Categorical Perception: A Groundwork for Deep Learning

Classification is one of the major tasks that deep learning is successfully tackling. Categorization is also a fundamental cognitive ability. A well-known perceptual consequence of categorization in humans and other animals, categorical per ception, is notably characterized by a within-category compression and a between-category separation: two items, close in input space, are perceived closer if they belong to the same category than if they belong to different categories. Elaborating on experimental and theoretical results in cognitive science, here we study categorical effects in artificial neural networks. We combine a theoretical analysis that makes use of mutual and Fisher information quantities and a series of numerical simulations on networks of increasing complexity. These formal and numerical analyses provide insights into the geometry of the neural representation in deep layers, with expansion of space near category boundaries and contraction far from category boundaries. We investigate categorical representation by using two complementary approaches: one mimics experiments in psychophysics and cognitive neuroscience by means of morphed continua between stimuli of different categories, while the other introduces a categoricality index that, for each layer in the network, quantifies the separability of the categories at the neural population level. We show on both shallow and deep neural networks that category learning automatically induces categorical perception. We further show that the deeper a layer, the stronger the categorical effects. As an outcome of our study, we propose a coherent view of the efficacy of different heuristic practices of the dropout regularization technique. More generally, our view, which finds echoes in the neuroscience literature, insists on the differential impact of noise in any given layer depending on the geometry of the neural representation that is being learned, that is, on how this geometry reflects the structure of the categories.

Composite receptive fields in the mouse auditory cortex

A central question in sensory neuroscience is how neurons represent complex natural stimuli. This process involves multiple steps of feature extraction to obtain a condensed, categorical representation useful for classification and behavior. It has previously been shown that central auditory neurons in the starling have composite receptive fields composed of multiple features when probed with conspecific songs. Whether this property is an idiosyncratic characteristic of songbirds, a group of highly specialized vocal learners, or a generic characteristic of central auditory systems in different animals is, however, unknown. To address this question, we have recorded responses from auditory cortical neurons in mice, and characterized their receptive fields using mouse ultrasonic vocalizations (USVs) as a natural and ethologically relevant stimulus and pitch-shifted starling songs as a natural but ethologically irrelevant control stimulus. We have found that auditory cortical neurons in the mouse display composite receptive fields with multiple excitatory and inhibitory subunits. Moreover, this was the case with either the conspecific or the heterospecific vocalizations. We then trained the sparse filtering algorithm on both classes of natural stimuli to obtain statistically optimal features, and compared the natural and artificial features using UMAP, a dimensionality-reduction algorithm previously used to analyze mouse USVs and birdsongs. We have found that the receptive-field features obtained with the mouse USVs and those obtained with the pitch-shifted starling songs clustered together, as did the sparse-filtering features. However, the natural and artificial receptive-field features clustered mostly separately. These results indicate that composite receptive fields are likely a generic property of central auditory systems in different classes of animals. They further suggest that the quadratic receptive-field features of the mouse auditory cortical neurons are natural-stimulus invariant.

The categorical use of a continuous time representation

The abstract concept of time is mentally represented as a spatially oriented line, with the past associated with the left space and the future associated with the right. Although the line is supposed to be continuous, most available evidence is also consistent with a categorical representation that only discriminates between past and future. The aim of the present study was to test the continuous or categorical nature of the mental timeline. Italian participants judged the temporal reference of 20 temporal expressions by pressing keys on either the left or the right. In Experiment 1 (N = 32), all words were presented at the center of the screen. In Experiment 2 (N = 32), each word was presented on the screen in a central, left, or right position. In Experiment 3 (N = 32), all text was mirror-reversed. In all experiments, participants were asked to place the 20 temporal expressions on a 10-cm line. The results showed a clear Spatial–TEmporal Association of Response Codes (STEARC) effect which did not vary in strength depending on the location of the temporal expressions on the line. However, there was also a clear Distance effect: latencies were slower for words that were closer to the present than further away. We conclude that the mental timeline is a continuous representation that can be used in a categorical way when an explicit past vs. future discrimination is required by the task.

Typological Profiles of Depression of College Students: Latent Classes and Controlling for Response Bias of the Baptista Depression Short-Scale (EBADEP-short)

Response styles and non-linearity might reduce the validity of scores on depression inventories. To address both issues, we explored the latent class structure of the Baptista’s Depression Scale (EBADEP), and the influence of extreme response bias. In total, 1,137 Brazilian college students (M = 26 years, SD = 7.3) participated in this study. Taxometric analysis yielded ambiguous results, without clear support for either a dimensional or a categorical representation of the data. We found three latent classes: one comprising participants with a tendency to endorse items about sadness, angst, pessimism, and low self-efficacy; another with individuals scoring low on all symptoms; and a third with intermediate scores. We found no relationship between the composition of latent classes and extreme response. Participants who reported having received a diagnostic of depression were more likely to belong to the first latent class. These findings validate the clinical usefulness of a latent class structure for the EBADEP.

CDE++: Learning Categorical Data Embedding by Enhancing Heterogeneous Feature Value Coupling Relationships

Categorical data are ubiquitous in machine learning tasks, and the representation of categorical data plays an important role in the learning performance. The heterogeneous coupling relationships between features and feature values reflect the characteristics of the real-world categorical data which need to be captured in the representations. The paper proposes an enhanced categorical data embedding method, i.e., CDE++, which captures the heterogeneous feature value coupling relationships into the representations. Based on information theory and the hierarchical couplings defined in our previous work CDE (Categorical Data Embedding by learning hierarchical value coupling), CDE++ adopts mutual information and margin entropy to capture feature couplings and designs a hybrid clustering strategy to capture multiple types of feature value clusters. Moreover, Autoencoder is used to learn non-linear couplings between features and value clusters. The categorical data embeddings generated by CDE++ are low-dimensional numerical vectors which are directly applied to clustering and classification and achieve the best performance comparing with other categorical representation learning methods. Parameter sensitivity and scalability tests are also conducted to demonstrate the superiority of CDE++.

Flexible Representations of Odour Categories in the Mouse Olfactory Bulb

SUMMARYThe ability to group sensory stimuli into categories is crucial for efficient interaction with a rich and ever-changing environment. In olfaction, basic features of categorical representation of odours were observed as early as in the olfactory bulb (OB). Categorical representation was described in mitral cells (MCs) as sudden transitions in responses to odours that were morphed along a continuum. However, it remains unclear to what extent such response dynamics actually reflects perceptual categories and decisions therein. Here, we tested the role of learning on category formation in the mouse OB, using in vivo two-photon calcium imaging and behaviour. We imaged MCs responses in naïve mice and in awake behaving mice as they learned two tasks with different classification logic. In one task, a 1-decision boundary task, animals learned to classify odour mixtures based on the dominant compound in the mixtures. As expected, categorical representation of close by odours, which was evident already in naïve animals, further increased following learning. In a second task, a multi-decision boundary task, animals learned to classify odours independent of their chemical similarity. Rather, odour discrimination was based on the meaning ascribed to them (either rewarding or not). Following the second task, odour representations by MCs reorganized according to the odour value in the new category. This functional reorganization was also reflected as a shift from predominantly excitatory odour responses to predominantly inhibitory odour responses. Our data shows that odour representations by MCs is flexible, shaped by task demands, and carry category-related information.

Categorical representation from sound and sight in the ventral occipito-temporal cortex of sighted and blind

Is vision necessary for the development of the categorical organization of the Ventral Occipito-Temporal Cortex (VOTC)? We used fMRI to characterize VOTC responses to eight categories presented acoustically in sighted and early blind individuals, and visually in a separate sighted group. We observed that VOTC reliably encodes sound categories in sighted and blind people using a representational structure and connectivity partially similar to the one found in vision. Sound categories were, however, more reliably encoded in the blind than the sighted group, using a representational format closer to the one found in vision. Crucially, VOTC in blind represents the categorical membership of sounds rather than their acoustic features. Our results suggest that sounds trigger categorical responses in the VOTC of congenitally blind and sighted people that partially match the topography and functional profile of the visual response, despite qualitative nuances in the categorical organization of VOTC between modalities and groups.