Abstract: Apparatus and methods for training and controlling of e.g., robotic devices. In one implementation, a robot may be utilized to perform a target task characterized by a target trajectory. The robot may be trained by a user using supervised learning. The user may interface to the robot, such as via a control apparatus configured to provide a teaching signal to the robot. The robot may comprise an adaptive controller comprising a neuron network, which may be configured to generate actuator control commands based on the user input and output of the learning process. During one or more learning trials, the controller may be trained to navigate a portion of the target trajectory. Individual trajectory portions may be trained during separate training trials. Some portions may be associated with robot executing complex actions and may require additional training trials and/or more dense training input compared to simpler trajectory actions.

Abstract: Apparatus and methods for plasticity in spiking neuron networks. In various implementations, the efficacy of one or more connections of the network may be adjusted based on a plasticity rule during network operation. The rule may comprise a connection depression portion and/or a potentiation portion. Statistical parameters of the adjusted efficacy of a population of connections may be determined. The statistical parameter(s) may be utilized to adapt the plasticity rule during network operation in order to obtain efficacy characterized by target statistics. Based on the statistical parameter exceeding a target value, the depression magnitude of the plasticity rule may be reduced. Based on a statistical parameter being below the target value, the depression magnitude of the plasticity rule may be increased. The use of adaptive modification of the plasticity rule may improve network convergence while alleviating a need for manual tuning of efficacy during network operation.

Abstract: Apparatus and methods for a modular robotic device with artificial intelligence that is receptive to training controls. In one implementation, modular robotic device architecture may be used to provide all or most high cost components in an autonomy module that is separate from the robotic body. The autonomy module may comprise controller, power, actuators that may be connected to controllable elements of the robotic body. The controller may position limbs of the toy in a target position. A user may utilize haptic training approach in order to enable the robotic toy to perform target action(s). Modular configuration of the disclosure enables users to replace one toy body (e.g., the bear) with another (e.g., a giraffe) while using hardware provided by the autonomy module. Modular architecture may enable users to purchase a single AM for use with multiple robotic bodies, thereby reducing the overall cost of ownership.

Abstract: Apparatus and methods for developing robotic controllers comprising parallel networks. In some implementations, a parallel network may comprise at least first and second neuron layers. The second layer may be configured to determine a measure of discrepancy (error) between a target network output and actual network output. The network output may comprise control signal configured to cause a task execution by the robot. The error may be communicated back to the first neuron layer in order to adjust efficacy of input connections into the first layer. The error may be encoded into spike latency using linear or nonlinear encoding. Error communication and control signal provision may be time multiplexed so as to enable target action execution. Efficacy associated with forward and backward/reverse connections may be stored in individual arrays. A synchronization mechanism may be employed to match forward/reverse efficacy in order to implement plasticity.

Abstract: Apparatus and methods for controlling attention and training of autonomous robotic devices. In one approach, attention of the robot may be manipulated by use of a spot-light device illuminating a portion of the aircraft undergoing inspection in order to indicate to inspection robot target areas requiring more detailed inspection. The robot guidance may be aided by way of an additional signal transmitted by the agent to the robot indicating that the object has been illuminated and attention switch may be required. Responsive to receiving the additional signal, the robot may initiate a search for the signal reflected by the illuminated area requiring its attention. Responsive to detecting the illuminated object and receipt of the additional signal, the robot may develop an association between the two events and the inspection task. The light guided attention system may influence the robot learning for subsequent actions.

Abstract: Apparatus and methods for processing inputs by one or more neurons of a network. The neuron(s) may generate spikes based on receipt of multiple inputs. Latency of spike generation may be determined based on an input magnitude. Inputs may be scaled using for example a non-linear concave transform. Scaling may increase neuron sensitivity to lower magnitude inputs, thereby improving latency encoding of small amplitude inputs. The transformation function may be configured compatible with existing non-scaling neuron processes and used as a plug-in to existing neuron models. Use of input scaling may allow for an improved network operation and reduce task simulation time.

Abstract: A method and an apparatus for shaping of lawns and hedges into desired 3D patterns or shapes. The apparatus consists of a bStem and/or other computational device comprising storage, a motorized platform, and trimmer end effectors. The computational device instructs the end effectors to extend or retract as the platform moves along at a steady pace, thus producing a target pattern (e.g., a company logo) in a hedge, lawn, a wall or a ground-cover of any material suitable for such shaping. The apparatus may be configured to operate autonomously based on a pre-loaded pattern file. Software (e.g., such as BrainOS) may be used to provide real-time feedback to trimmers regarding the process and the results, and possibly to train the inverse model accordingly. The apparatus may learn to minimize predicted or current mismatches between the desired pattern and the one being produced. Users compete for the best designs.

Abstract: Artificial retina may be implemented. A retinal apparatus may comprise an input pixel layer, hidden photoreceptive layer, an output neuron layer, and/or other components. Individual cones of the photoreceptive layer may be configured to receive input stimulus from one or more cones within the cone circle of confusion. The cone dynamic may be described using a diffusive state equation characterized by two variables configured to represent membrane voltage and current. Diffusive horizontal coupling of neighboring cones may effectuate non-separable spatiotemporal response that is configured to respond to contrast reversing and/or coherent moving stimulus. The photoreceptive layer high-pass filtered output may facilitate contrast detection by suppressing time-invariant component of the input and reducing sensitivity of the retina to the static inputs.

Abstract: Object recognition apparatus and methods useful for extracting information from sensory input. In one embodiment, the input signal is representative of an element of an image, and the extracted information is encoded in a pulsed output signal. The information is encoded in one variant as a pattern of pulse latencies relative to an occurrence of a temporal event; e.g., the appearance of a new visual frame or movement of the image. The pattern of pulses advantageously is substantially insensitive to such image parameters as size, position, and orientation, so the image identity can be readily decoded. The size, position, and rotation affect the timing of occurrence of the pattern relative to the event; hence, changing the image size or position will not change the pattern of relative pulse latencies but will shift it in time, e.g., will advance or delay its occurrence.

Abstract: Apparatus and methods for developing parallel networks. In some implementations, a network may be partitioned into multiple partitions, wherein individual portions are being executed by respective threads executed in parallel. Individual portions may comprise multiple neurons and synapses. In order to reduce cross-thread traffic and/or reduce number of synchronization locks, network may be partitioned such that for given network portion, the neurons and the input synapses into neurons within the portion are executed within the same thread. Synapse update rules may be configured to allow memory access for postsynaptic neurons and forbid memory access to presynaptic neurons. Individual threads may be afforded pairs of memory buffers configured to effectuate asynchronous data input/output to/from thread. During an even iteration of network operation, even buffer may be utilized to store data generated by the thread during even iteration.

Abstract: Apparatus and methods for training of robotic devices. Robotic devices may be trained by a user guiding the robot along target trajectory using an input signal. A robotic device may comprise an adaptive controller configured to generate control commands based on one or more of the user guidance, sensory input, and/or performance measure. Training may comprise a plurality of trials. During first trial, the user input may be sufficient to cause the robot to complete the trajectory. During subsequent trials, the user and the robot's controller may collaborate so that user input may be reduced while the robot control may be increased. Individual contributions from the user and the robot controller during training may be may be inadequate (when used exclusively) to complete the task. Upon learning, user's knowledge may be transferred to the robot's controller to enable task execution in absence of subsequent inputs from the user.

Abstract: A data processing apparatus may utilize an artificial neuron network configured to reduce dimensionality of input data using a sparse transformation configured using receptive field structure of network units. Output of the network may be analyzed for temporally persistency that is characterized by similarity matrix. Elements of the matrix may be incremented when present activity unit activity at a preceding frame. The similarity matrix may be partitioned based on a distance measure for a given element of the matrix and its closest neighbors. Stability of learning of temporally proximal patterns may be greatly improved as the similarity matrix is learned independently of the partitioning operation. Partitioning of the similarity matrix using the methodology of the disclosure may be performed online, e.g., contemporaneously with the encoding and/or similarity matrix construction, thereby enabling learning of new features in the input data.

Abstract: Adaptive controller apparatus of a plant may be implemented. The controller may comprise an encoder block and a control block. The encoder may utilize basis function kernel expansion technique to encode an arbitrary combination of inputs into spike output. The controller may comprise spiking neuron network operable according to reinforcement learning process. The network may receive the encoder output via a plurality of plastic connections. The process may be configured to adaptively modify connection weights in order to maximize process performance, associated with a target outcome. The relevant features of the input may be identified and used for enabling the controlled plant to achieve the target outcome.

Abstract: Apparatus and methods for a modular robotic device with artificial intelligence that is receptive to training controls. In one implementation, modular robotic device architecture may be used to provide all or most high cost components in an autonomy module that is separate from the robotic body. The autonomy module may comprise controller, power, actuators that may be connected to controllable elements of the robotic body. The controller may position limbs of the toy in a target position. A user may utilize haptic training approach in order to enable the robotic toy to perform target action(s). Modular configuration of the disclosure enables users to replace one toy body (e.g., the bear) with another (e.g., a giraffe) while using hardware provided by the autonomy module. Modular architecture may enable users to purchase a single AM for use with multiple robotic bodies, thereby reducing the overall cost of ownership.

Abstract: Robots have the capacity to perform a broad range of useful tasks, such as factory automation, cleaning, delivery, assistive care, environmental monitoring and entertainment. Enabling a robot to perform a new task in a new environment typically requires a large amount of new software to be written, often by a team of experts. It would be valuable if future technology could empower people, who may have limited or no understanding of software coding, to train robots to perform custom tasks. Some implementations of the present invention provide methods and systems that respond to users' corrective commands to generate and refine a policy for determining appropriate actions based on sensor-data input. Upon completion of learning, the system can generate control commands by deriving them from the sensory data. Using the learned control policy, the robot can behave autonomously.

Abstract: A robotic vehicle may be operated by a learning controller comprising a trainable convolutional network configured to determine control signal based on sensory input. An input network layer may be configured to transfer sensory input into a hidden layer data using a filter convolution operation. Input layer may be configured to transfer sensory input into hidden layer data using a filter convolution. Output layer may convert hidden layer data to a predicted output using data segmentation and a fully connected array of efficacies. During training, efficacy of network connections may be adapted using a measure determined based on a target output provided by a trainer and an output predicted by the network. A combination of the predicted and the target output may be provided to the vehicle to execute a task. The network adaptation may be configured using an error back propagation method. The network may comprise an input reconstruction.

Abstract: Robotic devices may be trained by a user guiding the robot along target action trajectory using an input signal. A robotic device may comprise an adaptive controller configured to generate control signal based on one or more of the user guidance, sensory input, performance measure, and/or other information. Training may comprise a plurality of trials, wherein for a given context the user and the robot's controller may collaborate to develop an association between the context and the target action. Upon developing the association, the adaptive controller may be capable of generating the control signal and/or an action indication prior and/or in lieu of user input. The predictive control functionality attained by the controller may enable autonomous operation of robotic devices obviating a need for continuing user guidance.

Abstract: Apparatus and methods for encoding sensory input information into patterns of pulses and message multiplexing. In one implementation, the patterns of pulses are polychronous (time-locked by not necessary synchronous), and a retinal prosthetic encodes the input signal into the polychronous patterns for delivery via stimulating electrodes. Different polychronous patterns simultaneously encode different sensory signals; (such as different features of the image), thus providing for message multiplexing. Increasing data transmission capacity allows for a reduction in the number of electrodes required for data transmission. In one implementation, an adaptive feedback mechanism is employed to facilitate encoder operation. In another aspect, a computer vision system is described.

Abstract: Sensory encoder may be implemented. Visual encoder apparatus may comprise spiking neuron network configured to receive photodetector input. Excitability of neurons may be adjusted and output spike may be generated based on the input. When neurons generate spiking response, spiking threshold may be dynamically adapted to produce desired output rate. The encoder may dynamically adapt its input range to match statistics of the input and to produce output spikes at an appropriate rate and/or latency. Adaptive input range adjustment and/or spiking threshold adjustment collaborate to enable recognition of features in sensory input of varying dynamic range.

Abstract: Apparatus and methods for arbitration of control signals for robotic devices. A robotic device may comprise an adaptive controller comprising a plurality of predictors configured to provide multiple predicted control signals based on one or more of the teaching input, sensory input, and/or performance. The predicted control signals may be configured to cause two or more actions that may be in conflict with one another and/or utilize a shared resource. An arbitrator may be employed to select one of the actions. The selection process may utilize a WTA, reinforcement, and/or supervisory mechanisms in order to inhibit one or more predicted signals. The arbitrator output may comprise target state information that may be provided to the predictor block. Prior to arbitration, the predicted control signals may be combined with inputs provided by an external control entity in order to reduce learning time.