Reinforcement Learning with an Evolved Feature Observer
We define an artificial intelligence consisting of a feature observer which preprocesses all inputs to a reinforcement learner. This AI is then made to control a creature in an artificial environment which, after each step, presents a world state and action reward.
The feature observer is a recurrent neural network under adaptation through evolution using the NEAT algorithm while the online reinforcement learning component uses Sarsa-lambda with a neural network, linear function or tabular representation of the state-action value function. This experiment protocol involves comparing the value of temporal adaptation through the use of recurrence in the preprocessing network against temporal adaptation through the use of larger values of lambda. The reward function is also an interesting area of exploration.
Qualia is currently exploring machine vision through representation of complex objects as nested models. Segmentation initially takes place through Grabcut, seeded with a color histogram to automate the initial trimap. Segments are then represented through invariants including Fourier Descriptors and Hu Moments. A novel Bayesian approach to relationships between features allows a joint distribution over spatial and morphology coordinates to establish probability of class correspondence after training, with lower probabilities resulting in iterative recombinant segmentation until a maximum likelihood is found. Objects described in this way are inspired by the concept of K-Lines in the work of Marvin Minsky.
ASIRRA Image Classification (Cat/Dog)
The AI pipeline for our initial attack on this problem has three stages. Stage one involves an evolutionary search for adequate image preprocessing filters (Textons, Color Histograms, Canny or Gabor responses, etc.). Stage 2 reduces the dimensionality
of the filtered output using a stack of restricted Boltzmann machines (Hinton Deep Belief Net or DBN). Stage 3 classifies the DBN output using traditional supervised machine learning.
All three stages were distributed on the QMulus cloud computer allowing for massively parallel operation. Articulated (full body) as well as selectively pre-processed (facial pose) data sets have been used to date. Accuracy rates in the 80%+ range match state of the art results but leave an obvious great deal of work ahead.
Classification of Human Motion
Qualia applied Hierarchical Temporal Memory (HTM) from Numenta Corporation to the problem of classifying human actions. We used the proprioceptive representation model as developed at CMU as a sensor modality. Because of different test subjects and variations in acquired habits of each, classification involved complex, interactive spatial and temporal patterns. Our solution ultimately achieved 99.7% accuracy through innovative application of temporal-spatial pattern conversion using Toeplitz matrices. Presented at the Numenta HTM Workshop in 2008.
Presented with the need to predict future outcomes given only ambiguous past events, Qualia implemented a Bayesian belief network to model the underlying hidden causes associated with observations. An innovative sampling technique is then used to predict future outcomes given a tradeoff of performance (processing demands) and confidence (prediction accuracy). This approach leverages the power of Bayesian inference without intensely application-dependent calculations. The solution is now in beta.
Genetic Compression of Invariant Representations
Qualia applied genetic algorithms to the problem of determining the saddle point for which the smallest training set best encapsulates domain invariance at the bottom of an AI processing stack. We characterized surprisingly small training requirements with superior accuracy as compared to processor-intensive tuning approaches that did not include genetic training selection. Processor intensity of this approach makes it prohibitive for general use but establishment of a lower-bound on complexity of required training was an interesting result that may be qualitatively projected onto other domains.
Exploration of Neuro-Evolution of Adaptive Topologies (NEAT)
The NEAT algorithm and its derivatives have shown interesting results in a bio-plausible incarnation of machine intelligence. Specifically, the topological nature of emergent behavior in NEAT systems builds on our knowledge of what may some day constitute practical AGI. In an effort to understand this further, Qualia applied NEAT on the QMulus cloud to experimental problems in a predator/prey domain as well as card games, temporal sequence learning and other non-Markov processes. This is an ongoing area of research.
Deep Belief Nets (DBN)
Implemented efficient parallel Restricted Boltzmann Machine (RBM) for use in Deep Belief Network (DBN) stacks that serve as a classification, dimensionality reduction and generative modeling engine. This implementation has enabled research on the nature and topology of the DBN itself as well as visualization, feature extraction and temporal patterns in signal data.
Object Tracking in Video
Qualia implemented a video augmentation system which learns a specified object and then highlights its subsequent movements. The solution pipeline included texture quantization encoding as a preprocessor followed by other techniques operating on a composite representation of statistical measures representing high-dimension pseudo-pixels.
Actor/Environment Problem Coupling Framework
Qualia implemented an actor/environment framework on which to test a number of different problem statements (environments) and Artificial Intelligences (actors). Environments created in this way can provide Markov and non-Markov states, enabling experimentation with diverse intelligences and observation of not just effectiveness but also meta-results such as the speed of convergence, suitability of the solution, sensitivity to parameterization, solution complexity and so on.
Cloud Computing Platform for Distributed Artificial Intelligence Research
QMulus is a cloud computing platform that was developed by Qualia Labs. It operates over the public internet and enables volunteer computers to contribute CPU cycles toward artificial intelligence research. Central servers manage large A.I. jobs spanning days or weeks across hundreds of distributed machines. Existing cloud infrastructure solutions were not usable because of the MIMD nature of algorithm research as compared to other highly parallelizable SIMD efforts such as SETI@Home.
QMulus includes an appropriate security model, self-service installation, consumer portal, developer tools, maintenance processes and automated algorithm distribution. It has been downloaded to over 1000 machines in 30 countries.
Genetic Algorithm Experimentation Platform
Qualia created a flexible genetic algorithm (GA) framework leveraging the QMulus cloud computing infrastructure for use in the study of many different problems, tuning processes and algorithms. The GA framework in and of itself opens interesting research directions but it is generally used to manage and automate many generations of tens of thousands of trial experiments involving higher parts of an overall AI stack. The GA framework includes different evolutionary models, swarming, visualization tools and population analysis.
Proprioceptive Motion Capture
Engineered integration of motion capture cameras to create a machine learning data set based on the human proprioceptive system as specified by the Carnegie Mellon University artificial intelligence lab. This work resulted in a data set of 250,000 records related by deep temporal patterns.