Suppose you are told that in an alien language the strings AA, MM, DD, RR are all valid words, whereas BF, QG, CE, TM are not. You are then asked if you think EE is a valid word. Most people identify EE as a valid word; they are sensitive to the fact that all the valid words consist of two identical letters, whereas the invalid words do not. This is known as an identity effect, and has been observed in artificial language learning experiments and in a diverse range of natural languages. I give a formal proof that many popular learning framework, including methods for training neural network of arbitrary number of layers, cannot learn such identity effects. The proof exploits symmetries in the architectures and their training regimes to show that such learners cannot perform with human-like behaviour on these grammatical judgment tasks.

When attempting to correctly interpret signals from noise, many sources of noise are not random, only unwanted. Thesecan be discounted by observing cues that predict the noise and canceling or adjusting accordingly. We trained participantsto classify artificial bird calls of different pitches. Pitch was affected by the intended message or word the bird wascommunicating, as well as the size of the bird (larger birds were given lower pitch overall). Participants could hear thecall and also see an image indicating the size of the bird, allowing them to predict and counteract the effect of size, whichserved as noise when trying to interpret communication. At test, we probed many pitches and sizes outside the range oftraining stimuli, and we analyze the patterns by which participants not only compensate for noise, but extrapolate andgeneralize their compensation to new situations.

We propose a novel framework for the analysis of learning al-gorithms that allows us to say when such algorithms can andcannot generalize certain patterns from training data to testdata. In particular we focus on situations where the rule thatmust be learned concerns two components of a stimulus beingidentical. We call such a basis for discrimination an identity-based rule. Identity-based rules have proven to be difficult orimpossible for certain types of learning algorithms to acquirefrom limited datasets. This is in contrast to human behaviouron similar tasks. Here we provide a framework for rigorouslyestablishing which learning algorithms will fail at generalizingidentity-based rules to novel stimuli. We use this frameworkto show that such algorithms are unable to generalize identity-based rules to novel inputs unless trained on virtually all possi-ble inputs. We demonstrate these results computationally witha multilayer feedforward neural network.

Categorical speech content can often be perceived directlyfrom continuous auditory cues in the speech stream, buthuman-level performance on speech recognition tasksrequires compensation for contextual variables like speakeridentity. Regression modeling by McMurray and Jongman(2011) has suggested that for many fricative phonemes, acompensation scheme can substantially increasecategorization accuracy beyond even the information from 24un-compensated raw speech cues. Here, we simulate thesame dataset instead using a neurally rather than abstractlyimplemented model: a hybrid dynamic neural field model andconnectionist network. Our model achieved slightly loweraccuracy than McMurray and Jongman’s but similar accuracypatterns across most fricatives. Results also comparedsimilarly to more recent models that were also less neurallyinstantiated but somewhat closer fitting to humans inaccuracy. An even less abstracted model is an immediatefuture goal, as is expanding the present model to additionalsensory modalities and constancy/compensation effects.

Often in language and other areas of cognition, whether twocomponents of an object are identical or not determine whetherit is well formed. We call such constraints identity effects.When developing a system to learn well-formedness from ex-amples, it is easy enough to build in an identify effect. But canidentity effects be learned from the data without explicit guid-ance? We provide a simple framework in which we can rig-orously prove that algorithms satisfying simple criteria cannotmake the correct inference. We then show that a broad classof algorithms including deep neural networks with standardarchitecture and training with backpropagation satisfy our cri-teria, dependent on the encoding of inputs. Finally, we demon-strate our theory with computational experiments in which weexplore the effect of different input encodings on the ability ofalgorithms to generalize to novel inputs.

Clear speech is a vocal style used when a speaker wishes to im-prove comprehension, usually due to the presence of externalnoise, less-than-optimal listening conditions, or when they aresimply instructed to speak clearly. Clear speech has many dis-tinguishing features, including increased duration, pitch, andamplitude, as well as the exaggeration of articulatory move-ment. We use game theory to model the phenomenon of clearspeech, and make predictions of how it changes under differ-ent circumstances. We view the behaviours of speakers andhearers when communicating as optimal strategies in commu-nication games. When comprehension becomes more difficult,the optimal strategies of the games shift so that speakers exertmore energy to improve the likelihood of accurate communica-tion. We discuss how our models correspond to experimentalobservations and see what predictions are made for future ex-periments.

The problem of stability has long been a limiting factor in de-veloping neural networks that can grow in size and complex-ity. Outside of particular, narrow parameter ranges, changesin activity can easily result in total loss of control. Humancognition must have reliable means of acting to stay withinthe stable ranges of sensitivity and activation. Learning is onesuch mechanism, and population dynamics are another. Here,we focus on another, often overlooked stability mechanism:cellular homeostasis through metabolism dynamics. We ran avisual change detection experiment designed to strain networkstability while minimizing any learnable patterns. We fit thedata using models with and without cellular energy levels as afactor, finding that the model influenced by its past history ofenergy use was a closer fit to the human data.

Our goal is to develop an efficient framework for fitting stochastic continuous-time models to experimental data incognitive psychology. As a simple test problem, we consider data from an eye-tracking study of attention in learning. For eachsubject, the data for each trial consists of the sequence of stimulus features that the subject fixates on, together with the durationof each fixation. We fit a stochastic differential equation model to this data, using the Approximate Bayesian Computationframework. For each subject we infer posterior distributions for the unknown parameters in the model.