2024 SR&ED Claim

2024 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 reasoning software for autonomous robot control. The approach is based on the development of a deep symbolic neural network (ANN). Key requirements include: its decisions must be explainable, grounded on sensors and action devices, and it continuously learns.

Current robot control software is not general purpose. It is specifically developed to function in predefined environments and to perform predefined tasks. This is also known as narrow AI. An example is Tesla’s self-driving car whose learning capability is limited to driving. General purpose artificial intelligence (AGI) architectures exist but are not designed for robot control, i.e., they are not grounded on sense and action devices. Examples include ACT-R, SOAR, CLARION, LIDA, and HTM. These AGI architectures are also applied only in specifically defined environments. The Boston Dynamics Atlas robot is a good example. It performs well in a closely controlled environment (e.g. https://www.youtube.com/watch?v=-9EM5_VFlt8) but cannot transfer its experience to similar environments without reprogramming. These architectures make heavy use of stochastic inference which results in behavior that is not explainable. Also, their ability to continuously adapt to their environment is seriously limited.

This research is developing a new cognitive architecture (named Adaptron) to solve these problems. Currently there are no simple deep symbolic ANNs based on simple binary nodes that are feed-forward, explainable and use reinforcement learning. There are no ANNs that grow nodes in a hierarchy for learning to recognize objects and events and for learning to control action habits. Even deep learning ANNs, such as convolutional neural networks, do not add nodes dynamically as they learn.

More specifically the project is attempting to develop a new deep symbolic ANN 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 (non-symbolic) into symbolic representations,
  • Learns hierarchies of action habits, performs them in sequence and integrates them with the perception ANN,
  • Learns to reason, and
  • Has explainable behavior.

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)

Research approach:

  1. Use cognitive science, psychology and artificial intelligence theories to design the perception, reasoning and action ANNs and algorithms.
  2. 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.
  3. Run regression tests on already working lower complexity features.
  4. 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 structures, traces and processes.
  5. Rethink if the theory needs changing and try out alternate designs. More details about the research are available at www.adaptroninc.com.

The following two software applications are being used to perform the research:

  1. A general-purpose robot and grid world with walls and doors and robot simulation environment called DAVE (Deterministic Autonomous Vehicle Environment) was built last year. It is being used for testing the perception and action habit learning and control algorithms in the Adaptron perception-thinking-action ANN software (called ONWARD in B below). DAVE can be configured with multiple sensors (switches, cameras, microphones, buttons, etc.) and action devices (speakers, lights, wheels etc.) such that the robot can interact with the world. Proprioceptive feedback sensing was added to the robot and the interaction rules between sensing and action devices were improved. Additional windows were added to monitor the input and output of stimulus and action data.

Documents:  DAVE User Requirements documents
                     DAVE Architecture Document
                     World and Robot simulations
Software:      \PycharmProjects\DAVE-Software
Versions:      Maintained in Github folders

  1. The robot is controlled by the Adaptron perception-thinking-action ANN which runs as an independent program called ONWARD (Ongoing NetWork for Adaptron Research and Development) communicating via sockets with DAVE, the world and robot simulation software.
    1. The detailed structure of the ANN was changed several times to properly incorporate shape and repetition patterns for perception.
    2. The processing cycle (perception, attention, thinking and action) was modified several times.
    3. We added the process of focusing attention and its distraction by novel and interesting stimuli.
    4. We built screens and algorithms to calculate and display the accuarcy of the ANN expectations and to test its ability to make predictions.
    5. We continue to modify our theories and design for the learning and control of action habits that are integrated with the perception process to continue the execution of action sequences provided the expected outcomes continue to occur. Actions get triggered by the perception process when the current situation is recognized, and a known desirable outcome is predicted. The decision to act results in the activation of all the action habits that are known to produce the outcome but the only one that is executed is the one that is based on the current situation.
    6. Options for controlling actions that override Adaptron’s normal action selection were added.
    7. The same action control and learning architecture is also being used to control and learn mental operations for reasoning and decision making.
    8. The short-term working memory structure and process for recognizing temporal sequences was added and modified several times.

Documents:  ONWARD Requirements Documents
                     Architecture Documents (ONWARD)
                     Many ONWARD implementation documents
                     Test result documents
Software:      \PycharmProjects\ONWARD
Versions:      Maintained in Github folders

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

  1. In cognitive science there are several theories as to how attention functions. We divided the perceptual field of all the sensors of all senses into the attention field and background (non-attention) field. They were treated differently during perception. In the background known percepts were encoded and recognized. In the attention field new percepts were formed and memorized. We found that of the several cognitive science theories as to how attention functions, the late attention processing is the best approach for perception.
  1. We discovered that since actions cause changes in the world, they should be associated with instantaneous change events and not the event patterns that comprise the trigger and goal states.
  1. We found that proprioceptive feedback from action devices is critical for making action decisions. If it is not available there is no way to recognize what was done in given situations that produced desirable results. Without it a robot will continue to try previous actions even though the results have been explored before.