Neural networks trained on biased datasets tend to inadvertently learn spurious correlations, hindering generalization.

Of the useful features, which ones does the model use? Of the task-irrelevant features, which ones does the model represent?

Neural networks trained on biased datasets tend to inadvertently learn spurious correlations, hindering generalization.

High-resolution neural datasets enable foundation models for the next generation of brain-computer interfaces and neurological treatments. The community requires rigorous benchmarks to discriminate between competing modeling approaches, yet no standardized evaluation frameworks exist for intracranial EEG (iEEG) recordings. To address this gap, we present Neuroprobe: a suite of decoding tasks for studying multi-modal language processing in the brain. Unlike scalp EEG, intracranial EEG requires invasive surgery to implant electrodes that record neural activity directly from the brain with minimal signal distortion. Neuroprobe is built on the BrainTreebank dataset, which consists of 40 hours of iEEG recordings from 10 human subjects performing a naturalistic movie viewing task. Neuroprobe serves two critical functions. First, it is a mine from which neuroscience insights can be drawn. Its high temporal and spatial resolution allows researchers to systematically determine when and where computations for each aspect of language processing occur in the brain by measuring the decodability of each feature across time and all electrode locations. Using Neuroprobe, we visualize how information flows from the superior temporal gyrus to the prefrontal cortex, and the progression from simple auditory features to more complex language features in a purely data-driven manner. Second, as the field moves toward neural foundation models, Neuroprobe provides a rigorous framework for comparing competing architectures and training protocols. We found that the linear baseline is surprisingly strong, beating frontier foundation models on many tasks. Neuroprobe is designed with computational efficiency and ease of use in mind. We make the code for Neuroprobe openly available and maintain a public leaderboard, aiming to enable rapid progress in the field of iEEG foundation models, at https://neuroprobe.dev/

In-context learning (ICL) research often considers learning a function in-context through a uniform sample of input-output pairs. Here, we investigate how presenting a compositional subtask curriculum in context may alter the computations a transformer learns. We design a compositional algorithmic task based on the modular exponential-a double exponential task composed of two single exponential subtasks and train transformer models to learn the task in-context. We compare (a) models trained using an in-context curriculum consisting of single exponential subtasks and, (b) models trained directly on the double exponential task without such a curriculum. We show that models trained with a subtask curriculum can perform zero-shot inference on unseen compositional tasks and are more robust given the same context length. We study how the task and subtasks are represented across the two training regimes. We find that the models employ diverse strategies modulated by the specific curriculum design.

When neural networks are trained on biased datasets, they tend to inadvertently learn spurious correlations, leading to challenges in achieving strong generalization and robustness. Current approaches to address such biases typically involve utilizing bias annotations, reweighting based on pseudo-bias labels, or enhancing diversity within bias-conflicting data points through augmentation techniques. We introduce DeNetDM, a novel debiasing method based on the observation that shallow neural networks prioritize learning core attributes, while deeper ones emphasize biases when tasked with acquiring distinct information. Using a training paradigm derived from Product of Experts, we create both biased and debiased branches with deep and shallow architectures and then distill knowledge to produce the target debiased model. Extensive experiments and analyses demonstrate that our approach outperforms current debiasing techniques, achieving a notable improvement of around 5% in three datasets, encompassing both synthetic and real-world data. Remarkably, DeNetDM accomplishes this without requiring annotations pertaining to bias labels or bias types, while still delivering performance on par with supervised counterparts. Furthermore, our approach effectively harnesses the diversity of bias-conflicting points within the data, surpassing previous methods and obviating the need for explicit augmentation-based methods to enhance the diversity of such bias-conflicting points. The source code will be available upon acceptance.

Covert communication (also known as steganography) is the practice of concealing a secret inside an innocuous-looking public object (cover) so that the modified public object (covert code) makes sense to everyone but only someone who knows the code can extract the secret (message). Linguistic steganography is the practice of encoding a secret message in natural language text such as spoken conversation or short public communications such as tweets.. While ad hoc methods for covert communications in specific domains exist ( JPEG images, Chinese poetry, etc), there is no general model for linguistic steganography specifically. We present a novel mathematical formalism for creating linguistic steganographic codes, with three parameters: Decodability (probability that the receiver of the coded message will decode the cover correctly), density (frequency of code words in a cover code), and detectability (probability that an attacker can tell the difference between an untampered cover compared to its steganized version). Verbal or linguistic steganography is most challenging because of its lack of artifacts to hide the secret message in. We detail a practical construction in Python of a steganographic code for Tweets using inserted words to encode hidden digits while using n-gram frequency distortion as the measure of detectability of the insertions. Using the publicly accessible Stanford Sentiment Analysis dataset we implemented the tweet steganization scheme -- a codeword (an existing word in the data set) inserted in random positions in random existing tweets to find the tweet that has the least possible n-gram distortion. We argue that this approximates KL distance in a localized manner at low cost and thus we get a linguistic steganography scheme that is both formal and practical and permits a tradeoff between codeword density and detectability of the covert message.

In naturalistic learning problems, a model's input contains a wide range of features, some useful for the task at hand, and others not. Of the useful features, which ones does the model use? Of the task-irrelevant features, which ones does the model represent? Answers to these questions are important for understanding the basis of models' decisions, as well as for building models that learn versatile, adaptable representations useful beyond the original training task. We study these questions using synthetic datasets in which the task-relevance of input features can be controlled directly. We find that when two features redundantly predict the labels, the model preferentially represents one, and its preference reflects what was most linearly decodable from the untrained model. Over training, task-relevant features are enhanced, and task-irrelevant features are partially suppressed. Interestingly, in some cases, an easier, weakly predictive feature can suppress a more strongly predictive, but more difficult one. Additionally, models trained to recognize both easy and hard features learn representations most similar to models that use only the easy feature. Further, easy features lead to more consistent representations across model runs than do hard features. Finally, models have greater representational similarity to an untrained model than to models trained on a different task. Our results highlight the complex processes that determine which features a model represents.