General Synaptics

Safety_Critical_Machine_Learning

Probabilistic Adaptive Learning Mapper (PALM™) is the result of a research that gives answers to some issues raised in the DARPA L2M (Lifelong Learning Machines) and LwLL (Learning with Less Labelling) programs.

Lifelong Learning (L2) Systems are neural models capable of learning new data (in a supervised, unsupervised way or through reward / punishment mechanisms) during their normal inferential functioning.

It is essential to clarify that the learning algorithm of Error Back-Propagation on which Deep-Learning is based cannot be considered in this category as it requires multiple learning cycles (epochs) on all previously learned data. Learning iterations with just new data would lead to Catastrophic Forgetting situations.

Learning algorithms such as ART (Adaptive Resonance Theory - Stephen Grossberg & Gail Carpenter) and RCE (Restricted Coulomb Energy – Leon Cooper) on RBF (Radial Basis Function) architecture allow the continuous learning of new data avoiding catastrophic forgetting phenomena.

Learning algorithms such as ART (Adaptive Resonance Theory - Stephen Grossberg & Gail Carpenter) and RCE (Restricted Coulomb Energy - Leon Cooper) on RBF (Radial Basis Function) architecture allow the continuous learning of new data avoiding catastrophic forgetting phenomena.

Of course, there are always problems inherent in any learning algorithm. Overfitting on noisy data and lack of statistical consistency are two typical examples.

Both ART and RCE can be modified taking into account the data statistics (as in Probabilistic Neural Network - PNN). This and other more sophisticated mechanisms can lead to the achievement of the following results:

· Supervised learning when needed

· Continuous unsupervised learning that inherits from supervised learning

· Continuous adaptivity

· Statistical consistency

· STM (Short Term Memory)

· LTM (Long Term Memory)

· Smart statistical management of forgetting

· Smart statistical management of limited resources

· Statistically controlled balance of plasticity versus stability

We have obtained these results with the PALM™ (Probabilistic Adaptive Learning Mapper™) neural model.

Continuous learning can be achieved also with our ROCKET™ (Registers Optimized Chip Knowledge Emulation Technology™) which is an ultra high speed software simulator of a RBF-RCE neural chip.

Continuous learning can be achieved also with our LASER™ (L-Anti-Spoofing Engineered Recognizer™) algorithm.

Continuous learning with RULES-EXTRACTION can also be achieved with our SHARP™ (Systolic Hebb Agnostic Resonance Perceptron™) algorithm.

WhoWeAre WhatWeDo Technology

Customers Partners News

General Synaptics

Aerospace_and_Defence_Machine_Learning_Company

VAT NUMBER:_IT02670700992

REA NUMBER: GE-503104

NATO_CAGE_CODE: AK845

Email:_luca.marchese@synaptics.org