Developmental Network-2: The Autonomous Generation of Optimal Internal-Representation Hierarchy.

It is very challenging for machine learning methods to reach the goal of general-purpose learning since there are so many complicated situations in different tasks. The learning methods need to generate flexible internal representations for all scenarios met before. The hierarchical internal representation is considered as an efficient way to build such flexible representations. By hierarchy, we mean important local features in the input can be combined to form higher level features with more context. In this work, we analyze how our proposed general-purpose learning framework-the developmental network-2 (DN-2)-autonomously generates internal hierarchy with new mechanisms. Specifically, DN-2 incrementally allocates neuronal resources to different levels of representation during learning instead of handcrafting static boundaries among different levels of representation. We present the mathematical proof to demonstrate that optimal properties in terms of maximum likelihood (ML) are established under the conditions of limited learning experience and resources. The phoneme recognition and real-world visual navigation experiments that are of different modalities and include many different situations are designed to investigate general-purpose learning capability of DN-2. The experimental results show that DN-2 successfully learns different tasks. The formed internal hierarchical representations focus on important features, and the invariant abstract arise from optimal internal representations. We believe that DN-2 is in the right way toward fully autonomous learning.

Learning to recognize while learning to speak: Self-supervision and developing a speaking motor.

Traditionally, learning speech synthesis and speech recognition were investigated as two separate tasks. This separation hinders incremental development for concurrent synthesis and recognition, where partially-learned synthesis and partially-learned recognition must help each other throughout lifelong learning. This work is a paradigm shift-we treat synthesis and recognition as two intertwined aspects of a lifelong learning agent. Furthermore, in contrast to existing recognition or synthesis systems, babies do not need their mothers to directly supervise their vocal tracts at every moment during the learning. We argue that self-generated non-symbolic states/actions at fine-grained time level help such a learner as necessary temporal contexts. Here, we approach a new and challenging problem-how to enable an autonomous learning system to develop an artificial speaking motor for generating temporally-dense (e.g., frame-wise) actions on the fly without human handcrafting a set of symbolic states. The self-generated states/actions are Muscles-like, High-dimensional, Temporally-dense and Globally-smooth (MHTG), so that these states/actions are directly attended for concurrent synthesis and recognition for each time frame. Human teachers are relieved from supervising learner's motor ends. The Candid Covariance-free Incremental (CCI) Principal Component Analysis (PCA) is applied to develop such an artificial speaking motor where PCA features drive the motor. Since each life must develop normally, each Developmental Network-2 (DN-2) reaches the same network (maximum likelihood, ML) regardless of randomly initialized weights, where ML is not just for a function approximator but rather an emergent Turing Machine. The machine-synthesized sounds are evaluated by both the neural network and humans with recognition experiments. Our experimental results showed learning-to-synthesize and learning-to-recognize-through-synthesis for phonemes. This work corresponds to a key step toward our goal to close a great gap toward fully autonomous machine learning directly from the physical world.

Emergent neural turing machine and its visual navigation.

Traditional Turing Machines (TMs) are symbolic whose hand-crafted representations are static and limited. Developmental Network 1 (DN-1) uses emergent representation to perform Turing Computation. But DN-1 lacks hierarchy in its internal representations, and is hard to handle the complex visual navigation tasks. In this paper, we improve DN-1 with several new mechanisms and presenta new emergent neural Turing Machine - Developmental Network 2 (DN-2). By neural, we mean that the control of the TM has neurons as basic computing elements. The major novelty of DN-2 over DN-1 is that the representational hierarchy inside DN-2 is emergent and fluid. DN-2 grows complex hierarchies by dynamically allowing initialization of neurons with different ranges of connection. We present a complex task - vision guided navigation in simulated and natural worlds using DN-2. A major function that has not been demonstrated before is that the hierarchy enables attention that disregards distracting features based on the navigation context. In simulated navigation experiments, DN-2 can perform with no errors, and in real-world navigation experiments, the error rate is only 0.78%. These experimental results showed that DN-2 successfully learned rules of navigation with image inputs.

Motivated Optimal Developmental Learning for Sequential Tasks Without Using Rigid Time-Discounts.

Many methods for reinforcement learning use symbolic representations-nonemergent-such as Q-learning. We use emergent representations here, without human handcrafted symbolic states (i.e., each state corresponds to a different location). This paper models reinforcement learning for hidden neurons in emergent networks for sequential tasks. In this paper, their influences on sequential tasks (e.g., robot navigation in different scenarios) are investigated where the learned value and results of a behavior rely on not only the current experience just like in a pattern recognition (episodic) but also the prediction of future experiences (e.g., delayed rewards) and environments (e.g., previously learned navigational trajectories). We show that this new model of motivated learning amounts to the computation of the maximum-likelihood estimate through "life" where punishment and reward have increased weights. This new formulation avoids the greediness of time-discount in Q-learning. Its complex nonlinear sequential optimization has been solved in a closed-form procedure under the condition of the limited computational resources and limited learning experience so far, because we convert it into a simpler problem of incremental and linear estimation. The experimental results showed that the serotonin and dopamine systems speed up learning for sequential tasks, because not all events are equally important. As far as we know, this is the first work that studies the influences of reinforcers (via serotonin and dopamine) on hidden neurons (Y neurons) for sequential tasks in dynamic scenarios using emergent representations.

Modulation for emergent networks: serotonin and dopamine.

In autonomous learning, value-sensitive experiences can improve the efficiency of learning. A learning network needs be motivated so that the limited computational resources and the limited lifetime are devoted to events that are of high value for the agent to compete in its environment. The neuromodulatory system of the brain is mainly responsible for developing such a motivation system. Although reinforcement learning has been extensively studied, many existing models are symbolic whose internal nodes or modules have preset meanings. Neural networks have been used to automatically generate internal emergent representations. However, modeling an emergent motivational system for neural networks is still a great challenge. By emergent, we mean that the internal representations emerge autonomously through interactions with the external environments. This work proposes a generic emergent modulatory system for emergent networks, which includes two subsystems - the serotonin system and the dopamine system. The former signals a large class of stimuli that are intrinsically aversive (e.g., stress or pain). The latter signals a large class of stimuli that are intrinsically appetitive (e.g., pleasure or sweet). We experimented with this motivational system for two settings. The first is a visual recognition setting to investigate how such a system can learn through interactions with a teacher, who does not directly give answers, but only punishments and rewards. The second is a setting for wandering in the presence of a friend and a foe.

A computational developmental model for specificity and transfer in perceptual learning.

How and under what circumstances the training effects of perceptual learning (PL) transfer to novel situations is critical to our understanding of generalization and abstraction in learning. Although PL is generally believed to be highly specific to the trained stimulus, a series of psychophysical studies have recently shown that training effects can transfer to untrained conditions under certain experimental protocols. In this article, we present a brain-inspired, neuromorphic computational model of the Where-What visuomotor pathways which successfully explains both the specificity and transfer of perceptual learning. The major architectural novelty is that each feature neuron has both sensory and motor inputs. The network of neurons is autonomously developed from experience, using a refined Hebbian-learning rule and lateral competition, which altogether result in neuronal recruitment. Our hypothesis is that certain paradigms of experiments trigger two-way (descending and ascending) off-task processes about the untrained condition which lead to recruitment of more neurons in lower feature representation areas as well as higher concept representation areas for the untrained condition, hence the transfer. We put forward a novel proposition that gated self-organization of the connections during the off-task processes accounts for the observed transfer effects. Simulation results showed transfer of learning across retinal locations in a Vernier discrimination task in a double-training procedure, comparable to previous psychophysical data (Xiao et al., 2008). To the best of our knowledge, this model is the first neurally-plausible model to explain both transfer and specificity in a PL setting.

Spatio-temporal adaptation in the unsupervised development of networked visual neurons.

There have been many computational models mimicking the visual cortex that are based on spatial adaptations of unsupervised neural networks. In this paper, we present a new model called neuronal cluster which includes spatial as well as temporal weights in its unified adaptation scheme. The "in-place" nature of the model is based on two biologically plausible learning rules, Hebbian rule and lateral inhibition. We present the mathematical demonstration that the temporal weights are derived from the delay in lateral inhibition. By training with the natural videos, this model can develop spatio-temporal features such as orientation selective cells, motion sensitive cells, and spatio-temporal complex cells. The unified nature of the adaptation scheme allows us to construct a multilayered and task-independent attention selection network which uses the same learning rule for edge, motion, and color detection, and we can use this network to engage in attention selection in both static and dynamic scenes.

Multilayer in-place learning networks for modeling functional layers in the laminar cortex.

Currently, there is a lack of general-purpose in-place learning networks that model feature layers in the cortex. By "general-purpose" we mean a general yet adaptive high-dimensional function approximator. In-place learning is a biological concept rooted in the genomic equivalence principle, meaning that each neuron is fully responsible for its own learning in its environment and there is no need for an external learner. Presented in this paper is the Multilayer In-place Learning Network (MILN) for this ambitious goal. Computationally, in-place learning provides unusually efficient learning algorithms whose simplicity, low computational complexity, and generality are set apart from typical conventional learning algorithms. Based on the neuroscience literature, we model the layer 4 and layer 2/3 as the feature layers in the 6-layer laminar cortex, with layer 4 using unsupervised learning and layer 2/3 using supervised learning. As a necessary requirement for autonomous mental development, MILN generates invariant neurons in different layers, with increasing invariance from earlier to later layers and the total invariance in the last motor layer. Such self-generated invariant representation is enabled mainly by descending (top-down) connections. The self-generated invariant representation is used as intermediate representations for learning later tasks in open-ended development.

Incremental hierarchical discriminant regression.

This paper presents incremental hierarchical discriminant regression (IHDR) which incrementally builds a decision tree or regression tree for very high-dimensional regression or decision spaces by an online, real-time learning system. Biologically motivated, it is an approximate computational model for automatic development of associative cortex, with both bottom-up sensory inputs and top-down motor projections. At each internal node of the IHDR tree, information in the output space is used to automatically derive the local subspace spanned by the most discriminating features. Embedded in the tree is a hierarchical probability distribution model used to prune very unlikely cases during the search. The number of parameters in the coarse-to-fine approximation is dynamic and data-driven, enabling the IHDR tree to automatically fit data with unknown distribution shapes (thus, it is difficult to select the number of parameters up front). The IHDR tree dynamically assigns long-term memory to avoid the loss-of-memory problem typical with a global-fitting learning algorithm for neural networks. A major challenge for an incrementally built tree is that the number of samples varies arbitrarily during the construction process. An incrementally updated probability model, called sample-size-dependent negative-log-likelihood (SDNLL) metric is used to deal with large sample-size cases, small sample-size cases, and unbalanced sample-size cases, measured among different internal nodes of the IHDR tree. We report experimental results for four types of data: synthetic data to visualize the behavior of the algorithms, large face image data, continuous video stream from robot navigation, and publicly available data sets that use human defined features.

Auditory learning: a developmental method.

Motivated by the human autonomous development process from infancy to adulthood, we have built a robot that develops its cognitive and behavioral skills through real-time interactions with the environment. We call such a robot a developmental robot. In this paper, we present the theory and the architecture to implement a developmental robot and discuss the related techniques that address an array of challenging technical issues. As an application, experimental results on a real robot, self-organizing, autonomous, incremental learner (SAIL), are presented with emphasis on its audition perception and audition-related action generation. In particular, the SAIL robot conducts the auditory learning from unsegmented and unlabeled speech streams without any prior knowledge about the auditory signals, such as the designated language or the phoneme models. Neither available before learning starts are the actions that the robot is expected to perform. SAIL learns the auditory commands and the desired actions from physical contacts with the environment including the trainers.

Autonomous mental development in high dimensional context and action spaces.

Autonomous Mental Development (AMD) of robots opened a new paradigm for developing machine intelligence, using neural network type of techniques and it fundamentally changed the way an intelligent machine is developed from manual to autonomous. The work presented here is a part of SAIL (Self-Organizing Autonomous Incremental Learner) project which deals with autonomous development of humanoid robot with vision, audition, manipulation and locomotion. The major issue addressed here is the challenge of high dimensional action space (5-10) in addition to the high dimensional context space (hundreds to thousands and beyond), typically required by an AMD machine. This is the first work that studies a high dimensional (numeric) action space in conjunction with a high dimensional perception (context state) space, under the AMD mode. Two new learning algorithms, Direct Update on Direction Cosines (DUDC) and High-Dimensional Conjugate Gradient Search (HCGS), are developed, implemented and tested. The convergence properties of both the algorithms and their targeted applications are discussed. Autonomous learning of speech production under reinforcement learning is studied as an example.

Vision-guided navigation using SHOSLIF.

This paper presents an unconventional approach to vision-guided autonomous navigation. The system recalls information about scenes and navigational experience using content-based retrieval from a visual database. To achieve a high applicability to various road types, we do not impose a priori scene features, such as road edges, that the system must use, but rather, the system automatically derives features from images during supervised learning. To accomplish this, the system uses principle component analysis and linear discriminant analysis to automatically derive the most expressive features (MEF) for scene reconstruction or the most discriminating features (MDF) for scene classification. These features best describe or classify the population of the scenes and approximate complex decision regions using piecewise linear boundaries up to a desired accuracy. A new self-organizing scheme called recursive partition tree (RPT) is used for automatic construction of a vision-and-control database, which quickly prunes the data set in the content-based search and results in a low time complexity of O(log(n)) for retrieval from a database of size n. The system combines principle component and linear discriminant analysis networks with a decision tree network. It has been tested on a mobile robot, Rome, in an unknown indoor environment to learn scenes and the associated navigation experience. In the performing phase, the mobile robot navigates autonomously in similar environments, while allowing the presence of scene perturbations such as the presence of passersby.