Sparse Predictive Hierarchies with Eric Laukien

This is a past event

20 people went

1909 (FKA Palm Beach Tech Space)

313 Datura St 2nd Floor · West Palm Beach, fl

How to find us

Walk down the main hall and take the elevator to the 2nd Floor. If outer door is locked, call the number on the posted sign.

Location image of event venue


Sparse Predictive Hierarchies (SPH) is a novel technology developed by Ogma Corp. It offers an alternative paradigm to that of Deep Learning (DL). SPH addresses some of the fundamental weaknesses of DL: lack of online/incremental learning, slow/inefficient learning, and biological implausibility. As the SPH approach operates fundamentally differently from the typical dense backpropagation networks of DL, we will describe SPH by first introducing concepts such as sparse coding, bidirectional hierarchies, predictions as context for further predictions, exponential memory, and world model building.

From there, we will discuss how we are able to perform both inference and learning in real-time on very limited hardware such as the Raspberry Pi Zero by exploiting online learning and sparsity. We will also discuss concerns such of pre-encoding and general usage of SPH, as well as extensions to reinforcement learning (RL). A demo of a tiny self-driving car will by either shown pre-recorded or demonstrated on site.

Company Info & Speaker Bio:

Ogma was founded on the idea that some of the fundamental limitations of Deep Learning (DL) require a different machine learning paradigm.
They research primarily biologically plausible approaches and abstract them such that their implementations are fast. The speed of their algorithms is one their main strengths, as they feel that DL requires computational resources of an infeasible type and quantity in order to achieve AGI. Ogma has its office in Jupiter, Florida.

Eric Laukien is one of the founders of Ogma. He has a background in game development which has been proved helpful in using technologies such as OpenCL to run software on GPUs. When it comes to machine learning, he is mostly self-taught, which is probably why the algorithms he develops are so novel compared to mainstream Deep Learning. In addition to the research he does at Ogma, he is interested in evolutionary algorithms, artificial life, and swarm intelligence.

Meetup starts at 6:30 and talk will begin around 7:00. Free pizza and refreshments will be served thanks to our sponsor NextEra Energy.