2. Research Themes
pervised (but backprop-based) Variational Auto-Encoder
(VAE) training [
20
]. For completeness,we considered two
Neural networks are said to be biologically inspired since supervised Hebbian learning variants (Supervised
Hebthey mimic the behavior of real neurons. However, sev- bian Classifiers – SHC, and Contrastive Hebbian Learning
eral processes in state-of-the-art neural networks, includ- – CHL), for training the final classification layer, which
ing Convolutional Neural Networks (CNNs), are far from were compared to Stochastic Gradient Descent (SGD)
the ones found in biological brains [
19
]. One relevant training. We also investigated hybrid learning
methoddiference is the training process. In state-of-the-art arti- ologies, where some network layers were trained
followifcial neural networks, the training process is based on ing the Hebbian approach, and others were trained by
backpropagation and Stochastic Gradient Descent (SGD) backprop. Fig. 1 shows an example of such hybrid neural
optimization. However, studies in neuroscience strongly network models. We tested our approaches on MNIST
suggest that this kind of processes does not occur in the [
21
], CIFAR10 and CIFAR100 [
22
], Tiny ImageNet [
23
],
biological brain [
19
]. Rather, learning methods based on and ImageNet [
24
] datasets. Our results suggest that
STDP or the Hebbian learning rule seem to be more plau- Hebbian learning is generally suitable for training early
sible, according to neuroscientists [
7
]. In [
12
] , we investi- feature extraction layers, or to retrain higher network
gated the use of the Hebbian learning rule when training layers in fewer training epochs than backprop.
Moreneural networks for image classification by proposing a over, our experiments show that Hebbian learning
outnovel weight update rule for shared kernels in CNNs. We performs VAE training, with HPCA performing generally
performed experiments using the CIFAR-10 dataset in better than HWTA. In [
15
] we further extended our
exwhich we employed Hebbian learning, along with SGD, periments to cover a broader spectrum of competitive
to train parts of the model or whole networks for the learning approaches, beyond HWTA, namely k-WTA and
task of image classification, and we discussed their per- soft-WTA strategies.
formance thoroughly considering both efectiveness and In [
9
] we considered even more biologically
realiseficiency aspects. tic models, focusing on the goal of adopting biological
In [
16
], we deepened our investigations on Hebbian neuronal cultures to solve AI tasks. Previous work has
learning strategies applied to DNN training. We con- shown that it was possible to train neuronal cultures on
sidered two unsupervised learning approaches, Hebbian MEA devices, to recognize very simple patterns.
HowWinner-Takes-All (HWTA) and Hebbian Principal Com- ever, this work was mainly focused to demonstrate that
ponent Analysis (HPCA). The Hebbian learning rules it was possible to induce plasticity in cultures, rather
were used to train the layers of a CNN in order to extract than performing a rigorous assessment of their pattern
features that were then used for classification, without recognition performance. In our paper, we addressed
requiring backpropagation (backprop). We performed this gap by developing a methodology that allowed us to
experimental comparisons with state-of-the-art unsu- assess the performance of neuronal cultures on a
learning task. Specifically, we proposed a digital model of the eficient and sustainable alternatives for complex model
real cultured neuronal networks; we identified biolog- training.
ically plausible simulation parameters that allowed us In the second scenario, in [
9
], we developed a simulator
to reliably reproduce the behavior of real cultures; we of neuronal cultures on MEA devices, which is available
used the simulated culture to perform handwritten digit online 2. This is part of a broader ongoing project, aimed
recognition and rigorously evaluate its performance; we at using biological neuronal cultures for solving AI tasks,
also showed that it is possible to find improved simula- with the goal of providing more sustainable solutions for
tion parameters for the specific task, which can guide the AI. We compared the behavior of the simulated culture
creation of real cultures. with that of biological cultures, tuning the simulation
parameters to make the simulated results as close as
possible to the real-world data. We validated the simulated
3. Applications culture using digit recognition as a test case. We also
found that, by appropriately modifying the simulation
paOur research theme on bio-inspired machine learning rameters, it was possible to further improve performance.
found application in two real-world contexts: 1) learning We showed that, through simulation, it is possible to
with scarce data, and 2) learning with biological neuronal obtain insights on the parameters and properties (such
cultures. as strength and range of excitatory and inhibitory
con
The first scenario is promising in the direction of Sus- nections) that a neuronal culture should have in order
tainable AI, because heavy contributions to the energy perform well at a given task. These insights can then
footprint of machine learning are given by complex train- be used by neuroscientists in order to develop biological
ing procedures which require to process huge amounts networks, by means of modern cultivation techniques,
of data. On the other hand, biological systems appear to with the desired properties.
be able to learn from little experience. In this perspec- It was observed that simulated neurons were able to
tive, taking inspiration from biological systems holds the develop feature extractors, encoded in their weights, that
promise to revolutionize learning algorithms towards are reminiscent of edges and shapes found in the patterns
higher sample-eficiency, from the MNIST [
21
] digit recognition dataset used
dur
In [
13, 14, 17
] we proposed to address the issue of sam- ing training. Fig. 2 shows a visualization of such feature
ple eficiency, in CNNs, with a semi-supervised training extractors.
strategy that combines Hebbian learning with gradient
descent: all internal layers (both convolutional and fully
connected) were pre-trained using an unsupervised ap- 4. Conclusions and Future Work
proach based on Hebbian learning, and the last fully
connected layer (the classification layer) was trained us- In conclusion, bio-inspired learning approaches represent
ing Stochastic Gradient Descent (SGD). In fact, as Heb- a promising direction of future research towards more
bian learning is an unsupervised learning method, its sustainable AI systems.
potential lies in the possibility of training the internal Future challenges in the field of bio-inspired learning
layers of a CNN without labels. Only the final fully con- will involve the practical realization of biological devices
nected layer has to be trained with labeled examples. to carry out AI computations at scale. In this perspective,
We realized a machine learning module implementing the use of MEA devices seems promising, but also costly.
this strategy, which is available online 1. We performed On the other hand, we are exploring also the
possibilexperiments on various object recognition datasets, in ity to realize the interfacing with biological cultures via
diferent regimes of sample eficiency, comparing our optogenetics, using LED screens and photodetectors as
semi-supervised (Hebbian for internal layers + SGD for interfaces with the neuronal cultures.
the final fully connected layer) approach with end-to-end
supervised backprop training, and with semi-supervised
learning based on VAEs. The results showed that, in Acknowledgments
regimes where the number of available labeled samples
was low, our semi-supervised approach outperformed
the other approaches in almost all the cases. Further
work on Hebbian algorithms [
18
] allowed us to obtain
an extreme performance improvement, up to 50 times in
training speed, by leveraging eficient GPU computations.
This highlights the promises of bio-inspired solutions as
This work was partially supported by:
Tuscany Health Ecosystem (THE) Project (CUP
I53C22000780001), funded by the National Recovery
and Resilience Plan (NRPP), within the NextGeneration
Europe (NGEU) Program
PNRR - M4C2 - Investimento 1.3, Partenariato Esteso
PE00000013 - ”FAIR - Future Artificial Intelligence
1https://github.com/GabrieleLagani/HebbianLearning
2https://github.com/GabrieleLagani/SpikingGrid
Research” - Spoke 1 ”Human-centered AI”, funded by
the European Commission under the NextGeneration
EU programme.
INAROS (INtelligenza ARtificiale per il mOnitoraggio
e Supporto agli anziani) project co-funded by Tuscany
Region POR FSE CUP B53D21008060008.