On-chip few-shot learning of heartbeat anomaly on a new neural processing chip

We are looking for an enthusiastic and self-driven Master’s student, to perform few-shot learning on a mixed-signal neuromorphic chip, which will be done for the first time, resulting in a high-impact publication.

Project description
As we are moving towards an ever-increased connected world, the amount of data and energy consumption for processing it is exponentially increasing. Moreover, the privacy of users is becoming a concern, specifically in wearable applications, as their bio-markers will be collected for personalization of their experience. To solve this, we have designed a learning chip, called ALIVE, which can receive data in real time and learn from it on chip locally, and in a power efficient fashion (< 1mW).

To benchmark with ALIVE, we have chose the problem of classifying heartbeat anomaly detection, which is an important application for wearable systems.

The core idea is that we will pre-train a network with an existing dataset, and adapt its last layer’s weight to a non-seen input in a few-shot manner. This last-layer adaption will be done on the ALIVE chip. We already have a pre-existing PyTorch code for training a spiking neural network (SRNN) to classify Electrocardiogram (ECG) signals between normal and abnormal classes. The students will modify the code to be compatible with the ALIVE chip and will interface inputs with the chip to perform few-shot learning.

Detailed project plan
The project can be broken down in several smaller milestones:
1. Understand the pre-existing ECG classification PyTorch code.
2. Understand the ALIVE chip and the testing framework.
3. Modify the ECG classification network to match the chip constraints.
4. Interface the output of the ECG classification network with the ALIVE chip.
5. Perform few-shot learning on chip to adapt ALIVE to unseen data.