Supervised local training of spiking neural networks on digital neuromorphic hardware

The FOX research group is part of the CRIStAL laboratory (University of Lille, CNRS), located in Lille, France. We focus on video analysis for human behavior understanding. Specifically, we develop spatio-temporal models of motions for tasks such as abnormal event detection, emotion recognition, and face alignment. We are also involved in IRCICA (CNRS), a research institute promoting multidisciplanary research. At IRCICA, we collaborate with computer scientists and experts in electronics engineering to create new models of neural networks that can be implemented on low-power hardware architectures. Recently, we designed state-of-the-art models for image recognition with single and multi-layer unsupervised spiking neural networks. We were among the first to succesfully apply unsupervised SNNs on modern datasets of computer vision. We also developed our own SNN simulator to support experiments with SNN on computer vision problems. Our work is published in major journals (Pattern Recognition, IEEE Trans. on Affective Computing) and conferences (NeurIPS, WACV, IJCNN) in the field.

The position is located in Lille, France. With over 110 000 students, the metropolitan area of Lille is one France's top education student cities. The European Doctoral College Lille Nord-Pas de Calais is headquartered in Lille Metropole and includes 3,000 PhD Doctorate students supported by university research laboratories. Lille has a convenient location in the European high-speed rail network. It lies on the Eurostar line to London (1:20 hour journey). The French TGV network also puts it only 1 hour from Paris, 35 mn from Brussels, and a short trips to other major centres in France such as Paris, Marseille and Lyon.

Artificial Neural Networks (ANNs) have gathered exponential attention across diverse domains in recent years [1]. However, ANN training suffers from high and inefficient energy consumption on modern computers based on the von Neumann architecture [2]. Spiking Neural Networks (SNNs) [3], when implemented on neuromorphic hardware [4, 5], have emerged as a promising solution to overcome the von Neumann bottleneck [2] and enable energy-efficient computing. However, direct training of SNNs on neuromorphic hardware faces a major constraint: implementing network-level communication is difficult and requires significant circuitry overhead [6]. As a result, the learning mechanisms should be local, i.e., with weight updates based only on the activity of the two neurons that the synapse connects.

Spike Timing-Dependent Plasticity (STDP) [7] is the most common learning rule for unsupervised training of SNNs. STDP is inspired by the principal form of plasticity observed in biological synapses [8] and is particularly attractive for its local properties, enabling on-chip implementation. Unsupervised feature learning with STDP has been extensively studied in the literature, particularly for image recognition tasks [9, 10, 11]. STDP can also be adapted for supervised learning by incorporating a third factor, taking the form of an error signal that is used to guide the STDP updates [12]. As a result, recent work enabled end-to-end SNNs to perform classification tasks by combining unsupervised STDP for feature extraction and supervised STDP for classification [13, 14]. The mentioned results were obtained using energy expensive simulation on CPUs. In order to achieve low energy footprint efforts should be considered for transferring the STDP-based models on neuromorphic chips. This is the main scope of this project.

SpiNNaker [15] is a digital neuromorphic architecture designed for simulating SNNs. Unlike various other neuromorphic hardware, SpiNNaker provides a framework to define the behavior of neurons and synapses, enabling users to develop custom models. Unsupervised STDP is already implemented on SpiNNaker, which helped previous projects hosted at IRCICA to perform on-chip training of SNNs. However, due to the lack of supervision, the performance of such models remains poor on classification tasks. Recently, a supervised STDP rule, Stabilized Supervised STDP (S2-STDP) [14], was introduced by the FOX team, demonstrating state-of-the-art performance on image classification tasks. While this rule was specifically designed with hardware implementation in mind, its evaluation so far has been limited to modern software architectures. As a result, it remains unclear whether S2-STDP can be effectively implemented on SpiNNaker, and what challenges may arise during this process. Additionally, its performance and behavior under hardware-specific constraints are yet to be investigated.

To address the shortcomings of unsupervised learning, supervised or modulated STDP approaches have been developed.

SpiNNaker has played a transformative role in implementing and exploring SNNs, particularly through STDP-based learning. While unsupervised STDP aligns closely with biological systems, its limitations in practical applications are evident. Supervised and modulated STDP approaches provide a solution by incorporating feedback and reinforcement mechanisms, thereby expanding the utility of SNNs in solving real-world problems.

In this project we expect to achieve supervised STDP implementations on dedicated SpiNNaker hardware in order to achieve both low energy consumption and efficient classification rates. More specifically, it is expected to :

1. Get familiar with SpiNNaker models of spiking neurons and synapses
2. Produce an implementation of the neuron and synapses models compatible with the S2-STDP learning mechanism
3. Propose and implement updates to the existing S2-STDP so that it relies on the SpiNNaker models only
4. Validate the updated model on some (pre-existing) experimental data.

This Project is supported by both FOX and EMERAUDE team. Initial effort has been deployed in exploring the challenges rose by using supervised STDP with SpiNNaker.

Experience in one or more of the following is a plus:

 • image processing, computer vision;

 • machine learning;

 • bio-inspired computing;

 • research methodology (literature review, experimentation…).

Candidates should have the following skills:

 • scientific writing;

 • programming (experience in C++ is a plus, but not mandatory).