About This Project

Edge machine learning refers to the process of running embedded ML models on site using devices capable of collecting, processing, and recognizing patterns within collections of raw data. This project seeks to train one of such devices (Nicla SenseME) with dog data from the Earth Species' Bio-logger Ethogram Benchmark (BEBE). The board will be used as a smart dog collar, with its ML inferences controlling haptics vibrations in a bracelet wore by a person.

Ask the Scientists

What is the context of this research?

"Growing symbiotic new senses for humans" sets the context here. In experimenting with haptic interfaces to grant humans a direct perceptual experience of a dog's activities, we started to consider how edge ML could help to identify patterns in accelerometer data from a dog's tracker.

We've been able to create a prototype that demonstrates our vision for the use of edge sensing and embedded ML processing. However, we have used an open source ML model that was trained on human data. So the next step is to design an edge ML model with actual dog data. More specifically, data from "Movement Sensor Dataset for Dog Behavior Classification", Vehkaoja et al (2021), as described by the Bio-logger Ethogram Benchmark.

What is the significance of this project?

In our short paper "A Human-Canine Interface From Sensors, Haptics, and AI", we have outlined our larger vision and its main applications. Namely, the possibilities of such an interface to expand our direct perception of reality in relation to dogs, which could have immediate implications for:

1. Blind people who rely on guide dogs for locomotion.

2. Professionals who rely on dog's olfactory abilities to detect objects and even diseases (yes, dogs can detect 28+ diseases in real-time, often with higher accuracy than our best diagnostics).

Furthermore, Edge ML represents a great promise to non-human understanding from sensory data, especially due to on-site data processing. From conservation to sensory addition, this has been a key factor and our work is well placed to explore this frontier.

What are the goals of the project?

Our project begins by accessing and preprocessing canine data from "Movement Sensor Dataset for Dog Behavior Classification" by Vehkaoja et al (2021). Using Edge Impulse Studio, we'll train a machine learning model, ensuring data aligns with the "Data acquisition formatspecification". We'll test our model with a smart tracker on Maniçoba, specifically focusing on data from the neck and torso, as done in the original paper.

A pivotal aspect of our exploration is to discern the long-term effects of vibro-tactile feedback derived from canine behavior on human sensory perception. This could be especially relevant in the context of enhanced experiences for the visually impaired.

Our journey and findings will be summarized in a paper, detailing our workflow from data collection to haptic feedback.