NicePaper HERESubmitted on 21 Jul 2024]Few-Shot Transfer...

Nice

Paper HERE


Submitted on 21 Jul 2024]

Few-Shot Transfer Learning for Individualized Braking Intent Detection on Neuromorphic Hardware

Nathan Lutes, Venkata Sriram Siddhardh Nadendla, K. Krishnamurthy

Objective: This work explores use of a few-shot transfer learning method to train and implement a convolutional spiking neural network (CSNN) on a BrainChip Akida AKD1000 neuromorphic system-on-chip for developing individual-level, instead of traditionally used group-level, models using electroencephalographic data. The efficacy of the method is studied on an advanced driver assist system related task of predicting braking intention. Main Results: Efficacy of the above methodology to develop individual specific braking intention predictive models by rapidly adapting the group-level model in as few as three training epochs while achieving at least 90% accuracy, true positive rate and true negative rate is presented. Further, results show an energy reduction of over 97% with only a 1.3x increase in latency when using the Akida AKD1000 processor for network inference compared to an Intel Xeon CPU. Similar results were obtained in a subsequent ablation study using a subset of five out of 19 channels.

Significance: Especially relevant to real-time applications, this work presents an energy-efficient, few-shot transfer learning method that is implemented on a neuromorphic processor capable of training a CSNN as new data becomes available, operating conditions change, or to customize group-level models to yield personalized models unique to each individual.

5. Conclusion
The results show that the methodology presented was effective to develop individual-level models deployed on a state-of-the-art neuromorphic processor with predictive abilities for ADAS relevant tasks, specifically braking intent detection.

This study explored a novel application of deep SNNs to the field of ADAS using a neuromorphic processor by creating and validating individual-level braking intent classification models with data from three experiments involving pseudo-realistic conditions. These conditions included cognitive atrophy through physical fatigue and real-time distraction
and providing braking imperatives via commonly encountered visual stimulus of traffic lights. The method presented demonstrates that individual-level models could be quickly created with a small amount of data, achieving greater than 90% scores across all three classification performance metrics in a few shots (three epochs) on average for both the ACS and FCAS. This demonstrated the efficacy of the method for different participants operating under non-ideal conditions and using realistic driving cues and further suggests that a reduced data acquisition scheme might be feasible in the field.
Furthermore, the applicability to energy-constrained systems was demonstrated through comparison of the inference energy consumed with a very powerful CPU in which the Akida processor offered energy savings of 97% or greater. The Akida processor was also shown to be competitive in inference latency compared to the CPU. Future work could
include implementation of the method presented on a larger number of participants, other neuromorphic hardware, different driving scenarios, and in real-world scenarios where individual-level models are created by refining previously developed group-level models in real time.