1:45 PM - 2:15 PM
Information
Dr. Anthony Lewis
Brainchip
Fast Online Recognition of Gestures using Hardware Efficient Spatiotemporal Convolutional Networks via Codesign
The Temporal Neural Networks (TENNs) developed by Brainchip can be used to tackle a wide range of vision tasks, including object detection, eye tracking and gesture recognition. Here, we will show how the codesign of model architecture, training pipeline and hardware implementation can combine to achieve SOTA performance, using a gesture recognition task example.
The TENNs architecture leverages multiple techniques to improve its efficiency on compatible hardware (such as the Akida chip). First, although effectively offering a 3D convolution, it uses spatially and temporally separable convolutions to make the model lighter in parameter count with equivalent computational power. Second, when deployed on dedicated hardware, temporal inputs are buffered efficiently to minimize memory usage and data movement. Finally, it is possible to reduce model computation even further by adding regularization to boost sparsity of information transiting in the already slim network (achieving more than 90% average activation sparsity in some layers) and thus further improve the efficiency on compatible hardware.
We apply a lightweight TENN model to a gesture recognition task, showing that it can accurately classify the movements performed by a range of actors with SOTA accuracy. The efficiency of the model is then pushed further with virtually no cost to accuracy by applying regularization of activations.
Looks like Tony will be taking the stage and delivering some magic on behalf of the Brainchip Family !!
In just a little over 7.5 hours from now, so work out your own time zones if you're interested, I am happy he has attended this event,
as our CTO you must respect the fact, yes it's a team thing, BUT Tony is now our leader on the technical side, as Peter has retired, as such.
Comon Brainchip, we are worth way more than our current ASX share price indicates, it's an absolute joke.......Tech x
brainsandmachines.net
(released on February 7), which is now available on multiple podcast platforms:
www.linkedin.com
