A domain-agnostic approach for characterization of lifelong learning systems.

Despite the advancement of machine learning techniques in recent years, state-of-the-art systems lack robustness to "real world" events, where the input distributions and tasks encountered by the deployed systems will not be limited to the original training context, and systems will instead need to adapt to novel distributions and tasks while deployed. This critical gap may be addressed through the development of "Lifelong Learning" systems that are capable of (1) Continuous Learning, (2) Transfer and Adaptation, and (3) Scalability. Unfortunately, efforts to improve these capabilities are typically treated as distinct areas of research that are assessed independently, without regard to the impact of each separate capability on other aspects of the system. We instead propose a holistic approach, using a suite of metrics and an evaluation framework to assess Lifelong Learning in a principled way that is agnostic to specific domains or system techniques. Through five case studies, we show that this suite of metrics can inform the development of varied and complex Lifelong Learning systems. We highlight how the proposed suite of metrics quantifies performance trade-offs present during Lifelong Learning system development - both the widely discussed Stability-Plasticity dilemma and the newly proposed relationship between Sample Efficient and Robust Learning. Further, we make recommendations for the formulation and use of metrics to guide the continuing development of Lifelong Learning systems and assess their progress in the future.

Unsupervised learning of vowel categories from infant-directed speech.

Infants rapidly learn the sound categories of their native language, even though they do not receive explicit or focused training. Recent research suggests that this learning is due to infants' sensitivity to the distribution of speech sounds and that infant-directed speech contains the distributional information needed to form native-language vowel categories. An algorithm, based on Expectation-Maximization, is presented here for learning the categories from a sequence of vowel tokens without (i) receiving any category information with each vowel token, (ii) knowing in advance the number of categories to learn, or (iii) having access to the entire data ensemble. When exposed to vowel tokens drawn from either English or Japanese infant-directed speech, the algorithm successfully discovered the language-specific vowel categories (/I, i, epsilon, e/ for English, /I, i, e, e/ for Japanese). A nonparametric version of the algorithm, closely related to neural network models based on topographic representation and competitive Hebbian learning, also was able to discover the vowel categories, albeit somewhat less reliably. These results reinforce the proposal that native-language speech categories are acquired through distributional learning and that such learning may be instantiated in a biologically plausible manner.

Success and failure of new speech category learning in adulthood: consequences of learned Hebbian attractors in topographic maps.

The influence of a native language on learning new speech sounds in adulthood is addressed using a network model in which speech categories are attractors implemented through interactive activation and Hebbian learning. The network has a representation layer that receives topographic projections from an input layer and has reciprocal excitatory connections with deeper layers. When applied to an experiment in which Japanese adults were trained to distinguish the English /r/-/l/ contrast (McCandliss, Fiez, Protopapas, Conway, & McClelland, 2002), the model can account for many aspects of the experimental results, such as the time course and outcome of the learning, how it varies as a function of feedback, the relative efficacy of adaptive and initially easy training stimuli versus nonadaptive and difficult stimuli, and the development of a discrimination peak at the acquired category boundary. The model is also able to capture some aspects of the individual differences in learning.

Network analysis of toxic chemicals and symptoms: implications for designing first-responder systems.

The rapid and accurate identification of toxic chemicals is critical for saving lives in emergency situations. However, first-responder systems such as WISER typically require a large number of inputs before a chemical can be identified. To address this problem, we used networks to visualize and analyze the complex relationship between toxic chemicals and their symptoms. The results explain why current approaches require a large number of inputs and help to identify regularities related to the co-occurrence of symptoms. This understanding provides implications for the design of future first-responder systems, with the goal of rapidly identifying toxic chemicals in emergency situations.

Perceptuomotor bias in the imitation of steady-state vowels.

Previous studies suggest that speakers are systematically inaccurate, or biased, when imitating self-produced vowels. The direction of these biases in formant space and their variation may offer clues about the organization of the vowel perceptual space. To examine these patterns, three male speakers were asked to imitate 45 self-produced vowels that were systematically distributed in F1/F2 space. All three speakers showed imitation bias, and the bias magnitudes were significantly larger than those predicted by a model of articulatory noise. Each speaker showed a different pattern of bias directions, but the pattern was unrelated to the locations of prototypical vowels produced by that speaker. However, there were substantial quantitative regularities: (1) The distribution of imitation variability and bias magnitudes were similar for all speakers, (2) the imitation variability was independent of the bias magnitudes, and (3) the imitation variability (a production measure) was commensurate with the formant discrimination limen (a perceptual measure). These results indicate that there is additive Gaussian noise in the imitation process that independently affects each formant and that there are speaker-dependent and potentially nonlinguistic biases in vowel perception and production.