Towards a Trade-off Between Accuracy and Computational Cost for Embedded Systems: A Tactile Sensing System for Object Classification

Amin, Youssef; Gianoglio, Christian; Valle, Maurizio

doi:10.1007/978-3-031-16281-7_15

Cited by 2 publications

(2 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The method works by maintaining the non-differentiable step function during the inference of the RSNN (forward pass with trained parameters) and replaces it with a differentiable function to compute its gradient during the backward pass and update the learnable weights of the network (W ij & V ij ) using the chain rule [53]. We used the gradient (partial derivative) of the fast sigmoid function σ(x) (7) during the backward pass in this work:…”

Section: A Encodingmentioning

confidence: 99%

“…These systems have important applications in prosthetics [3] [4] and robotics [5]. This electronic skin (e-skin) endows the robots with cutting-edge abilities to augment interaction with their surroundings such as perceiving texture [6], stiffness [7], and shape [8], [9]. Tremendous efforts have been made to establish an adequate artificial replication of the human skin behaviour [10]- [12].…”

Section: Introduction a Motivationmentioning

confidence: 99%

See 1 more Smart Citation

Neuromorphic Tactile Sensing System for Textural Features Classification

Al Haj Ali,

Abbass,

Gianoglio

et al. 2024

IEEE Sensors J.

View full text Add to dashboard Cite

Artificial tactile sensing systems have gained significant attention in recent years due to their potential to enhance human-machine interaction. Numerous initiatives have been introduced to shift the computational paradigms of these systems towards a more biologically inspired approach, by incorporating neuromorphic computing methods. Despite the significant advances made by these systems, dependence on complex offline methods for classification (i.e. hand-crafted encoding features), remains a limitation for their real-time deployment. In this work, we present a neuromorphic tactile PVDF-based sensing system for textural features classification, that employs raw signals directly for classification. We first converted raw signals into spikes and then trained recurrent spiking neural networks (RSNNs) using Backpropagation Through Time (BPTT) with surrogate gradients to perform classification. We proposed an optimization method based on tuning the refractory period of the encoding neurons, to explore a potential trade-off between the computational cost and classification accuracy of the RSNN. The proposed method effectively identified two RSNNs with refractory period configurations that achieved a tradeoff between the two evaluation metrics. Following this, we reduced the inference time steps of the selected RSNN during inference using a rate-coding based method. This method succeeded in saving around 26.6% out of the total original time steps. In summary, the proposed system paves the way for establishing an end-to-end neuromorphic approach for tactile textural features classification, through deploying the selected RSNNs on a dedicated neuromorphic hardware device for real-time inferences.

show abstract

Section: A Encodingmentioning

confidence: 99%