SYSTEM AND METHOD FOR CONSCIOUS MACHINES

Abstract

Consciousness is widely considered to be a mysterious and uniquely human trait, which cannot be achieved artificially. On the contrary, a system and method are disclosed for a computational machine that can recognize itself and other agents in a dynamic environment, in a way that seems quite similar to biological consciousness in humans and animals. The machine comprises an artificial neural network configured to identify correlated temporal patterns and attribute causality and agency. The machine is further configured to construct a virtual reality environment of agents and objects based on sensor inputs, to create a coherent narrative, and to select future actions to pursue goals. Such a machine may have application to enhanced decision-making in autonomous vehicles, robotic agents, and intelligent digital assistants.

Claims

1. A computational system configured to interact with an environment, comprising: at least one input port configured to receive information relating to a self agent, other agents, and objects in an environment; at least one output port configured to produce an output to change a relationship between the computational system and the environment; a goal defining processor, configured to establish at least one goal based on at least a set of rules for interacting between the self agent and the other agents and objects in the environment; a simplified representation processor, configured to periodically update a simplified representation of the environment which provides a coherent narrative of the self agent interacting with the other agents and the objects in the environment, based on at least the received information, and store a series of the simplified representations in a memory; a predictive causal model, configured to predict a plurality of alternative future states of the self agent, the other agents, and the objects within the environment; and a decision processor, configured to decide between prospective alternate actions of the self agent, based on at least the at least one goal, the stored simplified representations of the environment, and the plurality of alternative future states.

2. The computational system according to claim 1, wherein the output is dependent on the decision between the prospective alternate actions.

3. The computational system according to claim 1, wherein the output is dependent on the simplified representation processor and independent of the decision processor.

4. The computational system according to claim 1, further comprising a dynamic temporal pattern recognition processor, configured to process the received environmental information and recognize changes in the self agent, the other agents, and the objects in the environment, wherein the predictive causal model is responsive to the recognized change.

5. The computational system according to claim 1, further comprising a feedback processor, configured to refine the predictive causal model using success in meeting the at least one goal represented in the received information as a basis for feedback.

6. The computational system according to claim 1, further comprising a reference clock configured to define a period for the periodic updates of the representation.

7. The computational system according to claim 6, further comprising a retrieval processor configured to retrieve respective simplified representations of the environment by time-stamp and by a content of the respective simplified representations of the environment.

8. The computational system according to claim 1, wherein the environment comprises a real physical environment and the self agent further comprises a navigational process for navigating in the real physical environment.

9. The computational system according to claim 8, where the self agent comprises at least one of an autonomous vehicle and a robot.

10. The computational system according to claim 1, wherein the other agents are autonomous navigation mobile devices.

11. The computational system according to claim 1, wherein the environment comprises a virtual environment in which the self agent can navigate, for training of the computational system.

12. The computational system according to claim 1, further comprising a communication process configured to communicate with the other agents in the environment comprising human agents and computational agents.

13. The computational system according to claim 1, wherein the decision processor comprises an artificial neural network having a plurality of layers comprising at least one hidden layer.

14. The computational system according to claim 13, wherein the artificial neural network comprises an automated processor selected from the group consisting of at least one of a tensor processing unit, a graphics processing unit, a field programmable gate array, and a neuromorphic chip.

15. A computational method for interacting with an environment, comprising: receiving information relating to a self agent, other agents, and objects in an environment; establishing at least one goal based on at least a set of rules for interacting between the self agent and the other agents and objects in the environment; periodically updating a simplified representation of the environment which provides a coherent narrative of the self agent interacting with the other agents and the objects in the environment, based on at least the received information, and store a series of the simplified representations with at least one automated processor; predicting a plurality of alternative future states of the self agent, the other agents, and the objects within the environment with a predictive causal model implemented on the at least one automated processor; and deciding between prospective alternate actions of the self agent with the at least one automated processor, based on at least the at least one goal, the stored simplified representations of the environment, and the plurality of alternative future states; produce an output to change a relationship between the computational system and the environment.

16. The computational method according to claim 15, wherein the output is dependent on the decision between the prospective alternate actions.

17. The computational method according to claim 15, wherein the output is dependent on the simplified representation processor and independent of the decision.

18. The computational method according to claim 15, further comprising processing the received environmental information to recognize changes in the self agent, the other agents, and the objects in the environment, wherein the predictive causal model is responsive to the recognized change.

19. The computational method according to claim 15, further comprising refining the predictive causal model using success in meeting the at least one goal represented in the received information as a basis for feedback.

20. A method for controlling interactions of a respective agent within an environment, comprising: defining a set of rules for interacting between the respective agent and the environment; recognizing active other agents in the environment, objects in the environment, and the respective agent in the environment; recognizing changes in the environment with a dynamic temporal pattern recognition system based on an environmental status input; sequentially generating and storing a series of simplified representations of the environment comprising other agents, the respective agent, and the objects; defining at least one goal for the respective agent based on at least the set of rules; making alternative predictions of future states of the other agents, the respective agent, and objects within the environment using a predictive causal model, based on alternate different actions by the respective agent and past representations stored in the memory; generating decisions based on the goals and the alternate predictions; and refining the predictive causal model using success in meeting the goals as a basis for feedback.

Description

BRIEF DESCRIPTION OF THE FIGURES

[0089] FIG. 1 shows a classic von Neumann computer architecture of the prior art.

[0090] FIG. 2 shows a conceptual example of an artificial neural network of the prior art.

[0091] FIG. 3 shows a crude block diagram of a conscious mind and an unconscious mind interacting with the environment, which may be emulated in a conscious machine.

[0092] FIG. 4 shows a preferred embodiment of a conscious computing module and an unconscious computer module, together with input and output structures.

[0093] FIG. 5 shows an agent moving in an environment and causing an object to move.

[0094] FIG. 6 shows a conceptual diagram of a virtual reality environment with the self, other agents, and objects.

[0095] FIG. 7 presents a preferred embodiment of a memory structure and time base for a conscious machine.

[0096] FIG. 8 presents a flow chart for a preferred method for designing and training a conscious machine.

[0097] FIG. 9 shows alternative modes of programming artificial neural networks (ANNs) for conscious machines.

DETAILED DESCRIPTION OF THE INVENTION

[0098] Although consciousness has been difficult to define, most researchers in artificial intelligence would agree that AI systems to date have not exhibited anything resembling consciousness. Conventional views of consciousness are mostly illusory, so that a new definition of consciousness may provide a basis for developing a conscious machine. The key is pattern recognition of correlated events in time, leading to the identification of a unified self-agent.

[0099] The sense of consciousness may represent the real-time activation of neural circuits linking the present self with the past self. Such a conscious system can create a simplified virtual environment, edit it to reflect updated sensory information, and segment the environment into self, other agents, and relevant objects. It can track recent time sequences of events, predict future events based on models and patterns in memory, explore possible results of future actions, extrapolate based on past trends or experiences, and attribute causality to events and agents. It can make rapid decisions based on incomplete data, and can dynamically learn new responses based on appropriate measures of success and failure. In this view, the primary function of consciousness is the generation of a dynamic narrative, a real-time story of a self-agent pursuing goals in a virtual reality.

[0100] A conscious machine of this type may be implemented using appropriate neural networks linked to episodic memories. Near-term applications may include autonomous vehicles and flexible robotic assistants.

[0101] A major illusion of human consciousness is one of unified top-level control. FIG. 3 shows a diagram representing a conscious mind C above an unconscious mind U, which in turn is above the environment E. We believe that our conscious mind controls most of what we do, and that our conscious mind is in direct communication and control of the environment. But psychological research has shown that in many cases, actions are triggered slightly before the conscious mind is aware of them, suggesting that these may actually be directed unconsciously. The conscious mind may take credit for these actions, which form part of a narrative of conscious action. There are other internal experiences of consciousness, including perception of time and space, identification of a unified self linked to past memories, emotions, and actions, and identification of other agents and objects. A consistent explanation of the mind needs to account for all of these. Further, neither C nor U is fully rational, but U creates a consistent simplified narrative that C experiences. The present application proposes that similar aspects may be emulated in a conscious machine, with a similar organization.

[0102] Many aspects responsible for human consciousness are hidden from view, and may not be evident either in the structure of the brain or in the internal experience of consciousness. But as disclosed here, consciousness seems to involve a self-identified agent following a continuous, causal narrative in a dynamic virtual environment. The environment must integrate various sensory modalities with relevant emotions and memories. This is shown in FIG. 4, which shows consciousness as a virtual reality (VR) construct created from filtered input data, and representing a simplified dynamic model of the reality. The interaction with the external environment occurs via filtered inputs and outputs to the unconscious mind, where most of the detailed coordination and decision takes place. The VR environment represents the self, acting in a simplified environment, comprised of objects and other agents. Previous frames of the VR are saved in memory, and may be retrieved as needed.

[0103] A conscious visual representation of an object (such as a rose) is not just a portion of a larger two-dimensional image. Rather, it is an object embedded in a three-dimensional space, which represents the concept of a rose and links to abstract attributes and memories of images and aromas (even if ancillary sensory stimulation, such as smell, is not currently present). This may be analogous to a PDF file of a scanned document which is subjected to optical character recognition (OCR). Such a file contains an image of a given page, but the image of each character is also associated with the functional representation of the character. Now imagine that the document also contains embedded links to definitions of words and other relevant documents, and one starts to have the richness necessary for consciousness. In contrast to a static environment of a document, the VR environment is highly dynamic, and is rapidly updated in time.

[0104] Another aspect of consciousness is the subjective sense of agency. This is created by activation of an adaptive neural net, primed to recognize self-agency and causality, and also to recognize external agents. The dynamic VR is built around such actors, as shown in FIG. 6. Note that recognition of agency is really a form of temporal pattern recognition, as suggested in FIG. 5, which shows an agent moving up to a fixed object, which then starts to move. We live in a world that is continuous and causal, so that simplified causal models based on observed temporal correlations are generally highly functional. Furthermore, this is a case of dynamic learning; the mind learns to generate and refine a simplified model which maintains effective representation of the external environment.

[0105] A further aspect of consciousness is a sense of urgency or priorities. Often, the issues have different objective classifications, and the ranking requires some normalization of different classes of issues. A cost or distance function may therefore be included within the model. Consistent with our understanding of consciousness, this ranking or ranking function may be highly dynamic, and have what are apparently irrational characteristics, with biases, systematic “errors”, and other “personal” characteristics. While in some cases, such as an autonomous vehicle controller is desired to avoid unpredictable action or irrational behavior, in other cases, it is exactly these attributes that make the system better appear as being conscious.

[0106] In a further key aspect, this view of a conscious mind or a conscious machine requires both a clock and a memory structure. Consider a short-term dynamic memory module, containing the recent past, and a predictive model of one or more near futures, as shown in FIG. 7. A clock time-stamps the present frame and shifts it to the past. Two alternative futures may be presented, based on present actions. When one action is selected, this shifts that future into the present. This ensures that perceived time is a central element in consciousness. The subjective sense of self is associated with the repeated activation of self-recognition circuits, mapping the present self onto the past self.

[0107] Longer term memories are also stored, possibly in another location, based on a different technology. These may be retrieved either by time reference or by content. The memory elements may be VR frames, but other parameters abstracted from one or more prior experiences may be independently stored and retrieved.

[0108] The presence of a conscious VR module does not preclude a machine or mind from also incorporating unconscious perception and control via neural circuits that do not interact with consciousness. For example, detailed coordination of specific muscles in walking or running is an essential part of brain activity, but is generally completely inaccessible to the conscious mind. Similarly, a conscious machine will incorporate a hybrid approach, assigning to the conscious module only those aspects that cannot be achieved efficiently by a more conventional machine learning approach.

[0109] Indeed, we are accustomed to think of consciousness as a superior form of brain activity, but it might be better to think of a conscious mind as a specialized niche processor that should only be used as a last resort. For example, consider a car driving along a road, with a large puddle in the road. Is this a shallow puddle that one should simply drive through, or a deep pothole that should be avoided? If the answer is obvious, this need not rise to the level of consciousness. But if it requires a more complex assessment of weather conditions and the local status of road repair, the decision needs to be made at a higher level. There is a danger in this—such high level decisions based on incomplete data can be slow, at a time when a rapid decision is necessary. And the decision can still turn out to be the wrong one. However, any system that significantly increases the likelihood of improved decision making has substantial value, for a self-driving car or other autonomous system. For example, a conscious system might further consider ramifications of error (or success). If the puddle is shallow, and the vehicle swerves to avoid it, what are the risks of accident or discomfort for the passengers? If the puddle is deep and the tire hits it, what damage could occur? If the car brakes suddenly, what is the risk of skidding or a rear-end accident? Etc.

[0110] As with other neural network systems, the initial design and training of a conscious machine are essential, and a simplified flowchart of steps in initializing a conscious module are presented in FIG. 8. The first step is to design an initial model for the machine moving in its environment. What are the important sensors and actuators needed, and what are the relevant timescales? Second, the neural network must be configured to evaluate sequences in time and space, and designed to use temporal pattern recognition to identify agents in the environment. The most important agent is the “self”, which is correlated with the sensors and actuators. Other agents and objects can also be identified, and a virtual reality environment can be generated based on the interactions of the self with these other agents and objects. This machine can be trained initially using a simplified external environment. In some cases, the initial training environment might be a computerized VR system. Finally, the machine can be transferred to the real world, where learning will continue. Throughout the training and learning, a method to monitor the internal VR environment would be highly advantageous. This would be preferable to simply observing the functional behavior. One can imagine that training young children or animals would be much easier if we could really read their minds!

[0111] It is known in the prior art that deep neural networks may be implemented in a variety of technologies, including digital and analog circuits, biological neurons, CMOS transistors, superconducting Josephson junctions, memristors, phase-change memories, and resistive RAMs, based on pulse or voltage-level logic. They may be implemented in a variety of circuit architectures, including not only conventional processors, but also special-purpose processors such as GPUs, TPUs, and FPGAs. The deep neural networks comprising a conscious machine may in principle be implemented on any device technology and architecture that may be configured to support temporal pattern recognition, and to identify agents of change in the virtual environment created by the set of sensory inputs.

[0112] This temporal pattern recognition represents a repeated process activated by a periodic clock which establishes a time base. The frequency of this updating can be relatively slow if the systems deals with slow changes in typical human environments; for comparison, the alpha rhythm in human brains is of order 10 Hz. But electronic systems are capable of much faster operation, so that faster updating might be appropriate for application to a rapidly moving autonomous vehicle, for example.

[0113] In order to be able to act with a sufficient level of sophistication in a complex environment, the neural network of a conscious machine must comprise a large number of neurons with an even larger number of interconnections. For example, the network must have at least millions of neurons, with at least billions of interconnections. The strengths of the interconnections representing memories and associations must preferably be non-volatile over long periods of time, or at least fully backed up in case of power failure. The memories must be capable of repeated adjustment and readout, with high reliability and very low rate of failure. The system should be able to be temporarily turned off, so that repairs or upgrades may be implemented, and then turned back on in a way that remembers past performance.

[0114] The neural networks comprising a conscious machine must be capable of learning during an initial supervised or unsupervised training period, but should also continue to learn during full operation in the field. In this way, a conscious machine comprising a robot or autonomous vehicle could be optimized for operation in an environment with distinct but unpredicted characteristics.

[0115] Furthermore, in some cases, the initial training period may be accelerated by pre-programming some of the interconnections, based on a readout of the internal interconnections of another conscious machine that has already been trained. This effectively comprises implanting a set of artificial memories and experiences into a given machine consciousness. It may be advantageous to use such a procedure to create a plurality of identically trained conscious machine twins or clones. In this way, a standard machine with reliable and validated performance may be efficiently mass-produced. Further adaptation to custom environments may be obtained in the field.

[0116] Biological neural networks are not normally designed to read out their internal connections, but if such a readout were available, it could in principle enable an artificial system to emulate aspects of the mind of a biological organism. That could enable, for example, the creation of an artificial pet that could better emulate some aspects of the behavior of a biological pet.

[0117] These alternative modes of programming artificial neural networks (ANNs) for conscious machines are illustrated in the block diagram of FIG. 9. The memory weights in the interconnections of the ANN may be determined by simulating a conscious machine in a virtual environment, or by training the machine in a natural environment, or even by reading out the states of an analogous biological neural network (should that become feasible). These can then be written to the states of untrained ANNs of a plurality of conscious machines. These machines can then be distributed to customers, where further learning in a variety of natural environments can continue.

[0118] Biological behavior in social animals may be governed in part by issues of cooperation and competition. Cooperation depends in part on empathy, the ability to project the perspective of another agent on one's own perspective, in order to better predict the action of other agents. This may be associated with “mirror neurons” in biological neural nets. Competition in social animals may depend on dominance hierarchies, which provide rules to negotiate differences in order to avoid conflict. In an environment which may comprise a plurality of both people and conscious machines, it may be important that the machines incorporate mechanisms of both cooperation and competition. Similar issues have been anticipated in the literature of science fiction, for example in the Three Rules of Robotics of Isaac Asimov (en.wikipedia.org/wiki/Three_Laws_of_Robotics).

[0119] Indeed, in an environment of interacting conscious machines, an exchange of virtual reality environments between cooperative agents may occur, so that one machine can better predict outcomes of interactions with other machines, and improve its own virtual reality environment. Such inter-machine communication could be acoustic or visual, but it could also occur via wireless rf communication channels. Of course, in some cases, agents are competitive, and this exchange would be disfavored.

[0120] One aspect of biological consciousness that is not generally considered for artificial intelligence is the role of sleep and dreams. But sleep and dreams are universal among animals, and their deprivation is highly deleterious, suggesting that they must serve an important function, even if it not well understood. Some recent research has suggested that neural interconnections in animals grow dramatically during the day, but are selectively pruned back during sleep. This can be a form of consolidating learning, which may be important to emulate in machine systems. Furthermore, dreams are associated with rapid eye movement (REM), which is also common among animals, suggesting that they, too, dream during sleep. Dreams are a form of virtual reality that is somewhat similar to the real-world experience. The difference, of course, is that there is no sensory input during dreams, but there is still a sense of self and a dynamic narrative. Dreams may also represent part of the learning process. A conscious machine could also operate without sensor input, but with some low-level activation of memories. Perhaps we will know that we have made conscious machines when we can observe them dream.

[0121] A conscious machine might present another issue that is normally restricted to brains: mental illness. For example, malfunctioning of the VR generator may present distorted perceptions or narratives, which might be analogous to paranoia or schizophrenia. Furthermore, memory activation thresholds that are too high or too low might cause hyperactivity, depression, or obsessive behavior. While this is purely speculative at present, it may be important to monitor a conscious machine for abnormal behavior or thoughts. From another point of view, a malfunctioning conscious machine might even represent a model for human mental behavior.

[0122] Finally, a reliable, inexpensive, conscious machine would enable a wide range of potential applications, many of which have not even been considered. Going beyond autonomous vehicles or intelligent digital assistants, one can imagine a variety of security and military applications, from surveillance to monitoring the Internet to active defense. Similarly, one can envision intelligent medical systems for monitoring and diagnosing patients, in the absence of medical personnel. Alternatively, one might have an artificial pet, a companion without some of the requirements and shortcomings of dogs or cats. The only limit is our own imaginations.

[0123] In a medical environment, the conscious machine may seek to model the behavior of a patient. As characteristics of the patient are successfully modelled and their predictive nature verified, it may then be possible to analyze the virtual reality environment to determine deviations from normal, and thus the system could be part of a diagnostic or therapeutic device.

[0124] More generally, the system may be used to implement user models, typically within a particular knowledge domain, to provide assistive agents.

[0125] While the present invention has been described with respect to a limited number of embodiments, those skilled in the art will appreciate numerous modifications and variations therefrom. It is intended that the appended claims cover all such modifications and variations as fall within the scope of this present invention.