian LiDAR sensors as well.
brain processes information in a multitude models; [these] are particularly adapted for “Another major application target
of timescales. And this is the very property the ‘small data’ world, which has a lot of domain is sensor-based health-care and
which we would like to exploit in the MeM- uncertainty,” she said. “We recently proposed lifestyle systems such as smart patches,
Scales project.” a machine-learning technique that exploits smart wristbands, smart glasses, and even
The MeM-Scales project aims to elevate resistive memory variability.” Vianello and smart shoes,” said Vianello. Here, too, “we
neuromorphic computing in microproces- her fellow researchers described the work in a can make use of sensory fusion by com-
sors, with close interactions among experts recent Nature Electronics article. 2 bining a heterogeneous set of sensors for
in nanoelectronic device engineering, circuit Vianello explained that we can imag- collecting information such as ECG, EMG,
and microprocessor design, manufacturing ine a series of spatial and temporal filters bio-impedance streams, and potentially
technology, and computer science. distributed with a variety of time and space also brain signals through EEG sensors and
“The idea is not to raise the computational constants as processing elements in the brain. neuro-probes.”
envelope but to co-develop a novel class of These elements can be passive RC filters that Artificial neural networks, software,
devices, circuits, and algorithms that repro- filter and integrate information, as well as and/or hardware systems that mimic the
duce multi-timescale processing of biological active elements that introduce non-linearity. functioning of neurons in the human brain
neural systems,” said Vianello. “These sys- “Low-power neuromorphic systems make use are at the heart of theoretical and practi-
tems can process real-world sensory signals of sparsity in time, only consuming energy cal developments in artificial intelligence.
efficiently and in real time, without increasing when an ‘event’ arrives at the input,” she There is still much to be learned about how
the computational envelope.” said. “These systems are not clocked and are a biological brain works, but researchers and
The goal of the project is to develop asynchronous and event-based.” Therefore, scientists, with a multidisciplinary approach,
devices, circuits, and algorithms to enable the systems “make use of the physics of the are increasingly able to understand how
both learning and inference at the edge. substrate to implement time constants [e.g., cognitive processes take place and, with
Vianello pointed out that the project is RC circuits].” innovative technological insights, are suc-
focused on event-based, or spiking, neural One challenge for the implementation, ceeding in producing systems with a high
network or neuromorphic applications in Vianello added, is “the large range of time level of emulation. ■
which diverse timescales are present and constants.”
required. “So we mostly aim at streaming The MeM-Scales goal is to develop compact REFERENCES
applications in which the input/sensor data non-conventional memories and devices 1 A. Tavanaei et al. Deep Learning in Spiking
is sampled in a near-continuous stream and with controllable retention time. “In other Neural Networks. Revised 2019.
information of the past has to be stored across words, we want to exploit the physics of the arxiv.org/abs/1804.08150
multiple time horizons,” said Vianello. memories and devices — not just use them as 2 T. Dalgaty et al. In situ learning using intrinsic
A spiking neural network aims to simulate conventional digital elements but [use them] memristor variability via Markov chain Monte
natural neural networks more accurately. In to implement neural dynamics,” said Vianello. Carlo sampling. Nature Electronics 4 151–161.
addition to the synaptic and neuronal states, Vianello also told us that as part of this 2021. go.nature.com/3ofw3wP
such a network incorporates the concept of project, a solution was proposed to exploit 3 Y. Demirag et al. YPCM-trace: Scalable Synaptic
time into its operational model. the drift behavior of phase-change memory Eligibility Traces with Resistivity Drift of Phase-
Analog circuits and resistive memories (PCM) devices to implement eligibility traces Change Materials. 2021. arxiv.org/abs/2102.07260
are noisy. The simplest technique to address (ETs) covering behavioral timescales. The
3
that is to average over space (populations of use of this new solution improves the effi- Maurizio Di Paolo Emilio is editor-in-chief of
neurons) or over time (calculated mean rate), ciency of the 10× area compared with existing Power Electronics News and EEWeb.

www.eetimes.eu | JUNE 2021