Intel Neuromorphic Newsletter: Jan '24
Major releases improved Lava performance and enabled Loihi 2 features
Lava 0.8.0 (release notes) accelerated multi-chip network compilation by over 10x, added support for saving and loading compiled processes, introduced the Lava Peripherals package with support for Prophesee event cameras, and added Loihi 2 hardware features such as sparse connectivity, synaptic delays, and multiple inputs to µcode neurons.
Lava 0.9.0 (release notes) added continual learning prototypes (CLP) tutorials demonstrating 3-factor learning, added support for basic vector operations to enable VSA and normalization of a population of neurons, merged infrastructure to enable high-speed I/O to Loihi 2 chips, and fixed bugs in convolution, delay dense, and the install process.
Lava-DL 0.5.0 (release notes) added tiny YOLOv3 inference tutorial:
Team Clairaudience was awarded first prize in the Intel Neuromorphic DNS Challenge Track 1
Clairaudience claimed the top $15,000 prize for the Intel Neuromorphic Deep Noise Suppression (NDNS) Challenge. 19 models from 8 teams were evaluated for algorithmic innovation using an audio denoising dataset of human speech including English, German, French, Spanish, Russian and various categories of noise. The winning team’s submission improved audio quality by up to 8 dB using a novel Gated Spiking Unit neuron model, a sophisticated network architecture integrating both full-band and sub-band processing modules, and adversarial training techniques rarely applied so far in the neuromorphic field (winning presentation).