Natural neural systems tightly couple a huge diversity of neuron types and connections to perform extremely efficient pattern recognition processing. Perfected through a billion years of natural selection, an ant can track a scent trail, a duckling will imprint on its parent, and a young child can recognize a general category of non-natural objects such as cars after seeing only a few such examples.

On the other hand, most neural networks in use today must be trained with hundreds or thousands of supervised images. Almost all current artificial neural network (NN) research has been focused on homogeneous systems of identical neuron types in mostly multilayered perceptron models. This owes partly to the availability of powerful GPU machines, which can process data very quickly but are limited to a single-instruction, multiple-data (SIMD) paradigm.

New machine architectures can support MPMD (Multiple-Program, Multiple Data) models, allowing researchers to develop more powerful models of machine learning that emulate the awesome learning and recognition abilities of the brain evolved through natural selection. These sparse and heterogeneous neural network models pipeline data through specialized processing algorithms to perform specific interpretive tasks. Running on MPMD architecture, Heterogeneous Neural Network (HNN) models can save a lot of training time and computational resources compared to homogeneous networks.

Workshop on Heterogeneous Neural Networks

In partnership with the California Institute for Telecommunications and Information Technology (Calit2) at UC San Diego, KnuEdge is sponsoring the Heterogeneous Neural Network (HNN) Workshop, to be held in April, 2017 in San Diego, Calif. The event will also include a KnuEdge-sponsored research paper competition, challenging participants to enable the next-generation of machine learning performance and efficiency through developing heterogeneous neural network algorithms.

  • Date: April 26-28, 2017
  • Location: Calit2 Theater, Atkinson Hall on the campus of UC San Diego
  • Attendance fee: $250; $119 for members of academia; $50 for students or free with abstract presented in workshop
  • Speakers will include a variety of experts from UCSD and KnuEdge, including Larry Smarr, Dan Goldin and Ken Kreutz-Delgado
  • Sessions will be archived and accessible via this site 1 week after the event
  • Conference Tracks will be announced in Q4, 2016; actual topics, speakers and other agenda items will be set and communicated in January, 2017

Research Papers Wanted

We welcome submissions of abstracts for papers that demonstrate the benefits of sparse networks in both application and theory from the computer science, cognitive science, neuroscience, biophysics, natural science, and other communities. Selected applicants will be flown out to present their work at our Symposium in beautiful San Diego, California, in April of 2017, in front of an audience of colleagues and experts!

The workshop is intended to highlight the benefits of applied sparse and heterogeneous NN models; for example, in terms of computational resource efficiency in model training. The applications and objectives of the your research paper should be of your own choosing, but it should focus on applications and design of models and illuminate the comparative benefits of a sparse or heterogeneous NN approach relative to a homogeneous or traditional NN model.

Appropriate topics may include, for example:

Pattern recognition applicationsMachine Learning processesNeural network architecture
Facial identificationRecurrent modelsModel construction
Voice and language processingBayesian networksModel tuning and pruning
Handwriting processing and identificationInformation-theoretic learning modelsSelf-organizing models
Text string pattern search and identificationOther learning modelsCognitive networks
Image and video processingMeta-learning algorithms
Other sensory classificationFeature selection
Evolutionary models

Specialized MPMD Computational Resources Available

KnuEdge, parent company of KNUPATH, is making its proprietary KNUPATH Hermosa MPMD processors optimized for sparse HNN processing available to workshop participants. If desired, applicants can develop applications in Hermosa-enabled C and submit processing jobs to the Hermosa servers on a space-available basis. KNUPATH will award special prizes to the best papers that include applications developed on the Hermosa architecture.

To register for this event, please click here.

Questions on the workshop or submissions can be directed toward