From neuroscience to AI

In machine learning, feature selection involves finding a reduced subset of features that captures most of the information required to train an accurate and efficient model. However, as datasets grow in size and complexity - particularly with the rise of high-dimensional text embeddings from large language models - traditional and recent feature selection methods struggle to scale effectively on the volume of enterprise data. Evolutionary algorithms like Particle Swarm Optimization (PSO) become computationally prohibitive when applied to large datasets. Hybrid approaches that combine filter and wrapper methods offer some relief, but they rely on arbitrary or fixed feature cutoffs, which can lead to suboptimal performance and inefficiency. This work presents FeatureCuts, a novel feature selection algorithm that adaptively selects the optimal feature cutoff after performing filter-based ranking. Evaluated on 15 publicly available datasets and one industry dataset, FeatureCuts achieved on average, up to 14 percentage points more feature reduction and 99.6 % less computation time while maintaining model performance, compared to state-of-the-art methods. When the selected features are used in a wrapper method such as PSO, it enables 25 percentage points more feature reduction, requires 66 % less computation time, and maintains model performance when compared to PSO alone. The minimal overhead of FeatureCuts makes it scalable for large datasets typically seen in enterprise applications.

Unsupervised representation learning on mixed data is highly challenging but rarely explored. It has to tackle significant challenges related to common issues in real-life mixed data, including sparsity, dynamics and heterogeneity of attributes and values. This work introduces an effective and efficient unsupervised deep representer called Mix2Vec to automatically learn a universal representation of dynamic mixed data with the above complex characteristics. Mix2Vec is empowered with three effective mechanisms: random shuffling prediction, prior distribution matching, and structural informativeness maximization, to tackle the aforementioned challenges. These mechanisms are implemented as an unsupervised deep neural representer Mix2Vec. Mix2Vec converts complex mixed data into vector space-based representations that are universal and comparable to all data objects and transparent and reusable for both unsupervised and supervised learning tasks. Extensive experiments on four large mixed datasets demonstrate that Mix2Vec performs significantly better than state-of-the-art deep representation methods. We also empirically verify the designed mechanisms in terms of representation quality, visualization and capability of enabling better performance of downstream tasks.

Background Phosphenes induced by transcranial magnetic stimulation (TMS) are a subjectively described visual phenomenon employed in basic and clinical research as index of the excitability of retinotopically organized areas in the brain. Objective Phosphene threshold estimation is a preliminary step in many TMS experiments in visual cognition for setting the appropriate level of TMS doses; however, the lack of a direct comparison of the available methods for phosphene threshold estimation leaves unsolved the reliability of those methods in setting TMS doses. The present work aims at fulfilling this gap. Methods We compared the most common methods for phosphene threshold calculation, namely the Method of Constant Stimuli (MOCS), the Modified Binary Search (MOBS) and the Rapid Estimation of Phosphene Threshold (REPT). In two experiments we tested the reliability of PT estimation under each of the three methods, considering the day of administration, the participants' expertise in phosphene perception and the sensitivity of each method to the initial values used for the threshold calculation. Results We found that MOCS and REPT have comparable reliability when estimating phosphene thresholds, while MOBS estimations appear less stable. Conclusions Based on our results, researchers and clinicians can estimate phosphene threshold according to MOCS or REPT equally reliably, depending on their specific investigation goals. We suggest several important factors for consideration when calculating phosphene thresholds and describe strategies to adopt in experimental procedures.

Background. Two experiments investigated the effect of features of human behaviour on the quality of interaction with an Embodied Conversational Agent (ECA). Methods. In Experiment 1, visual prominence cues (head nod, eyebrow raise) of the ECA were manipulated to explore the hypothesis that likeability of an ECA increases as a function of interpersonal mimicry. In the context of an error detection task, the ECA either mimicked or did not mimic a head nod or brow raise that humans produced to give emphasis to a word when correcting the ECA’s vocabulary. In Experiment 2, presence versus absence of facial expressions on comprehension accuracy of two computer-driven ECA monologues was investigated. Results. In Experiment 1, evidence for a positive relationship between ECA mimicry and lifelikeness was obtained. However, a mimicking agent did not elicit more human gestures. In Experiment 2, expressiveness was associated with greater comprehension and higher ratings of humour and engagement. Conclusion. Influences from mimicry can be explained by visual and motor simulation, and bidirectional links between similarity and liking. Cue redundancy and minimizing cognitive load are potential explanations for expressiveness aiding comprehension.

Background Transcranial magnetic stimulation (TMS) is a popular functional mapping tool in cognitive and sensory neuroscience. While strong TMS typically degrades performance, two recent studies have demonstrated that weak TMS, delivered to visual cortex, can improve performance on simple visual tasks. The improvement was interpreted as the summation of visually-evoked and TMS-elicited neuronal activity in visual cortex, but the nature of this interaction remains unclear. Objective The present experiments sought to determine whether these weak pulses of TMS assist subjects to see the visual stimulus itself or create a distinct “melded” percept that may not be recognizable as the visual stimulus. Methods We measured contrast thresholds in an orientation discrimination task in which participants reported the orientation (left or right) of gratings tilted 45° from vertical. Results Weak TMS improved sensitivity for identifying gratings, suggesting that TMS sums with but preserves orientation information so that the subject can recognize the visual stimulus. We explain the effect using a mechanism of non-linear transduction of sensory signals in the brain. Conclusions The capability of low-intensity TMS to augment the neural signal while preserving information encoded in the stimulus can be employed as a novel approach to study the neural correlates of consciousness by selectively “pushing” an unconscious stimulus into consciousness.

A dissociation between visual awareness and visual discrimination is referred to as “blindsight”. Blindsight results from loss of function of the primary visual cortex (V1) which can occur due to cerebrovascular accidents (i.e. stroke-related lesions). There are also numerous reports of similar, though reversible, effects on vision induced by transcranial Magnetic Stimulation (TMS) to early visual cortex. These effects point to V1 as the “gate” of visual awareness and have strong implications for understanding the neurological underpinnings of consciousness. It has been argued that evidence for the dissociation between awareness of, and responses to, visual stimuli can be a measurement artifact of the use of a high response criterion under yes-no measures of visual awareness when compared with the criterion free forced-choice responses. This difference between yes-no and forced-choice measures suggests that evidence for a dissociation may actually be normal near-threshold conscious vision. Here we describe three experiments that tested visual performance in normal subjects when their visual awareness was suppressed by applying TMS to the occipital pole. The nature of subjects’ performance whilst undergoing occipital TMS was then verified by use of a psychophysical measure (d') that is independent of response criteria. This showed that there was no genuine dissociation in visual sensitivity measured by yes-no and forced-choice responses. These results highlight that evidence for visual sensitivity in the absence of awareness must be analysed using a bias-free psychophysical measure, such as d', in order to confirm whether or not visual performance is truly unconscious.

The underlying mechanisms of action of transcranial magnetic stimulation (TMS) are still a matter of debate. TMS may impair a subject's performance by increasing neural noise, suppressing the neural signal, or both. Here, we delivered a single pulse of TMS (spTMS) to V5/MT during a motion direction discrimination task while concurrently manipulating the level of noise in the motion stimulus. Our results indicate that spTMS essentially acts by suppressing the strength of the relevant visual signal. We suggest that TMS may induce a pattern of neural activity that complements the ongoing activation elicited by the sensory signal in a manner that partially impoverishes that signal.

To calibrate the intensity of transcranial magnetic stimulation (TMS) at the occipital pole, the phosphene threshold is used as a measure of cortical excitability. The phosphene threshold (PT) refers to the intensity of magnetic stimulation that induces illusory flashes of light (phosphenes) on a proportion of trials. The existing PT estimation procedures lack the accuracy and mathematical rigour of modern threshold estimation methods. We present an improved and automatic procedure for estimating the PT which is based on the well-established Ψ Bayesian adaptive staircase approach. To validate the new procedure, we compared it with another commonly used procedure for estimating the PT. We found that our procedure is more accurate, reliable, and rapid when compared with an existing PT measurement procedure. The new procedure is implemented in Matlab and works automatically with the Magstim Rapid2 stimulator using a convenient graphical user interface. The Matlab program is freely available for download.

Under conditions of inattention or deficits in orienting attention, special classes of stimuli (e.g. faces, bodies) are more likely to be perceived than other stimuli. This suggests that biologically salient visual stimuli automatically recruit attention, even when they are task-irrelevant or ignored. Here we report results from a behavioral experiment with female and male subjects and two magnetoencephalography (MEG) experiments with male subjects only, in which we investigated attentional capture with face and hand stimuli. In both the behavioral and MEG experiments, subjects were required to count the number of gender-specific targets from either face or hand categories within a block of stimuli. In the behavioral experiment, we found that male subjects were significantly more accurate in response to female than male face target blocks. There was no corresponding effect found in response to hand target blocks. Female subjects did not show a gender-based difference in response to face or hand target blocks. MEG results indicated that the male subjects' responses to face stimuli in primary visual cortex (V1) and the face-selective part of the fusiform gyrus (FG) were reduced when male face stimuli were not relevant to the task, whereas female faces maintained a strong response in these areas in both task-relevant and task-irrelevant conditions. These results suggest that within the male brain, female face stimuli are more resilient to suppression than male faces, once attention is drawn to the part of the visual field where the face appears.

Despite the complete imprint of a visual scene on the retina, the brain selects particular items for further processing. However, there is considerable debate about when and where the first attentional effects take hold in the cortex. We examined the timing of face specificity and attentional influences in the primary/secondary visual cortex (V1/V2) and in the fusiform gyrus (FG) in two experiments using magnetoencephalography (MEG). In experiment 1, using a passive viewing task, we identified three components in response to "Face," "Hand," and "Shoe" stimuli bilaterally in the FG: M(FG)100, M(FG)170, and M(FG)200-all showing a stronger preference for faces. The timing of these three activations of the FG is consistent with earlier studies claiming distinct stages of processing of visual stimuli in the first 300 ms. In experiment 2, subjects performed a gender-discrimination task on either faces or hands, drawing attention to only one of the two object categories. In addition to the previously identified three components in FG, here we found object-selective attentional enhancement first appearing in V1/V2 at around 170 ms, and then in FG at around 200 ms, i.e. concurrent with the third component. No attentional effects were evident on the first or second magnetoencephalography components. These findings may indicate that the visual input for an object is first encoded and matched to an attended "cue" object held in mind. When the attended and encoded objects match, a third stage involving attentive processing is enhanced.