Daniel Situnayake: AI on the Edge

Feature Engineering

Feature engineering for edge AI presents unique challenges, particularly when dealing with high-frequency sensor data like audio. Daniel Situnayake explains that traditional methods of feature engineering, which involve manipulating columns in tables, are not effective for this type of data. Instead, signal processing tools are essential to extract meaningful signals from raw inputs, such as converting time domain signals to frequency domain to focus on relevant frequencies for human hearing 1. This process involves a trade-off between computational cost and model performance, as Daniel notes:

If you're doing that, you're probably going to be leaving performance on the table.

--- Daniel Situnayake

Additionally, collecting representative data sets for specific devices is crucial, as models often operate in streaming contexts rather than discrete data chunks, posing further optimization challenges 2.

   

Performance Constraints

Balancing performance and constraints in edge AI involves significant trade-offs, particularly with model compression techniques like quantization. Daniel Situnayake highlights that quantization reduces model size and computational cost by converting 32-bit floating point data to 8-bit integers, often with minimal performance loss 3. This allows models to run efficiently on small devices, but it can also lead to a loss of nuanced information in the data. Daniel Bashir adds that understanding these trade-offs is crucial for edge AI practitioners:

How much do we care about performance?

--- Daniel Bashir

Moreover, the heterogeneity of hardware means that the balance between digital signal processing and model performance varies across devices, requiring tailored solutions for each application 4.