2017 SR&ED Claim

2017 Scientific Research and Experimental Development tax credit claim

T661 - Part 2 – Project Information

Section B – Project Description

Line 242 -What scientific or technological uncertainties did you attempt to overcome- uncertainties that could not be removed using standard practice? (Maximum 350 words)

The project is attempting to develop general purpose learning and thinking software for the autonomous robot control. The approach is based on the development of a new artificial neural network (ANN).

Current robot control software is not general purpose. It is specifically developed to function in a predefined environment with predefined tasks to be performed. An example is Google’s self-driving car control software. Learning capability, if present, is limited to this environment. General purpose artificial intelligence (AGI) architectures exist but are not designed for robot control i.e. they are not grounded on sensory input and device activation. Examples include ACT-R, SOAR, CLARION, LIDA, SIGMA and HTM.  These AGI architectures are also applied only in specifically defined environments. The DARPA Robotics Challenge provides good examples of the state of the art robotic control software: https://en.wikipedia.org/wiki/DARPA_Robotics_Challenge.  When such architectures are grounded on real sensory input they also make heavy use of stochastic inference that becomes complex and is a heavy computational overhead. One of the reasons for the complexity of these architectures is that they do not use the same structures for pattern recognition and motor action control.

This research is developing a new architecture to solve these problems. However there are no simple hierarchical (deep learning) ANNs based on binary nodes that are feed-forward and use reinforcement learning. There are also no ANNs that grow nodes in a hierarchy for both recognizing objects and sequences as well as for learning new action habits. Even the deep learning pattern recognition ANNs, such as convolutional neural networks, do not add nodes dynamically as they learn.

More specifically the project is attempting to develop a new hierarchical structure of ANN nodes that:

  • Uses binary neurons (binons) with reinforcement learning based on intrinsic motivation.
  • Grows the ANN by adding nodes for parallel (spatial) and sequential (temporal) pattern recognition.
  • Learns continuously - does not separate the training phase from the testing phase.
  • Converts magnitude sensor readings (sub-symbolic) into symbolic stimuli.
  • Learns action habits and integrates them into the ANN.
  • Executes action habits subconsciously while thinking about other actions (simulation based on experience).

Line 244 – What work did you perform in the tax year to overcome the scientific or technological uncertainties described in Line 242? (Summarize the systematic investigation) (Maximum 700 words)

Approach: Increase the complexity of the ANN structure and algorithms to process the requirements at greater levels of complexity while still handling the lower complexity features. Run regression tests on already working lower complexity features. Determine its success at learning and thinking based on the observation of its actions in artificial test environments simulated in software and by inspection of its internal memory traces and processes. More details about the research are available at www.adaptroninc.com.

Three pieces of software are being used to validate the research:

A) Morse code is being used for sequential (temporal) pattern recognition from multi-sensor senses producing magnitude (ratio scale) and discrete (symbolic / nominal) stimulus readings using a hierarchy of binons. This software uses a state and spike driven binon activation approach.

B) Hand written digits are being used for object (spatial) pattern recognition using a topological activation structure and binon hierarchy for long term memory of patterns/objects and their associated labels. This treats the two dimensional array of pixels as a one dimensional row of stimuli.

C) Over 50 test cases and simulated robot bodies in a maze world are being used to test the pattern recognition and reinforcement learning based on intrinsic motivation (exploration).

In area A (Morse code recognition):

A.1 – Modified the activation tree and short term memory structure to recognize multi-modal sensory input. The 3 sense properties are duration of the dit-dahs and quiet spaces, the intensity of the dit-dahs versus the quiet spaces and the letter labeling the Morse code.

A.2 – Interspersed the recognition binons with response binons. These were either covert responses for orienting purposes or covert actions for predicting the letter before the letter label was provided.

In area B (Handwritten digit recognition):

B.1 – Continued applying the state and spike driven binon approach from area A to spatial pattern recognition of hand drawn digits.

B.2 – Tried an approach to separate patterns from parts. This was based on the idea that parts (objects) had to be combined next to each other to form a pattern and only patterns of parts could be combined using overlaps. Once a pattern of parts became a part due to repetition it could no longer be combined using overlaps like a pattern.

B.3 – Made rows of pixels into patterns before combining the rows together to incorporate some of the two dimensional properties.

B.4 – Added the MNIST handwritten digits for recognition.

B.5 – Applied the center-surround receptive field calculation as a contrast detector to enhance edge detection for blurry handwritten digit images.

B.6 – Used a position to size ratio to represent the gaps between two patterns to recognize groups of objects according to the Gestalt laws.

In area C (Stimulus-response and reinforcement learning):

C.1 – Used the state and spike driven binon approach from area A to test the 50 test cases.

C.2 – Added 2 more robot body configurations to see how the existing and new robots would behave in a simulated maze.

Line 246 – What scientific or technological advancements did you achieve as a result of the work described in line 244? (Maximum 350 words)

1/ Applying the state and spike driven binon approach for the temporal learning of Morse code also worked for the spatial learning of handwritten digits.

2/ Found out that recognizing parts and patterns did not increase the prediction accuracy of handwritten digits compared with just recognizing patterns.

3/ Prediction success and familiarity can be used to stop the formation of unnecessary binons at higher levels of complexity that don’t add anything to better prediction accuracy. This effectively stops the growth of the hierarchy until novel patterns are experienced.

4/ Discovered that interspersing recognition binons with response binons did not slow down the learning rate and produced both a forward model of motor control (stimulus to response) and a feedback control (response to stimulus) in Morse code recognition and robot motion.

5/ The use of the center-surround receptive field calculation for edge detection on blurred images works better than the calculation of contrast velocity and acceleration for edge detection.