Instrumentation & Measurement Magazine 25-9 - 23

Fig. 3. Time evolution of the ground reaction force signals for the front right leg: comparison between the normalized real
and predicted signal when the robot sensory system is fully operative and in presence of a fault in the joint sensors of the
front right leg (i.e., time window from 200 to 800 samples).
information at the level of the joint torque sensors. This architecture
can also be extended to include multiple read-out
maps, following the paradigm of neural reuse, to accomplish
other tasks such as terrain classification [12].
Sequence Learning
This second case of study reports a high-level approach to
action-oriented perception, inspired by the basic fly brain
structure, with the addition of accessory neural areas which
boost the overall computational structure capabilities, as experimentally
found in bees.
As reported above, KCs are sparsely activated by the
upcoming input at the level of the calyx and generate spatio-temporal
spiking activity on the MB-lobes. The relevant
dynamics needed to perform a behavioral task can be extracted
through extrinsic neurons used to transfer the decoded
information to the motor area. This pretty much resembles,
once again, the LSMs scheme, augmented with multiple
read-out maps, that when subject to learning are exploited
concurrently both for classification and motor learning purposes.
Although we have access to the whole connectome
through the fruit fly brain project (Fig. 1), we downscaled the
complexity of the neural network to reduce its size in view of
robotic applications.
The LSM structure described in Fig. 2 is adopted here with
the introduction of further elements inspired by more complex
insect brain structures, introduced in and shown in Fig. 4a. The
spiking neuron lattice, representing the KCs, elaborates the
December 2022
sensory information coming
from the input layer
(e.g., the antennal lobes
(AL)), creating a dynamic
map to be exploited through
the outer layers. The lateral
horn (LH) has a dominant
inhibitory role on the KCs
lattice activity after a given
time window, to guarantee
the processing of the newly
acquired stimuli. Different
from the previous case (i.e.,
GRFs estimation), where a
regression problem is addressed,
here a sequence
learning task is executed,
followed by a selection task,
based on the identification
of a specific sequence of previously
learned actions.
The complex internal
dynamics generated within
the LSM layer are extracted
and classified in terms of
periodic signals through a
supervised learning process
performed at the read-out
map level. The different frequencies associated with the LSM
output signals are detected through a pool of resonant neurons,
here modeled using the Morris-Lecar (ML) neuron. An
unsupervised growing mechanism is adopted to guarantee the
formation of new ML neurons, corresponding to new classes,
when needed.
In addition to this classification layer, a context layer is included
in the form of a pool of spiking neurons, topologically
organized in concentric rings, mimicking the arrangement of
the parallel fibers along the MB lobes in bees [13]. Each ring
of the context layer is stimulated when specific input is classified
through the read-out map: the neural activity diffuses
from the inner to the outer rings through time. The presence of
lateral inhibition among neurons belonging to the same ring
generates competition, filtering out external disturbances.
Feedback connections from the context layer to the resonant
neurons are subject to learning. A pool of end sequence neurons
is also employed to reset the context layer activity, either
when a rewarding signal is provided, or when a time-out event
is triggered. The output layer selects a behavior for the robot,
based on the winning (i.e., maximally activated) neuron. At
every presentation of a new stimulus, the winning neuron in
the outermost ring of the context layer and the previous winner
neuron are subject to an STDP learning step to reinforce
their connection weights and contribute to the formation of
sequences.
The context neurons and the resonant neurons compete
to generate a motor command that could be either directly
IEEE Instrumentation & Measurement Magazine
23

Instrumentation & Measurement Magazine 25-9

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-9