Artificial intelligence (AI) has recently seen explosive growth and remarkable successes in several application areas. However, it is becoming clear that the methods that have made this possible are ...subject to several limitations that might inhibit progress towards replicating the more general intelligence seen in humans and other animals. In contrast to current AI methods that focus on specific tasks and rely on large amounts of offline data and extensive, slow, and mostly supervised learning, this
natural intelligence
is quick, versatile, agile, and open-ended. This position paper brings together ideas from neuroscience, evolutionary and developmental biology, and complex systems to analyze why such natural intelligence is possible in animals and suggests that AI should exploit the same strategies to move in a different direction. In particular, it argues that integrated embodiment, modularity, synergy, developmental learning, and evolution are key enablers of natural intelligence and should be at the core of AI systems as well. The analysis in the paper leads to the description of a biologically grounded
deep intelligence
(DI) framework for understanding natural intelligence and developing a new approach to building more versatile, autonomous, and integrated AI. The paper concludes that the dominant paradigm of AI today is unlikely to lead to truly natural general intelligence and that something like the biologically inspired DI framework is needed for that.
The discovery of place cells and other spatially modulated neurons in the hippocampal complex of rodents has been crucial to elucidating the neural basis of spatial cognition. More recently, the ...replay of neural sequences encoding previously experienced trajectories has been observed during consummatory behavior—potentially with implications for rapid learning, quick memory consolidation, and behavioral planning. Several promising models for robotic navigation and reinforcement learning have been proposed based on these and previous findings. Most of these models, however, use carefully engineered neural networks, and sometimes require long learning periods. In this paper, we present a self-organizing model incorporating place cells and replay, and demonstrate its utility for rapid one-shot learning in non-trivial environments with obstacles.
The whole-brain functional connectivity (FC) pattern obtained from resting-state functional magnetic resonance imaging data are commonly applied to study neuropsychiatric conditions such as autism ...spectrum disorder (ASD) by using different machine learning models. Recent studies indicate that both hyper- and hypo- aberrant ASD-associated FCs were widely distributed throughout the entire brain rather than only in some specific brain regions. Deep neural networks (DNN) with multiple hidden layers have shown the ability to systematically extract lower-to-higher level information from high dimensional data across a series of neural hidden layers, significantly improving classification accuracy for such data. In this study, a DNN with a novel feature selection method (DNN-FS) is developed for the high dimensional whole-brain resting-state FC pattern classification of ASD patients vs. typical development (TD) controls. The feature selection method is able to help the DNN generate low dimensional high-quality representations of the whole-brain FC patterns by selecting features with high discriminating power from multiple trained sparse auto-encoders. For the comparison, a DNN without the feature selection method (DNN-woFS) is developed, and both of them are tested with different architectures (i.e., with different numbers of hidden layers/nodes). Results show that the best classification accuracy of
is generated by the DNN-FS approach with 3 hidden layers and 150 hidden nodes (3/150). Remarkably, DNN-FS outperforms DNN-woFS for all architectures studied. The most significant accuracy improvement was
with the 3/150 architecture. The method also outperforms other feature selection methods, e.g., two sample
-test and elastic net. In addition to improving the classification accuracy, a Fisher's score-based biomarker identification method based on the DNN is also developed, and used to identify 32 FCs related to ASD. These FCs come from or cross different pre-defined brain networks including the default-mode, cingulo-opercular, frontal-parietal, and cerebellum. Thirteen of them are statically significant between ASD and TD groups (two sample
-test
< 0.05) while 19 of them are not. The relationship between the statically significant FCs and the corresponding ASD behavior symptoms is discussed based on the literature and clinician's expert knowledge. Meanwhile, the potential reason of obtaining 19 FCs which are not statistically significant is also provided.
We present the results of an ongoing collaboration between computer science and psychology researchers that employs Natural Language Processing (NLP) methods to examine the trajectory of semantic ...space used during group idea generation sessions. Specifically, we track and estimate the region of semantic space being used and the degree to which new ideas expand that space. We present a visualization of this space mapping endeavor and compare human ratings of creativity dimensions (i.e., novelty, task-relevance, and elaboration) to algorithm-based estimations of those same dimensions. The semantic space mapping and algorithm development projects can be used to deliver real-time feedback to human creative groups in order to optimize the collaborative creativity process. The overall goal of this research is to increase the "survival" of novel ideas and their elaboration in the collaborative ideation and subsequent decision processes.
Coordination of multiple degrees of freedom in the performance of dynamic and complex motor tasks presents a challenging neuromuscular control problem. Experiments have inferred that humans exhibit ...self-organized, preferred coordination patterns, which emerge due to actor and task constraints on performance. The purpose of this study was to determine if the set of effective coordination strategies that exist for a task centers on a small number of robust, invariant patterns of behavior.
Kinetic movement patterns computed from a cohort of 780 primarily female adolescent athletes performing a drop vertical jump (DVJ) task were analyzed to discover distinct groups into which individuals could be classified based on the similarity of movement coordination solutions.
Clustering of reduced-dimension joint moment of force time series revealed three very distinct, precisely delineated movement profiles that persisted across trials, and which exhibited different functional performance outcomes, despite no other apparent group differences. The same analysis was also performed on a different task-a single-leg drop landing-which also produced distinct movement profiles; however, the three DVJ profiles did not translate to this task as group assignment was inconsistent between these two tasks.
The task demands of the DVJ and single-leg drop-successful landing, reversal of downward momentum, and, in the case of the DVJ, vertical propulsion toward a maximally positioned target-constrain movement performance such that only a few successful outcomes emerge. Discovery of the observed strategies in the context of associated task constraints may help our understanding of how injury risk movement patterns emerge during specific tasks, as well as how the natural dynamics of the system may be exploited to improve these patterns.
DeepCPG Policies for Robot Locomotion Deshpande, Aditya M.; Hurd, Eric; Minai, Ali A. ...
IEEE transactions on cognitive and developmental systems,
12/2023, Letnik:
15, Številka:
4
Journal Article
Odprti dostop
Central Pattern Generators (CPGs) form the neural basis of the observed rhythmic behaviors for locomotion in legged animals. The CPG dynamics organized into networks allow the emergence of complex ...locomotor behaviors. In this work, we take this inspiration for developing walking behaviors in multi-legged robots. We present novel DeepCPG policies that embed CPGs as a layer in a larger neural network and facilitate end-to-end learning of locomotion behaviors in deep reinforcement learning (DRL) setup. We demonstrate the effectiveness of this approach on physics engine-based insectoid robots. We show that, compared to traditional approaches, DeepCPG policies allow sample-efficient end-to-end learning of effective locomotion strategies even in the case of high-dimensional sensor spaces (vision). We scale the DeepCPG policies using a modular robot configuration and multi-agent DRL. Our results suggest that gradual complexification with embedded priors of these policies in a modular fashion could achieve non-trivial sensor and motor integration on a robot platform. These results also indicate the efficacy of bootstrapping more complex intelligent systems from simpler ones based on biological principles. Finally, we present the experimental results for a proof-of-concept insectoid robot system for which DeepCPG learned policies initially using the simulation engine and these were afterwards transferred to real-world robots without any additional fine-tuning.
Understanding cognition has been a central focus for psychologists, neuroscientists and philosophers for thousands of years, but many of its most fundamental processes remain very poorly understood. ...Chief among these is the process of thought itself: the spontaneous emergence of specific ideas within the stream of consciousness. It is widely accepted that ideas, both familiar and novel, arise from the combination of existing concepts. From this perspective, thought is an emergent attribute of memory, arising from the intrinsic dynamics of the neural substrate in which information is embedded. An important issue in any understanding of this process is the relationship between the emergence of conceptual combinations and the dynamics of the underlying neural networks.
Virtually all theories of ideation hypothesize that ideas arise during the thought process through association, each one triggering the next through some type of linkage, e.g., structural analogy, semantic similarity, polysemy, etc. In particular, it has been suggested that the creativity of ideation in individuals reflects the qualitative structure of conceptual associations in their minds. Interestingly, psycholinguistic studies have shown that semantic networks across many languages have a particular type of structure with small-world, scale free connectivity. So far, however, these related insights have not been brought together, in part because there has been no explicitly neural model for the dynamics of spontaneous thought. Recently, we have developed such a model. Though simplistic and abstract, this model attempts to capture the most basic aspects of the process hypothesized by theoretical models within a neurodynamical framework. It represents semantic memory as a recurrent semantic neural network with itinerant dynamics. Conceptual combinations arise through this dynamics as co-active groups of neural units, and either dissolve quickly or persist for a time as emergent metastable attractors and are recognized consciously as ideas. The work presented in this paper describes this model in detail, and uses it to systematically study the relationship between the structure of conceptual associations in the neural substrate and the ideas arising from this system’s dynamics. In particular, we consider how the small-world and scale-free characteristics influence the effectiveness of the thought process under several metrics, and show that networks with both attributes indeed provide significant advantages in generating unique conceptual combinations.
Rapid and accurate identification of new essential genes in under-studied microorganisms will significantly improve our understanding of how a cell works and the ability to re-engineer ...microorganisms. However, predicting essential genes across distantly related organisms remains a challenge. Here, we present a machine learning-based integrative approach that reliably transfers essential gene annotations between distantly related bacteria. We focused on four bacterial species that have well-characterized essential genes, and tested the transferability between three pairs among them. For each pair, we trained our classifier to learn traits associated with essential genes in one organism, and applied it to make predictions in the other. The predictions were then evaluated by examining the agreements with the known essential genes in the target organism. Ten-fold cross-validation in the same organism yielded AUC scores between 0.86 and 0.93. Cross-organism predictions yielded AUC scores between 0.69 and 0.89. The transferability is likely affected by growth conditions, quality of the training data set and the evolutionary distance. We are thus the first to report that gene essentiality can be reliably predicted using features trained and tested in a distantly related organism. Our approach proves more robust and portable than existing approaches, significantly extending our ability to predict essential genes beyond orthologs.
Interactive or collaborative pick-and-place tasks occur during all kinds of daily activities, for example, when two or more individuals pass plates, glasses, and utensils back and forth between each ...other when setting a dinner table or loading a dishwasher together. In the near future, participation in these collaborative pick-and-place tasks could also include robotic assistants. However, for human-machine and human-robot interactions, interactive pick-and-place tasks present a unique set of challenges. A key challenge is that high-level task-representational algorithms and preplanned action or motor programs quickly become intractable, even for simple interaction scenarios. Here we address this challenge by introducing a bioinspired behavioral dynamic model of free-flowing cooperative pick-and-place behaviors based on low-dimensional dynamical movement primitives and nonlinear action selection functions. Further, we demonstrate that this model can be successfully implemented as an artificial agent control architecture to produce effective and robust human-like behavior during human-agent interactions. Participants were unable to explicitly detect whether they were working with an artificial (model controlled) agent or another human-coactor, further illustrating the potential effectiveness of the proposed modeling approach for developing systems of robust real/embodied human-robot interaction more generally.